Version4 Mrda Manual

User Manual:

Open the PDF directly: View PDF .
Page Count: 132 [warning: Documents this large are best viewed by clicking the View PDF Link!]

Meeting Recorder

Project:

Dialog Act

Labeling Guide

ICSI Technical Report TR-04-002

February 9, 2004

Rajdip Dhillon

Sonali Bhagat

Hannah Carvey

Elizabeth Shriberg

ACKNOWLEDGEMENTS

We especially thank Jeremy Ang for processing our data and Chuck Wooters for

providing us with the TableTrans software. We are also grateful to Don Baron and

Chris Oei for their assistance in preparing data for annotation. We are thankful to

Ashley Krupski for her annotation efforts; Barbara Peskin and Jane Edwards for their

assistance in using the corpus; and Dan Jurafsky and Andrei Popescu-Belis for

supplying us with their input.

This work was supported by a DARPA Communicator project, ICSI NSF ITR Award IIS-

0121396, SRI NASA Award NCC2-1256, SRI NSF IRI-9619921, SRI DARPA ROAR

project N66001-99-D-8504, and by an ICSI award from the Swiss National Science

Foundation through the research network IM2. The views represented herein are those

of the authors and do not represent the views of the funding agencies.

TABLE OF CONTENTS

Introduction......................................................................................................................1

Section 1: Quick Reference Information ..........................................................................2

1.1 Terminology.......................................................................................................2

1.2 Mapping Meeting Recorder DA (MRDA) Tags to SWBD-DAMSL Tags ............3

1.3 Meeting Recorder DA (MRDA) Tagset ..............................................................6

Section 2: Segmentation..................................................................................................8

Section 3: How to Label.................................................................................................15

3.1 Basic Format of DAs and Labels .....................................................................15

3.2 Label Construction...........................................................................................15

3.3 Annotating Utterances Containing Multiple DAs..............................................18

3.4 Disruption Forms .............................................................................................19

3.5 Quotes.............................................................................................................21

3.6 Using TableTrans (Annotation Interface).........................................................22

Section 4: Adjacency Pairs ............................................................................................25

4.1 Purpose and Definition ....................................................................................25

4.2 Labeling Adjacency Pairs ................................................................................25

4.3 Labeling Conventions......................................................................................26

4.4 Restrictions on Using Adjacency Pairs............................................................30

Section 5: Tag Descriptions...........................................................................................32

5.1 Preliminaries....................................................................................................32

5.2 Group 1: Statements .......................................................................................32

5.3 Group 2: Questions .........................................................................................33

5.4 Group 3: Floor Mechanisms ............................................................................43

5.5 Group 4: Backchannels and Acknowledgments ..............................................48

5.6 Group 5: Responses........................................................................................57

5.7 Group 6: Action Motivators ..............................................................................70

5.8 Group 7: Checks..............................................................................................76

5.9 Group 8: Restated Information ........................................................................80

5.10 Group 9: Supportive Functions........................................................................87

5.11 Group 10: Politeness Mechanisms..................................................................92

5.12 Group 11: Further Descriptions .......................................................................96

5.13 Group 12: Disruption Forms ..........................................................................110

5.14 Group 13: Nonlabeled ...................................................................................113

Appendix 1: Labeled Meeting Sample .........................................................................115

Appendix 2: Unused/Merged SWBD-DAMSL Tags .....................................................120

Appendix 3: Unique MRDA Tags.................................................................................123

Appendix 4: Final MRDA Tagset Revisions .................................................................126

Bibliography.................................................................................................................128

Index of Tags...............................................................................................................129

INTRODUCTION

This labeling guide is adapted from work on the Switchboard recordings and the

accompanying manual (Jurafsky et al. 1997). The Switchboard-DAMSL (SWBD-

DAMSL) manual for labeling one-on-one phone conversations provided a useful starting

point for the types of dialog acts (DAs) that arose in the ICSI meeting corpus. However,

the tagset for labeling meetings presented here has been modified as necessary to

better reflect the types of interaction we observed in multiparty face-to-face meetings.

This guide consists of five major sections: Quick Reference Information, Segmentation,

How to Label, Adjacency Pairs, and Tag Descriptions. The first section supplies

definitions for terms used throughout this guide and contains the correspondence of the

Meeting Recorder DA (MRDA) tagset, which is the tagset detailed within this guide, to

the SWBD-DAMSL tagset. This section also contains the entire MRDA tagset

organized into groups according to syntactic, semantic, pragmatic, and functional

similarities of the utterances they mark. The section entitled “Segmentation,” as its

name indicates, details the rules and guidelines governing what constitutes an utterance

along with how to determine utterance boundaries. The third section, “How to Label,”

provides instruction regarding label construction, the management of utterances

requiring additional DAs or containing quotes, and the use of the annotation software.

The section entitled “Adjacency Pairs” details how adjacency pairs are constructed and

the rules governing their usage. The section entitled “Tag Descriptions” provides

explanations of each tag within the MRDA tagset.

Two appendices are also found within this guide. The first provides a labeled portion of

a meeting and the second contains information regarding tags used for a select number

of meetings.

With regard to the examples from meeting data found throughout this guide, it must be

noted that the start and end times for each utterance within the examples do not reflect

the most recent time alignments. However, the start and end times are accurate to a

point which allows for them to be located within their corresponding audio files without

difficulty.

SECTION 1: QUICK REFERENCE INFORMATION

1.1 Terminology

Below is some rudimentary terminology used in dialog act labeling:

utterance: a segment of speech occupying one line in the transcript by

a single speaker which is prosodically and/or syntactically

significant within the conversational context

speech: a group of successive utterances or successive portions of

an utterance

turn: the period during which a speaker has the floor

label: the entire set of DAs and/or other tags applicable to an

utterance

dialog act (DA): the tag or sequence of tags pertaining to the function of an

utterance or portion of an utterance. Each DA contains at

least one general tag and may contain one or more specific

tags, depending upon the nature of the utterance

tag: the individual component(s) of a DA or label

general tag: the tag which represents the basic form of an utterance

(e.g., statement, question, backchannel, etc.)

specific tag: the tag which represents the function or a characteristic of

an utterance and is appended to the general tag (e.g.,

accepting, rejecting, acknowledging, rising tone, etc.)

disruption form: the tag which represents a disruption or otherwise

indiscernible utterance

1.2 Mapping Meeting Recorder DA (MRDA) Tags to

SWBD-DAMSL Tags

The following table shows the correspondence between Switchboard-DAMSL (SWBD-

DAMSL) dialog tags and those used to label Meeting Recorder DA (MRDA) data. The

tags within the table are ordered according to the categorical structure within the

SWBD-DAMSL manual, with tags unique to the MRDA tagset being inserted in

accordance with this categorical structure. The SWBD-DAMSL categories are not

explicitly marked within this table in order to avoid confusion with the categories of the

MRDA tagset.

Tags listed in italics are based upon SWBD-DAMSL tags but have had their meanings

altered for the purposes of the MRDA data. Tags in boldface are not in the original

SWBD-DAMSL manual but have been added to accurately characterize the MRDA

data. Tag titles in boldface correspond to names of MRDA tags. All other tag titles

correspond to names of SWBD-DAMSL tags.

Additionally, the reasoning behind why certain SWBD-DAMSL tags are not used in the

MRDA tagset is found in Appendix 2. Explanations regarding the presence of tags

unique to the MRDA tagset are found in Appendix 3.

TAG TITLE SWBD-DAMSL MRDA

Uninterpretable % %

bandoned %- %--

Interruption not marked %-

Nonspeech x x

Self-talk t1 t1

3rd-party-talk t3 t3

bout-task t t

bout-communication c not marked

Statement-non-opinion sd s

Statement-opinion sv s

Open-option oo not marked

Yes-No-question qy qy

Wh-Question qw qw

Open-Question qo qo

Or-Question qr qr

Or-Clause qrr qrr

Rhetorical-Question qh qh

Declarative-Question d d

Tag-Question g g

ction-directive ad co

Offer co cs

Commit cc cc

Conventional-opening fp not marked

Conventional-closing fc not marked

Explicit-performative fx not marked

Exclamation fe fe

Other-forward-function fo not marked

Thanks ft ft

Welcome fw fw

pology fa fa

Topic Change not marked tc

Floor Holder not marked fh

Floor Grabber not marked fg

ccept aa aa

ccept-part aap aap

Maybe am am

Reject-part arp arp

Reject ar ar

Hold before

answer/agreement h h

Signal-non-understanding br br

Continuer b b

Rhetorical-question

continuer bh bh

cknowledge-answer bk bk

Mimic other m m

Repeat not marked r

Collaborative completion 2 2

Reformulate/summarize bf bs

ssessment/appreciation ba ba

Sympathy by by

Downplayer bd bd

Correct-misspeaking bc bc

Misspeak Self-Correction not marked bsc

Understanding Check not marked bu

Defending/Explanation not marked df

"Follow Me" not marked f

Yes answers ny aa

No answers nn ar

ffirmative non-yes answers

na na

Negative non-no answers ng ng

Other answers no no

Expansions of y/n answers e e

Dispreferred answers nd nd

Quoted Material q not marked

Hedge h not marked

Continued from previous line

+ not marked

Humorous Material not marked j

Rising Tone not marked rt

Nonlabeled not marked z

1.3 Meeting Recorder DA (MRDA) Tagset

The categorization scheme for the Meeting Recorder DA (MRDA) tagset differs from the

scheme employed for the SWBD-DAMSL tags seen. The reasoning behind this is that,

in the process of adjusting the definitions of previously established SWBD-DAMSL tags

and creating new tags to assist in adequately assessing the MRDA data, the resulting

MRDA tagset could not be appropriately characterized when placed in direct relation to

the SWBD-DAMSL tagset, given the nature of the data for which the MRDA tagset was

employed. Consequently, the tags are not organized on a dimensional level, but rather

the correspondences for the MRDA tagset are listed on the tag level. Descriptions of

the individual tags within the MRDA tagset are found in Section 5.

Group 1: Statements

s Statement

Group 2: Questions

qy Y/N Question

qw Wh-Question

qr Or Question

qrr Or Clause After Y/N Question

qo Open-ended Question

qh Rhetorical Question

Group 3: Floor Mechanisms

fg Floor Grabber

fh Floor Holder

h Hold

Group 4: Backchannels and Acknowledgements

b Backchannel

bk Acknowledgement

ba Assessment/Appreciation

bh Rhetorical Question Backchannel

Group 5: Responses

Positive

aa Accept

aap Partial Accept

na Affirmative Answer

Negative

ar Reject

arp Partial Reject

nd Dispreferred Answer

ng Negative Answer

Uncertain

am Maybe

no No Knowledge

Group 6: Action Motivators

co Command

cs Suggestion

cc Commitment

Group 7: Checks

f "Follow Me"

br Repetition Request

bu Understanding Check

Group 8: Restated Information

Repetition

r Repeat

m Mimic

bs Summary

Correction

bc Correct Misspeaking

bsc Self-Correct Misspeaking

Group 9: Supportive Functions

df Defending/Explanation

e Elaboration

2 Collaborative Completion

Group 10: Politeness Mechanisms

bd Downplayer

by Sympathy

fa Apology

ft Thanks

fw Welcome

Group 11: Further Descriptions

fe Exclamation

t About-Task

tc Topic Change

j Joke

t1 Self Talk

t3 Third Party Talk

d Declarative Question

g Tag Question

rt Rising Tone

Group 12: Disruption Forms

% Indecipherable

%- Interrupted

%-- Abandoned

x Nonspeech

Group 13: Nonlabeled

z Nonlabeled

SECTION 2: SEGMENTATION

Utterance segmentation is one of the most debated topics in discourse analysis. The

function of dialog must always be considered when determining utterance boundaries.

Lengthy utterances containing multiple conjunctions, speaker rambling, and floor-

holding are just a few factors complicating the decisions regarding utterance

boundaries. In order to segment transcribed speech into distinguishable utterances, the

following factors are taken into consideration within the context of the conversation:

syntax, pragmatic function, and prosody.

Prior to determining how to segment transcribed speech, knowledge of how utterance

boundaries are marked within the transcript is necessary. There are two ways to mark

utterance boundaries within the transcript. When a speaker trails off or is interrupted

and consequently does not complete his utterance, an utterance boundary in the form of

<==> is marked at the end of the corresponding utterance in the transcript. In Example

1 on the following page, speaker c2 does not finish his utterance (speaker c3 adds the

remainder of c2's utterance shortly after) and an utterance boundary is signaled by the

<==> in the transcript. If a speaker's utterance is complete, an utterance boundary in

the form of < . > is marked at the end of the corresponding utterance in the transcript.

Returning to the factors involved in segmentation, in terms of syntax, utterance

boundaries are primarily derived on a phrasal level. This is not to say that an utterance

consists only of a noun phrase or a verb phrase, but rather that it is permitted for a

complete utterance to consist only of a noun phrase, a verb phrase, or both. In

Example 11, the noun phrase "jose" constitutes a complete utterance:

Example 2 and 3 depict instances where verb phrases, "got it" and "wants to conserve"

in Example 2 and "confused" in Example 3, behave as complete utterances:

1 Examples take a format in which the numerical values of the first column represent start and end times

of utterances, the second column indicates the channel, the third indicates the DA, and the fourth

presents the transcript.

xample 1: Bmr010

280.000-284.762 c2 s.%-- and i did some training on - on one

dialogue which was transcribed by ==

284.762-288.568 c2 s yeah we - we did a nons- - s- -

speech nonspeech transcription .

287.474-288.294 c3 s^2 jose.

The pragmatic function of an utterance is also an important consideration for utterance

boundary identification. Phrases or clauses that do not appear complete grammatically

may actually form complete utterances on account of having unique functions within

conversation. Although it may seem peculiar to segment utterances on a phrasal and

clausal level, such a method of segmentation is utilized for the purpose of maximizing

the amount of information derived from DAs.

Example 4 presents an utterance that appears complete grammatically, yet does not

maximize the amount of information which can be derived from DAs.

In Example 5, the same utterance from Example 4 is shown, however the utterance is

segmented at the clausal level so that more information may be provided by the DAs

that otherwise would not be present had the utterance not been segmented.

xample 2: Bed011

114.007-116.680 c2 s and um - i - i told it to stay on forever

and ever .

116.680-119.347 c2 s but if it's not plugged in it just doesn't

obey my commands .

119.120-119.320 c1 s^bk okay .

119.726-120.386 c2 s it has a mind .

121.961-122.331 c1 s^bk got it .

122.160-123.170 c4 s wants to conserve .

Example 3: Bed003

2950.850-2957.110 c3 s yeah the only like - possible

interpretation is that they are - like -

come here just to rob the museum or

something to that effect .

2952.260-2953.830 c2 s^2 confused .

xample 4: Bmr010

217.921-227.363 c6 s^cs that uh - if we had something that

worked for many cases before maybe

starting from there a little bit because

ultimately we're going to end up with

some s- - kind of structure like that.

Syntax and pragmatic function are both taken into account when encountering

conjunctions. Conjunctions such as "and," "or," "but," and "so" often behave as cues to

locations where a string of clauses might be segmented into separate utterances.

Rather than simply start a new utterance, a speaker might use one of these

conjunctions as a connection between two complete utterances, as seen in a pre-

segmented utterance in Example 6:

Example 7 depicts a correctly segmented version of Example 6:

Caution must be taken not to segment utterances upon the appearance of conjunctions

in every instance. Quite often, conjunctions are used to simply connect noun phrases

or verb phrases that would not constitute separate utterances in the context in which

they are used. In these cases, the utterance is not segmented at the conjunction.

xample 5: Bmr010

217.921-222.161 c6 s^cs that uh - if we had something that

worked for many cases before maybe

starting from there a little bit .

222.161-227.363 c6 s^df because ultimately we're going to end

up with some s- - kind of structure like

that.

xample 7: Bmr020

595.187-596.880 c6 s that's somewhat - that's somewhat

subject to error .

596.880-601.180 c6 s but still we - we uh don did some ha- -

hand checking .

601.310-604.837 c6 s and - and we think that - based on that

we think that the results are you know

valid .

604.837-608.363 c6 s although of course some error is going

to be in there .

xample 6:

mr020

595.187-608.363 c6 s that's somewhat - that's somewhat

subject to error but still we - we uh don

did some ha- - hand checking and –

and we think that - based on that we

hink that the results are you know valid

although of course some error is going

to be in there .

Example 8 and Example 9 demonstrate instances when an utterance is not segmented

upon the appearance of a conjunction:

On occasion, a speaker may have an extremely lengthy utterance with many

conjunctive clauses and parentheticals. In such situations, each clause or parenthetical

is segmented into a separate utterance. As with segmenting on a clausal or phrasal

level, segmenting parentheticals in such a way allows for the maximization of

information provided by DAs. In deciding how to segment such instances within

transcribed speech, it is helpful to determine whether a speaker actually had the whole

string of speech in mind or else unintentionally diverged from his original thoughts.

Example 10 depicts a rather lengthy utterance prior to segmentation and Example 11

presents a segmented version of the same utterance.

xample 8: Bro014

238.387-240.098 c2 s^e i mean it's like

one little text file you edit

and change those numbers .

Example 9: Bro014

302.417-305.275 c2 s now h t k's compiled for both the linux

and for um the sparcs .

xample 10: Bmr005

1012.960-1033.300 c4 s but i - i mean - i think also to some

extent its just educating the human

subjects people in a way because

there's if uh - you know - there's court

transcripts there's - there's transcripts

of radio shows i mean - people say

people's names all the time so i think

it - it can't be bad to say people's

names it's just that i mean - you're

right that there's more poten- - if we

never say anybody's name then there's

no chance of - of - of slandering

anybody .

Example 11: Bmr005

1012.960-1019.350 c4 s but i - i mean - i think also to some

extent its just educating the human

subjects people in a way .

Prosody is also of considerable importance in detecting utterance boundaries. To take

the prosody of an utterance into consideration is to take the aural cues such as the rise

and fall of pitch, the energy level, and duration of the words of the utterance as well as

the complete utterance into consideration. Utterances that appear complete

syntactically, whether they are quite lengthy or consist of short phrases or clauses, may

be incomplete prosodically. If the prosody of the end of an utterance consists of a pitch,

energy level, or duration that is incongruent with that of a complete utterance, then that

particular utterance is considered incomplete. General prosodic patters found within

complete utterances and prosodic patterns specific to certain speakers are necessary

factors in determining how to assess the prosody of a complete utterance.

Prosody is of use in determining whether an utterance is interrupted or abandoned. If a

speaker begins trailing off in pitch and the energy level begins to decrease, the

speaker's utterance is most likely to be marked as abandoned. Prosody can also help

distinguish between floor grabbers and backchannels, as floor grabbers tend to have a

higher energy level in contrast to the surrounding speech and backchannels do not.

Pauses also behave as signifiers to utterance boundaries. Oftentimes, the appearance

of a lengthy pause indicates that the segment of speech following the pause constitutes

a new utterance. If the portion of speech immediately preceding the pause is

incomplete, that portion may either be an abandoned utterance or the beginning of an

utterance of which the portion of speech following the pause is the end. If the former

applies, and the portion preceding the pause is actually abandoned, a change in DAs,

prosody, or both is an obvious signal that the pause is indicative of a boundary.

However, if the latter case is applicable, no such drastic change in the prosody between

the segment preceding and the segment following the pause will be present and both

portions of speech are to comprise one utterance. To reiterate with regard to the latter

case, an utterance boundary will not be marked at the pause. As a side note, it must be

mentioned that some speakers tend to speak slowly in such a manner that their

utterances are filled with frequent pauses. In such instances, pauses are not indicators

of utterance boundaries unless the segment of speech following a pause is incongruent

with the segment preceding.

1019.350-1025.740 c4 s^df because there's if uh - you know -

there's court transcripts there's -

there's transcripts of radio shows i

mean - people say people's names all

the time .

1026.390-1028.940 c4 s so i think it - it can't be bad to say

people's names .

1029.270-1033.300 c4 s^df it's just that i mean - you're right that

there's more poten- - if we never say

anybody's name then there's no

chance of - of - of slandering

anybody .

As difficulty in determining utterance boundaries is encountered when considering the

factors of syntax, prosodic function, prosody, and pauses, additional segmentation

issues occasionally arise with the applicability of certain tags, namely <fg>, <fh>, <h>,

<aa>, <ar>, <bk>, and <g>. Regarding <fg>, <fh>, and <h>, often the problem at hand

is whether to segment an utterance in which a speaker utters a string of <fg>s, <fh>s, or

<h>s, as seen in Example 12. If there exist significant pauses between each portion of

the string of <fg>s, <fh>s, or <h>s, the utterance is segmented upon each pause and

each resulting utterance is labeled appropriately as <fg>, <fh>, or <h>, depending upon

its nature. However, if no such significant pauses exist, then the entire utterance

remains intact and receives a suitable label. Additionally, it is far more difficult to judge

if a pause actually signifies an utterance boundary within strings of <fg>s, <fh>s, or

<h>s than within strings of fluent speech.

As a general convention, unless an utterance is comprised solely of floor holders, it is

not to end with a floor holder <fh>. In the case that a floor holder is found at the end of

an utterance, it is split from the utterance and either receives its own line or is merged

with the following utterance of the same speaker, depending primarily upon its prosody

and its temporal proximity to the following utterance. If the length of the floor holder is

incongruent to the length of the words of the following utterance, the floor holder is of a

different intonation in relation to the following utterance, or a significant pause exists

between the floor holder and the following utterance, the floor holder is not merged with

the following utterance. If the floor holder is merged with the following utterance and the

following utterance is not a floor holder, then it is permissible for the resulting utterance,

which consists of a floor holder and another DA, to contain multiple DAs. Additionally,

although a floor grabber and a hold do not occur mid-speech as a floor holder does,

these tags may also be merged with the following utterance if deemed necessary and

the resulting utterance will also contain multiple DAs. Section 3.3 specifies the manner

in which utterances with multiple DAs are treated.

After splitting a floor holder from an utterance, it must be decided whether the portion

which originally preceded the floor holder is complete or incomplete. Example 13

depicts an utterance ending with a floor holder and the same utterance is seen in

Example 14 with the exception that the utterance has been segmented so that the floor

holder receives its own line.

xample 12: Bmr012

1886.800-1891.3100 c1 s^cs and then just sort of have that as the -

and then you can have groups of

twenty people or whatever .

1891.310-1892.080 c1 fh and - and uh ==

Regarding the tags <aa>, <ar>, <bk>, and <g>, the largest problem is determining

whether or not an utterance boundary exists after speech labeled with the tag <aa>,

<ar>, or <bk>, that is if speech from the same speaker immediately follows, or if a

boundary exists before speech labeled with the tag <g>, that is if speech from the same

speaker immediately precedes the portion labeled with the tag <g>. This problem only

emerges if the speech surrounding the portions labeled with the tags previously

specified is such that the prosody bears no indication of a boundary between

utterances, the speaker speaks so quickly that a boundary cannot be discerned, or else

no significant pause is found to mark a boundary. When the issue arises that a

boundary cannot be marked between speech labeled with the previously mentioned

tags and the surrounding speech, then it is permissible for an utterance to have multiple

DAs. Section 3.3 details the format of labels for utterances which have multiple DAs.

Another issue regarding segmentation concerns otherwise complete utterances being

segmented in such a way that yields abandoned utterances. For instance, a complete

utterance may be quite lengthy and appear as though it ought to be segmented.

However, segmenting the utterance may yield incomplete utterances that would be

marked as abandoned. As the original intact utterance is complete and some of the

segmented portions are marked as being abandoned, it is clear that segmenting the

utterance in a way that yields abandoned utterances is incorrect.

As an addendum to the aforementioned system of segmentation, if uncertainty exists as

to whether or not to segment an utterance, a general guideline is to segment the

utterance regardless. Also, portions of speech that constitute one utterance but for

some reason, perhaps mistakenly, are segmented as multiple utterances are merged to

form one utterance.

xample 13: Bmr010

601.519-604.014 c0 s and if it's good enough we'll arrange

windows machines to be available

so ==

Example 14: Bmr010

601.519-602.707 c0 s and if it's good enough we'll arrange

windows machines to be available .

603.465-604.014 c0 fh so ==

SECTION 3: HOW TO LABEL

3.1 Basic Format of DAs and Labels

The basic format of a DA is as follows2:

The basic format of a label is as follows (depending upon the utterance, the portions

enclosed in brackets may or may not be necessary):

3.2 Label Construction

The general tag is a mandatory component of every label. Only one general tag is

present in each DA. Specific tags and disruption forms (which indicate when a speaker

has been interrupted, trails off, or else is indecipherable) are included within a label only

when an utterance cannot be sufficiently characterized by a general tag and when

further characterization is needed. Specific tags are appended to general tags when

necessary and are not used alone. For the purpose of uniformity among annotators,

when multiple specific tags are appended to a general tag, they are attached in

alphabetical order3.

In the following sets of tags, the first set contains general tags, the second set contains

specific tags, and the third set contains disruption forms. Detailed descriptions of the

tags in the three sets can be found in Section 5. Note that the tags found in Set 1 are

2 Throughout this manual, when discussing format, the convention of enclosing portions in brackets

denotes that, depending upon an utterance, those portions may or may not be necessary.

3 As specific tags are attached in alphabetical order, the tag <2> is the last tag within the alphabetically

ordered hierarchy, rather than the first.

<general tag> [ ^ specific tag ]

<general tag> [ [ ^ <specific tag> ] [ | <general tag> [ ^ <specific tag> ] ] [ . <disruption form> ] ]

only used as general tags, the tags found in Set 2 are only used as specific tags (in

conjunction with a general tag), and tags in Set 3 are only used as disruption forms.

Set 1: General Tags

s qy qw qr qrr qo qh b fg fh h

Set 2: Specific Tags

aa aap am ar arp ba bc bd bh bk br

bs bsc bu by cc co cs d df e f

fa fe ft fw g j m na nd ng no

r rt t tc t1 t3 2

Set 3: Disruption Forms

Disruptions

%- %--

Indecipherable

x %

Within a DA, when specific tags are necessary, they are attached to the general tag with

a caret (^), thus rendering the following depiction of a DA:

Disruption forms are attached to and separated from the end of a DA with a period < . >,

as seen in the following representation:

< general tag >^< specific tag 1 >^< specific tag 2 >^< specific tag 3 > ...^< specific tag n >

< general tag > [ ^ < specific tag 1 > ...^< specific tag n >] . < disruption form >

It must be noted that, in some cases, a disruption form is present within an utterance

without sufficient information to assign a DA to that utterance. In such instances, a label

comprised solely of a disruption form is necessary.

Additionally, if for some reason an utterance is not to be labeled with a DA, then that

particular utterance receives a label consisting only of the tag <z>. For instance, if an

utterance contains data that is not to be labeled on account of it containing digits,

containing pre- or post-meeting chatter, pertaining to a "bleeped" portion in the

corresponding audio file, or else is simply not relevant to the labeling task, a label

comprised solely of the tag <z> is used. As the tag <z> is used to mark utterances

which otherwise would be labeled with DAs but instead are intentionally not to be

labeled, it is clear why the tag <z> is not included within the other groups of tags (i.e.

general tags, specific tags, and disruption forms). The tag <z> does not provide any

information regarding the characteristics and functions of utterances as the tags of the

other groups do, and for this reason it is separated from those groups.

The following is a partial list of sample labels that are acceptable within the previously

established conventions for label construction:

s qy qr b fg %

s^bk qy^d^f^g^rt qr^rt b.% fh^rt %-

s^nd qy^bh qrr.%-- b.x h %--

s^aa^rt.%-- qy^bu.%- qh^rt.% b^rt z x

Listed below is an incomplete list of sample labels that are not acceptable within the

previously established conventions for label construction:

s^s aa^bk x.%-- %--.s^qy^d s^z

s^s^aa %.%-- %--.x b.%- z.%--

It is worthy of mention that other restrictions apply in constructing labels. Such

restrictions include particular specific tags which may only appear with certain general

tags, particular general tags which have a limited set of applicable specific tags, and

sets of specific tags which are prohibited from appearing in the same DA. Restrictions

applying to the usage of tags are discussed in the individual tag descriptions in Section

3.3 Annotating Utterances Containing Multiple DAs

In cases where one DA does not suffice to represent an utterance, two DAs are used.

Such a need arises in cases as those described in Section 2, usually with tags such as

<fg>, <fh>, <h>, <aa>, <ar>, <bk>, and <g> which correspond to short utterances.

Often, an utterance requires multiple DAs when a floor grabber <fg> or floor holder <fh>

is uttered at the beginning of a statement <s> or question, when a short answer of the

nature <aa>, <ar>, or <bk> is following by a longer explanation, or when a statement is

followed by a tag question <g>. In some cases, an utterance requires multiple DAs

when a statement <s> is followed by a short answer of the nature <aa>, <ar>, or <bk>.

In which case, the DAs can be separated in both the label and the portion of the

transcript containing the utterance with a pipe bar < | >.

The pipe bar < | > is only used when sequential portions of an utterance that operate

closely together require different characterizations. For instance, a pipe bar is not used

for an agreement <aa> and a question that immediately follows it. In fact, an agreement

followed by a question does not constitute an utterance but constitutes two separate

utterances instead. Rather, an agreement immediately followed by an explanation of

the agreement, a longer, narrative form of agreement, or a direct reference to what the

agreement regards would require a pipe bar so long as the prosody and lack of

significant pauses warrants such usage of a pipe bar.

The use of a pipe bar indicates that segmenting an utterance is not necessary, despite

that the initial portion of an utterance, or last portion in the case of <g>, has a different

DA than the rest of the utterance.

The pipe bar is indicated in the appropriate location within the label as well as within the

transcription. Within the label, the pipe bar separates the DAs. Within the transcript,

the pipe bar separates the portions of an utterance to which the different DAs apply.

This is done in such a manner that the DA to the left of the pipe bar in the label pertains

to the portion of the utterance to the left of the pipe bar in the transcript and the DA to

the right of the pipe bar in the label pertains to the portion of the utterance to the right of

the pipe bar in the transcript.

Example 1 demonstrates the correct usage of a pipe bar, whereas Example 2 and

Example 3 depict the incorrect usage of a pipe bar.

xample 1: Bmr012

94.861-99.771 c4 fg|s^t um - | everyone should have at least

two forms possibly three in front of you

depending on who you are .

3.4 Disruption Forms

Disruption forms are used to mark utterances that are indecipherable, abandoned, or

interrupted. Only one disruption form may be used per utterance.

Disruption forms are included in a label in one of three formats, depending upon the

nature of an utterance. When a DA is not detected, a disruption form alone may

comprise an entire label. When used in conjunction with a DA, disruption forms are

marked using either a period < . > or a pipe bar < | >.

If an utterance contains a disruption form and is too short to determine which DA

applies to it, then only the disruption form is marked in the label. An utterance that is

indecipherable may actually be quite lengthy, but because it cannot be deciphered, an

appropriate DA cannot be assigned to it and only the disruption form is marked.

Example 4 depicts a disrupted utterance which contains insufficient information to

provide a DA:

Exceptions occasionally apply to short utterances deemed indecipherable. Utterances

which appear to be backchannels, for instance, yet are indecipherable may be labeled

with the appropriate DA along with a period and the applicable disruption form. Such

treatment of indecipherable utterances is only employed when there is a high probability

that the specific DA applies to the utterance based upon the surrounding context of the

short utterance and the speaker's speech patterns. The following are two sample labels

pertaining to short indecipherable utterances:

xample 4: Bro014

1207.310-1207.880 c1 %- but i- ==

xample 2: Bmr0

94.861-99.771 c4 s^t|fg um - | everyone should have at least

two forms possibly three in front of you

depending on who you are .

Example 3: Bmr012

94.861-99.771 c4 fg|s^t um - everyone | should have at least

two forms possibly three in front of you

depending on who you are .

b.% b.x

A period or a pipe bar is used in conjunction with a disruption form if a disruption form is

indeed applicable to an utterance and if an utterance contains sufficient information to

assign to it a DA. For instance, if an utterance, such as a statement, is interrupted or

abandoned, the DA is marked and then followed by a period and the appropriate

disruption form, as seen in Example 5:

In the case of Example 5, the utterance contains sufficient information to determine that

it is indeed a statement, despite being abandoned. If an utterance does not contain

adequate information to decide which DA applies to it, then a DA is not marked.

Two types of instances exist in which an utterance containing a pipe bar requires a

disruption form. In the first, an utterance requiring a pipe bar, such as what is discussed

in Section 3.3, is either abandoned or incomplete. To the left of the pipe bar is a DA

containing a tag such as <fg> or <aa> and to the right is a statement or explanation of

some sort that is either incomplete or abandoned. Note that the disruption form only

applies to the DA to the right of the pipe bar. Keeping in mind that the portion of the

utterance to the right of the pipe bar contains sufficient information to assign to it a DA

and is also abandoned or incomplete, its DA is followed by a period and the appropriate

disruption form, as seen in Example 6:

In the second instance in which an utterance containing a pipe bar requires a disruption

form, the portion of the utterance to the right of the pipe bar does not contain sufficient

information to assign to it a DA. This portion may be abandoned, interrupted, or

indecipherable. The DA designated to the portion of the utterance to the left of the pipe

bar clearly begins upon the onset of the utterance and ends at the point where the pipe

bar is placed. The DA pertaining to the initial portion of the utterance is marked, a pipe

bar is placed after the DA in the label and at the point where that particular DA ends in

the transcript, and a disruption form is marked after the pipe bar, as seen in Example 7

and Example 8:

xample 5: Bro014

495.681-499.134 c4 s.%-- some people are arguing that it would

be better to have weights on ==

xample 6: Bro014

1897.760-1904.500 c0 s^bk|s.%-- yeah | hopefully i think what we want to

have is to put these features in s- -

some kind of ==

The distinction between the use of the pipe bar and a period exists in how an utterance

can be divided. An utterance divided by a pipe bar behaves in some ways as two

separate utterances. The segment of the utterance to the left of the pipe bar will be

annotated with a particular DA that is different from the DA used to annotate the right,

that is if it is possible to assign a DA. The pipe bar exists as a clear boundary which

marks where one DA ends and another begins in a single utterance. The portion to the

right of the pipe bar behaves as a separate utterance in that it alone is the specific

segment which is interrupted, abandoned, or indecipherable. The portion to the left is

complete.

With regard to periods, and even labels consisting solely of disruption forms, no clear

and comparable boundary as found in utterances requiring pipe bars exists. The exact

region within an utterance where the disruption form occurs does not behave as a

separate segment of the utterance that can be marked clearly with a mechanism such

as a pipe bar. It is also unnecessary to use a pipe bar to mark where an interruption

begins or where a speaker abandons his utterance, since the DA to the left of the pipe

bar may also apply to the other side where the disruption form is marked.

Additionally, the reasoning behind why a disruption form is not used as a tag within a

DA is that the tags used within a DA apply primarily to the function of an entire

utterance. Disruption forms, however, usually apply only to the end of the utterance.

For this reason, the use of periods with disruption forms is deemed necessary.

3.5 Quotes

Utterances that contain quoted material are to end with punctuation that reflects the DA

of the utterance overall. If a quoted question is embedded within a statement, a period,

rather than a question mark, is used at the end of the utterance in the transcript and no

other punctuation is used.

A colon in the label signifies that there is quoted material in the transcription. The DA to

the left of the colon characterizes the function of the entire utterance and the DA to the

right of the colon characterizes only the quote. If the quoted material only consists of a

xample 7: Bmr028

1187.370-1188.240 c1 fg|%- yeah | he ==

Example 8: Bro014

403.710-405.428 c2 s^aa|%-- yeah | it's uh ==

few words, such as a noun phrase, DA annotation of the quotation is unnecessary.

Example 9 demonstrates the manner with which quotes are handled:

3.6 Using TableTrans (Annotation Interface)

A. The Interface

There are three sections of TableTrans: the labeling and transcription section located at

the top, the time-segmented transcription located in the middle, and the waveform

located at the bottom.

xample 9: Bmr026

941.984-944.924 c1 s^cs:qw and just say an e- - just ask him that

you know wha- - what should you do .

945.464-947.864 c1 s:qy and in my answer back was are you

sure you just want one .

In the labeling and transcription section, the first and second columns on the left provide

the start and end times for each utterance and the third column denotes the speaker or

channel number. DA and adjacency pair (AP) labels are entered in the fourth and fifth

columns. The comment field is located in the sixth column and is primarily for an

annotator's notes regarding an utterance. The last column on the right, under the

"Trans" heading, provides the transcript of the utterances.

In order to label a meeting, the "Open Annotation File" command must be selected from

the "File" menu. A sub-menu will appear providing three formats that can be used.

"Table Format" is the format that is most widely used. A window will appear with a

"Feature List" and a "Delimiter" to which clicking the "OK" button is necessary. Shortly

after, the segment of the meeting to be annotated will appear.

Although the data within the fourth, fifth, sixth, and seventh columns may be altered

within the interface, the Time-Segmented section, which is the first two columns and

shows the annotator a series of utterances in chronological order, and the third column

denoting the speaker cannot be modified.

B. TableTrans Commands

COMMAND ACTION

Changing the Transcript

Ctrl-s Splits the current row at the location of the cursor in the

TRANS field.

Ctrl-m Merges the current row with the next row by the same speaker.

Moving within a Field

Ctrl-f or left-arrow Moves forward one character in a field.

Ctrl-b or right arrow Moves backward one character in a field.

Ctrl-p or up-arrow Moves up to previous row.

Ctrl-n or down-arrow Moves down to next row.

Shift + left-arrow Moves to previous field in the same row.

Shift + right-arrow Moves to next field in the same row.

right-click (In the Time-Segmented Transcription window) Opens up

Comment Field Window

Ctrl-1 Plays a segment

Ctrl-a Moves cursor to the beginning of a field

Ctrl-e Moves cursor to the end of a field

C. Printing Commands

Annotators can print out their comments using the program "csvcomment." The

command "csvcomment <csv_file>" is entered in the terminal window, where <csv_file>

is the name of the ".csv" file to print.

D. Playing the Sound File

To open up the wave file of a meeting to be labeled, a link command can be made from

the location where the sound file is saved in the annotator's home directory. After

returning to the TableTrans interface, "Open Sound File" is selected from the "File"

menu. The file can then be opened after browsing through the annotator's home

directory.

SECTION 4: ADJACENCY PAIRS

4.1 Purpose and Definition

Labeling adjacency pairs (AP) in meetings provides a means to extract the information

provided by the interaction between speakers. Adjacency pairs reflect the structure of

conversation and are paired utterances such as question-answer, greeting-greeting,

offer-acceptance, and apology-downplay. (Levinson 1983)

APs are defined as sequences of two utterances that are:

1. produced by different speakers

2. ordered with a first part (marked with “a”) and a second part (marked with “b”)

(Levinson 1983)

An example of an AP is shown below:

In Example 1, the utterances depict direct interaction between the two speakers.

4.2 Labeling Adjacency Pairs

Adjacency pairs consist of two parts, where each part is produced by a different

speaker. The basic form of an AP is seen below:

This format allows APs to be enumerated as: 1a, 1b, 2a, 2b, and so on. A different

number is assigned for each AP, yet every AP will contain an "a" part and a "b" part. A

labeled AP is seen in Example 2:

xample 1: Bro016

113.976-116.502 c4 s^bu but you were looking at mel cepstrum .

116.883-117.850 c5 s^aa yes .

Although APs are to be marked sequentially in ascending order, it is possible that the

numerical value of an AP jumps ahead of the numerical value of the previous AP by

more than a value of one (e.g., an AP has a numerical value of 5 and the following AP

has a numerical value of 7 instead of 6). However, such is only permitted so long as the

sequential order of the APs is preserved and the numerical values are not repeated or

used cyclically for entirely different APs.

4.3 Labeling Conventions

Specific labeling conventions have been established when marking APs in instances in

which an utterance contains multiple AP parts, an AP part consists of multiple

utterances, multiple speakers pertain to the same AP part, and an AP is overlooked.

A. Multiple AP Parts per Utterance

If an utterance functions as a "b" part of one AP and an "a" part of another AP, then

both APs are marked with a period < . > separating the two APs, as seen below:

A portion of a conversation in which APs are labeled is seen in Example 3:

xample 2: Bmr023

312.382-314.770 c2 qy^rt 30a are you implying that it's

currently disorganized ?

314.770-318.470 c3 s^na 30b in my mind .

xample 3: Bro021

66.555-68.227 c2 s^rt 4a well the first thing maybe

is that the p- -

eurospeech paper is uh

accepted .

69.904-70.928 c2 fh um ==

70.928-71.952 c2 fh yeah .

Refer to Section D for details regarding the treatment of utterances requiring three AP

parts.

B. Continued AP Parts

A continued AP part is an AP part consisting of multiple utterances by the same

speaker. When a continued AP part arises, a plus sign <+> is placed at the end of the

AP. Example 5 depicts an instance where an AP part consists of multiple utterances:

Additionally, an utterance consisting of a tag question <g> is included within an AP part,

assuming the utterance containing the statement <s> preceding it is a portion of the AP

part. In which case, the utterance containing the tag question will receive the

appropriate number of plus signs when labeled with an AP.

72.059-74.710 c5 qw^rt 4b.5a this is - what - what do

you uh - what's in the

paper there ?

74.702-81.090 c2 s^rt 5b.6a so it's the paper that

describe basically the um

system that were

proposed for the aurora .

80.320-82.794 c5 qy^bu^d^rt 6b.7a the one that we s- - we

submitted the last round ?

82.614-83.700 c2 s^aa 7b.8a right yeah .

83.110-83.750 c5 s^bk 8b uhhuh .

xample 5: Bro016

1494.110-1499.560 c1 qy^rt 20a do you have something

simple in mind for - i

mean vocal tract length

normalization ?

1497.570-1501.320 c5 s^ar|s^nd 20b uh no | i hadn't - i hadn't

thought - it was - thought

too much about it really .

1501.320-1503.200 c5 s^df^nd 20b+ it just - something that

popped into my head just

now .

1503.200-1505.070 c5 s.%-- 20b++ and so i - i ==

1505.690-1509.900 c5 s^cs 20b+++ i mean you could maybe

use the ideas - a similar

idea to what they do in

vocal tract length

normalization .

If an utterance contains multiple APs, where one or both is a continued AP part, a

period < . > is inserted between the two APs to separate them (e.g., 5b++.6a+).

C. Multiple Speakers per AP Part

In some cases, an AP part consists of two or more speakers. This occurs most often

with the "b" part and quite rarely with the "a" part. When such an occurrence arises, the

corresponding AP number and AP part are marked. Then each speaker contributing to

the same AP part receives a numerical value based upon the order in which the

speakers make their utterances. So the first speaker to contribute to an AP part

receives a value of 1, the second a 2, and so on. A hyphen <-> followed by a speaker's

numerical value is then appended to the AP. The format of an AP consisting of multiple

speakers is seen below:

AP parts containing multiple speakers are seen in Example 5:

If, for instance, the speaker designated as c2 in Example 5 continued speaking so that a

continued AP part resulted, then his next utterance would be labeled as 9b-2+, the next

9b-2++, and so on as necessary. When continued AP parts occur within AP parts

consisting of multiple speakers, each speaker retains his designated numerical value

and plus signs <+> are appended after the numerical values as necessary.

Additionally, if an utterance contains multiple APs, where one or both is an AP part

consisting of multiple speakers, a period < . > is inserted between the two APs to

separate them (e.g., 5b-1.6a+, 1b-3+.2a).

D. Handling Overlooked APs

As stated in Section 4.2, APs are to be marked sequentially in ascending order.

Occasionally, an AP is overlooked. If marking an overlooked AP with the next

numerical value in sequence results in a non-sequential ordering of APs then an

additional convention is implemented to handle the overlooked AP.

xample 5: Btr001

150.780-152.664 c5 s^bu 9a parentheses meaning

uncertainty .

151.730-152.365 c3 s^aa 9b-1 yeah .

152.467-153.164 c2 s^aa 9b-2 uhhuh .

For instance, if a meeting is labeled with APs in sequence starting with a numerical

value of 1 and ending with a value of 50 and an overlooked AP exists between an AP

with a numerical value of 34 and an AP with a numerical value of 35, the overlooked AP

is not to receive a numerical value of 51. Instead, the AP receives a numerical value of

34 followed by an underscore <_> and the appropriate AP part. The AP part is followed

by a hyphen with a numerical value and plus signs when necessary. An overlooked AP

located between two APs has the following format:

If a number of overlooked APs exist in sequence, for instance if three APs exist

between APs 34 and 35, then a slight modification of the above convention is

necessary. The first overlooked AP receives an AP in the format detailed above. The

second overlooked AP receives an AP in the same format but with two underscore <_>

symbols instead of one. The third overlooked AP receives an AP in the same format

but with three underscore symbols and so on, thus yielding the following format:

E. Labeled Meeting Sample

Example 6 depicts the labeling conventions discussed in Sections A through C. What is

particularly unique about this example is that it contains an utterance requiring two “a”

parts. Additionally, this utterance requires a total of three AP parts – two “a” parts and

one “b” part – when utterances usually require at most two.

<AP number of previous AP>_<AP part>[ - <numerical value>][+1, +2, …+n]

<AP number of previous AP>_1, _2, …_n <AP part>[ - <numerical value>][+1, +2, …+n]

xample 6: Bmr003

1594.720-1595.830 c3 qy^d 47b.48a you've already - you've

already done some ?

1595.360-1596.610 c2 s.%- 48b-1 she - she's done one –

she's one ==

1595.400-1595.950 c4 s^aa|s^na 48b-2 yes | i have .

1595.570-1597.070 c0 s^na 48b-3.49a.50a she's - she's done about

half a meeting .

1596.530-1597.570 c3 s^bk 49b-1 oh- - oh i see .

1597.130-1597.510 c2 s^bk 49b-2 right .

1597.570-1597.840 c3 s^bk 49b-1+ o_k .

1597.840-1598.100 c3 s^ba 49b-1++ good .

1597.760-1597.990 c2 s^bk 49b-2+ right .

1598.170-1598.360 c0 qy^d^g^rt 50a+ right ?

This utterance requires a “b” part as it contains the response to an earlier utterance,

which constitutes the “a” part of the AP with a numerical value of 48. The “a” part of the

AP with a numerical value of 49 only consists of one utterance and receives a number

of responses. The utterance requires another “a” part for the AP with a numerical value

of 50 as this utterance, along with the speaker’s following two utterances, comprise the

“a” part for yet another AP.

F. Complex Form of an AP

The following is a complex form of an AP, taking into account the aforementioned

conventions:

4.4 Restrictions on Using Adjacency Pairs

Certain restrictions apply to which tags can or cannot be labeled with an AP.

APs denote direct interaction between speakers. Backchannels , which serve

simply to encourage the current speaker, are never marked with APs. Backchannels

are not uttered directly to a speaker as a response and do not function in a way that

elicits a response either. Rhetorical question backchannels <bh>, receive APs when

uttered as acknowledgments and do not receive APs when uttered as backchannels.

Floor holders <fh> and floor grabbers <fg> are also never marked with APs, since they,

like backchannels, are not said directly to anyone. Holds <h>, however, are marked.

The definition of a hold entails that a speaker is given the floor and is expected to speak

in response to something and "holds-off" prior to making an utterance. As the speaker

is expected to speak and then utters a hold, which is usually followed by a response,

the hold is considered part of the response.

Mimics <m> and collaborative completions <2> are always marked with APs, as they

are always in direct reference to another speaker's utterance.

1598.580-1598.950 c2 s.%- i'm go- ==

1598.580-1598.980 c0 qy^d^rt 50a++ about half ?

1599.150-1600.160 c4 s^no 50b.51a s- -

i'm not sure if it's that's

much .

<AP number>[ _1 , _2 , …_n ]<AP part>[-<numerical value>][ +1, +2, …+n ][ .<AP number> >[ _1 , _2 , …_n ]<AP part> … ]

When indecipherable utterances appear, if the utterance can be characterized with a

DA and it appears as though the utterance functions within an AP, then an AP is

marked accordingly. Otherwise, no AP is marked.

In some cases, it is quite difficult to determine to which utterance a response refers. If

such difficulty arises, then an AP is not marked. For instance, a scenario may arise

where two or three speakers utter statements <s> simultaneously and another speaker

utters an acknowledgment <bk>. As an acknowledgment by one speaker to another

speaker is usually marked with an AP, if it cannot be determined whom a speaker is

acknowledging, then an AP is not marked.

SECTION 5: TAG DESCRIPTIONS

5.1 Preliminaries

This section provides a detailed description of each tag and the rules governing the

usage of each tag. The tags are categorized into thirteen groups according to syntactic,

semantic, pragmatic, and functional similarities of the utterances they mark. Beneath a

group heading will be a general description of the group along with explanations of the

tags within the group. Most tag descriptions will contain examples4 from data to further

elucidate a tag's usage.

With regard to the examples provided within this section, it is of much use to listen to

the corresponding audio portions, as some examples cannot be fully comprehended

otherwise. In particular, utterances marked as floor grabbers <fg>, floor holders <fh>,

holds <h>, backchannels , acknowledgements <bk>, and accepts <aa> share a

common vocabulary which renders examples of these tags in text insufficient in fully

communicating how utterances marked as such are identified.

5.2 Group 1: Statements

This group contains only one tag, <s>, and serves as the default general tag.

 Statement <s>

The <s> tag is the most widely used tag in the MRDA tagset. Unless an utterance is

completely indecipherable or else can be further described by a general tag as being a

type of question, backchannel, floor grabber, floor holder, or hold, then its default status

as a statement remains.

When necessary, specific tags are appended to the <s> tag to further characterize

utterances. The use of the <s> tag is seen in Example 1 through Example 4:

4 In some examples, when displaying surrounding context, unnecessary lines, such as those which are

irrelevant to characterizing a particular tag within the tag descriptions, may be edited out. The content

of utterances within the examples remains unchanged.

5.3 Group 2: Questions

This group contains all general tags pertaining to questions. The tag description for

elaborations <e> provides instructions regarding the treatment of questions followed by

elaborations.

 Y/N Question <qy>

This tag marks utterances in the form of yes/no questions if and only if they have the

pragmatic force along with the syntactic and prosodic indications of a yes/no question

(i.e. subject-auxiliary inversion or question intonation). Essentially, an utterance is

considered a yes-no question if it sounds as if it elicits a yes or no answer. This is not

to say that all yes/no questions will receive yes or no answers. A question may be

asked in a yes/no manner, but the response it receives may not be a simple yes or no.

Regardless of the answer, the utterance is still considered a yes/no question.

Basic yes/no questions are seen in Example 5 through Example 8:

xample 1: Bro004

578.567-585.527 c3 s if we exclude english um - there is not

much difference with the data .

Example 2: Bed016

70.600-71.470 c5 s^ba it's a great story .

Example 3: Bro021

3201.960-3204.850 c1 s^bu so this changes the whole mapping for

every utterance .

Example 4: Bro021

3204.850-3205.490 c1 s^bk okay .

xample 5: Bro016

58.863-61.782 c4 qy^rt do you think that would be the case for

next week also ?

The tag <qy> is also used as the general tag for tag questions <g> (e.g., "Yeah?", "Isn't

it?", etc.) and rhetorical question backchannels <bh> (e.g., "Really?", "Isn't that

interesting?", etc.). Many declarative questions <d> are also in the form of yes/no

questions. Example 9 through Example 11 exhibit these characteristics:

Additionally, a convention has been established in handling instances when a yes/no

question is followed by an elaboration <e> which requires its own line. In such cases,

the following elaboration could be considered a declarative yes/no question <qy^d>.

Instead, the elaboration receives a DA of <s^e>, along with any other necessary specific

tags. An instance of a yes/no question followed by an elaboration is seen in Example

12:

xample 6: Bmr027

2049.340-2051.730 c5 qy^rt did i say that ?

Example 7: Bmr027

1836.000-1838.580 c4 qy^bu^rt didn't they want to do language

modeling on you know recognition

compatible transcripts ?

Example 8: Bmr012

6.805-17.875 c1 qy^rt is this channel one ?

xample 9: Bro016

513.765-514.316 c4 qy^d^g^rt right ?

Example 10: Bmr027

2016.230-2017.440 c5 qy^bh oh really ?

Example 11: Bmr027

514.316-514.867 c4 qy^bu^d^rt the insertion number is quite high ?

xample 12: Bro021

316.709-319.202 c5 qy^rt wasn't there some experiment you

were going to try ?

In some cases, it may be difficult to determine whether an utterance is a yes/no

question or an "or" question <qr>. The tag description for <qr> details how distinguish

between the two tags in certain scenarios.

 Wh-Question <qw>

Wh-questions are questions that require a specific answer. These usually contain "wh"

words such as the following: what, which, where, when, who, why, or how. However,

not all questions containing a "wh" word are considered wh-questions. The section on

open-ended questions <qo> elucidates this point. Wh-questions are shown in Example

13 and Example 14:

Declarative wh-questions often appear as wh-questions prior to wh-movement. An

instance in which a declarative wh-question is used is seen in Example 15.

In some cases, utterances that do not contain wh-words are labeled as wh-questions

because they function as wh-questions. Such an instance is seen in Example 16:

319.202-325.216 c5 s^e.%-- where you did something differently for

each um uh - i don't know whether it

was each mel band or each uh um f f t

bin or someth- ==

xample 13: Bmr012

62.153-64.053 c3 qw^r^t3 why didn't you get the same results and

the unadapted ?

Example 14: Bmr012

231.944-233.704 c2 qw^t3 i guess - what time do we have to

leave ?

xample 15: Bed003

2889.130-2890.200 c1 qw what's the technical term ?

2890.330-2890.750 c3 qw^d^rt for which ?

2891.010-2892.820 c1 s^rt for the uh - nodes that are observable .

In Example 16, the utterance functions as a wh-question, in that "hm?" is akin to "what?"

as a request for repetition. "Huh?", "excuse me?", and "pardon?" also appear as wh-

questions in that they can also function in the same manner as what is exemplified in

Example 16. Caution must be taken to distinguish whether such utterances are indeed

wh-questions or if they are floor grabbers, floor holders, holds, backchannels, yes/no

questions that are rhetorical question backchannels, or acknowledgments.

Declarative wh-questions that do not contain "wh" words are often confused with

declarative forms of other questions because they appear the same syntactically.

Despite this syntactic similarity, they differ functionally based upon the response that the

question seeks. In determining whether an utterance is a declarative wh-question that

does not contain a "wh" word, the surrounding context, in particular the response the

question generates, is crucial to note. Most often, declarative wh-questions that do not

contain "wh" words are requests for repetition, such as those seen in Example 17

through Example 19.

xample 16: Bmr012

61.563-61.713 c0 qw^br^t3 hm ?

xample 17: Bmr031

947.610-948.925 c8 s it's still yeah two or three d v ds .

948.925-950.240 c8 %- but ==

949.569-951.874 c2 fg|s yeah | not if you have to distribute the

video also .

949.941-950.878 c5 qw^br^d two or three ?

951.125-953.860 c8 s^df if you use both sides and the two layer

and all that .

Example 18: Bro003

3193.230-3198.820 c2 fh|s^cc and um | for the broader class nets

we're - we're going to increase that .

3198.820-3204.400 c2 s^df because the um the digits nets only

correspond to about twenty phonemes .

3205.460-3208.780 c2 fh so .

3207.200-3207.840 c8 qw^br^d^rt broader class ?

3208.780-3210.430 c2 h|s um | the broader - broader training

corpus nets .

Example 19: Bro003

3400.840-3402.950 c8 qw^br^d^rt and - and you're saying about the

 Or Question <qr>

"Or" questions offer the listener at least two answers or options from which to choose.

Section 2 and Section 3.3, which deal with segmentation and multiple DAs within an

utterance, are quite helpful in determining if a question is actually an "or" question or if it

is a yes/no question <qy> followed by an "or" clause after a yes/no question <qrr>.

Select "or" questions can be seen in Example 20 through Example 23:

In terms of the responses "or" questions receive, the obvious response is one in which a

speaker selects one of the options posed within the "or" question. Sometimes the "or"

question is interrupted and answered as if it is a yes/no question. In these cases, the

question is marked as an "or" question if it seems as if the speaker would have

continued the question in an "or" question format if he had not been interrupted. In

other instances, the speaker asking the question might abandon his utterance, and the

speaker answering the question may respond as if the question were a yes/no question

without having interrupted the question at all.

spanish ?

3403.290-3404.350 c4 s the spanish labels .

3405.000-3409.590 c4 s that was in different format .

xample 20: Bmr001

305.466-307.826 c0 qr^rt are we going to - i mean - is it going to

be over there or is it going to be in

there ?

Example 21: Bed003

1214.120-1215.140 c4 qr are you assuming that or not ?

Example 22: Bmr001

339.042-342.612 c1 qr^rt do we have like a cabinet on order or

do we just need to do that ?

Example 23: Bmr007

165.987-167.447 cB qr is this the same as the e mail or

different ?

If a speaker abandons a question that is seemingly an "or" question, it is actually a

rather cumbersome task determining whether the question is indeed an "or" question or

not. The point where the speaker abandons his question is of crucial importance. If the

speaker abandons while posing at least a second option or after having posed at least

two options, the question can be considered an "or" question. If the speaker abandons

after saying the word "or" and has not issued a second option, the question could either

be an abandoned "or" question or a yes/no question followed by an "or" clause, as

mentioned above. If the speaker abandons at the word "or" abruptly, the utterance is

most likely an "or" question. If the speaker trails off at the word "or" so that the word

"or" is lengthened and sounds reminiscent of a floor holder <fh>, the "or" is segmented

from the utterance or else separated by a pipe bar and is labeled as an abandoned "or"

clause after a yes/no question <qrr.%--> and the remainder of the utterance is labeled

as a yes/no question.

Example 24 through Example 31 depict instances of interrupted and abandoned "or"

questions:

xample 24: Bed011

2776.460-2779.490 c1 qr.%- is that roughly the equivalent of - of

what i've seen in english or is it ?==

Example 25: Bmr005

2018.090-2023.710 c5 qr.%- you know - did she miss some

overlaps or did she ?==

Example 26: Bmr007

369.570-372.515 cB qr.%- is this uh just raw counts or is it ==?

Example 27: Bmr013

1987.000-1989.000 c2 qr.%-- well - oh wa- - in terms of the

speakers or the conditions or the ?==

Example 28: Bmr013

2064.000-2069.000 c1 qr^rt.%-- do the transcribers actually start wi- -

with uh - transcribing new meetings or

are they ?==

Example 29: Bmr014

582.763-585.270 c8 qr.%-- has that started or is that ?==

If an utterance is suspected to be an "or" question but the speaker abandons or is

interrupted before saying "or" and has not posed a second option, the utterance cannot

be considered an "or" question since there is insufficient evidence to label it with the

<qr> tag.

Furthermore, even with the presence of the word "or" along with a second option, it may

be difficult to determine whether an utterance is an "or" question or a yes/no question,

wh-question, or an open-ended question. If the question is actually presenting two

specific options, the question is an "or" question. The question is not an "or" question if

it presents one option and ends with a clause such as "or something." If a question

ends with such a clause, the clause is not labeled separately with the tag <qrr>.

Example 32 through Example 34 show instances when questions that are seemingly

"or" questions are to be labeled as otherwise:

xam

le 30: Bmr001

944.512-945.412 c8 qr^rt.%-- per channel or ?==

Example 31: Bmr009

1748.000-1751.000 c2 qr.%-- and north midland like like - uh illinois

or ?==

xample 32: Bmr005

3550.080-3551.680 c2 qy^d^rt^2 lapel mikes or something ?

Example 33: Bmr006

2057.610-2061.670 c0 qw what if there was a door slam or

something ?

Example 34: Bmr010

425.800-429.800 c6 qy is there a - a transformation uh - like

principal components transformation or

something ?

 Or Clause After Y/N Question <qrr>

This tag marks when a speaker adds an "or" clause to a yes/no question. The previous

description of "or" questions <qr> in conjunction with Section 2 and Section 3.3, which

deal with segmentation and multiple DAs within an utterance, are also quite useful in

determining whether a segment is an "or" clause and how to treat it.

As with the description of the tag <qr>, utterances marked with <qrr> must actually be

posing some sort of option, rather than being a wh-question, for instance, preceded by

the word "or."

Oftentimes, "or" clauses following yes/no questions are abandoned or else interrupted

and the entire utterance consists of the word "or." In these cases, the label for such an

utterance contains the <qrr> tag along with the appropriate disruption form.

Example 35 through Example 39 display in context instances where the tag <qrr> is

used:

xample 35: Bed003

1867.670-1868.970 c1 qy^rt do you have the true source files ?

1868.970-1870.270 c1 qrr or just the class ?

Example 36: Bmr018

405.920-411.860 c1 qy^rt the - i guess the question on my mind

is do we wait for the transcribers to

adjust the marks for the whole meeting

before we give anything to i b m ?

411.860-413.440 c1 qrr or do we go ahead and send them a

sample ?

Example 37: Bmr001

2178.450-2179.950 c0 qy^d^rt so - is it - it's going to disk ?

2179.950-2180.340 c0 qrr.%-- or is this ?==

Example 38: Bmr018

2722.490-2727.000 c1 qr did they ever try going - going the

other direction from simpler task to

more complicated tasks ?

2727.000-2728.000 c1 qrr.%-- or ?==

 Open-ended Question <qo>

An open-ended question places few syntactic or semantic constraints on the form of the

answer it elicits. A question containing a "wh" word and consequently appearing to be a

wh-question <qw> may actually be an open-ended question instead. Additionally, a

question that is seemingly a yes/no question or an "or" question may actually be an

open-ended question. As a wh-question, a yes/no question, and an "or" question

require a specific answer, an open-ended question, as its name suggests, does not

seek a specific answer at all. Rather, an open-ended question is asked in a broad

sense.

Open-ended questions are seen in Example 40 through Example 48:

xample 40: Bmr007

112.365-116.868 c3 fh|qo^d^rt um | and anything else ?

Example 41: Bmr007

117.088-118.018 c3 qo^d nothing else ?

Example 42: Bmr007

92.862-98.798 c3 fh|qo^d^rt

m | and anything else anyone wants to

talk about ?

Example 43: Bmr013

654.000-657.000 c3 qo^rt d- e- - anybody do you have any -

anybody have any opinion about that ?

Example 44: Bmr026

2307.190-2309.690 c5 qo anybody have any intuitions or

suggestions ?

xample 39: Bro004

1922.810-1928.020 c1 qy so do you - are you - w- - did you

have something going on - on the side

with uh - or on - on this ?

1928.020-1928.130 c1 qrr.%- or ?==

 Rhetorical Question <qh>

The tag <qh> marks questions to which no answer is expected. Such questions are

used by the speaker for rhetorical effect; they are essentially statements formulated as

questions. Although rhetorical questions and rhetorical question backchannels <bh>

are similar, <bh> lacks semantic content, functions mostly as a continuer, and is not

used by a speaker who has the floor. Rhetorical questions are seen in Example 49

through Example 55:

xample 49: Bed011

2204.540-2206.420 c2 qh^rt i mean is this realistic ?

Example 50: Bmr005

3802.380-3802.680 c4 qh^aa why not ?

Example 51: Bmr005

525.596-530.188 c4 qh^cs so why don't you - you start with that ?

Example 52: Bmr009

2089.900-2090.800 c3 qh s- - i mean who cares ?

xample 45: Bmr007

1681.390-1683.180 c3 fg|qo but - | what - what do you think about

that ?

Example 46: Bmr014

2691.750-2693.090 c4 qo^j how about them energy crises ?

Example 47: Bmr007

100.580-102.340 c0 qo^t what about the um - your trip

yesterday ?

Example 48: Bed006

666.870-667.530 c2 qo^d questions ?

5.4 Group 3: Floor Mechanisms

This group contains all general tags pertaining to mechanisms of grabbing or

maintaining the floor. The only disruption forms that can be appended to tags within this

section are the indecipherable tag <%> and the nonspeech tag <x>. Additionally, no

specific tag may be appended to the tags denoted as floor mechanisms. Section 2 and

Section 3.3 detail the issues regarding segmentation with floor mechanisms.

 Floor Grabber <fg>

Floor grabbers usually mark instances in which a speaker has not been speaking and

wants to gain the floor so that he may commence speaking. They are often repeated by

the speaker to gain attention and are used by speakers to interrupt the current speaker

who has the floor. Most often, floor grabbers tend to occur at the beginning of a

speaker's turn.

In some cases, none of the speakers will have the floor, resulting in multiple speakers

vying for the floor and consequently using floor grabbers to attain it. During such

occurrences, many speakers talk over one another without actually having the floor.

Floor grabbers are also used to mark instances in which a speaker who has the floor

begins losing energy during his turn and then uses a floor grabber to either regain the

attention of his audience or else because it seems as though he is relinquishing the

floor, which he does not wish to do. Such mid-speech floor grabbers are usually

followed by a change in topic.

Floor grabbers are generally louder than the surrounding speech. Although the energy

of a floor grabber is relative to the energy of the surrounding speech, it is also relative to

the energy of a speaker's normal speech.

xample 53: Bmr009

2512.610-2513.290 c1 qh^ba isn't that wonderful ?

Example 54: Bmr009

2778.960-2779.800 c0 qh^co why don't you read the digits ?

Example 55: Bmr012

1414.430-1415.430 c1 fh|qh uh - | but who knows ?

Common floor grabbers include, but are not limited to, the following: "well," "and," "but,"

"so," "um," "uh," "I mean," "okay," and "yeah." It is worth mentioning that the

identification of floor grabbers is not merely based purely on the vocabulary used, but

rather on the speaker's actual attempt, whether successful or not, to gain the floor.

As previously mentioned, floor grabbers are not to be identified solely based upon the

vocabulary used, as floor grabbers, floor holders <fh>, holds <h>, backchannels ,

acknowledgements <bk>, and accepts <aa> share a very similar vocabulary. In order

to properly distinguish whether an utterance is performing as a floor grabber, floor

holder, hold, backchannel, acknowledgement, or accept, it is necessary to take into

account the details provided within the individual tag descriptions and to listen to the

audio portions corresponding to the examples within those tag descriptions. Utterances

labeled with these tags tend to appear very similar in text yet emerge exceedingly

different in sound.

As floor grabbers and backchannels are often confused on the basis of having a similar

vocabulary, they are actually quite distinct in sound. The main distinctions between the

two is that backchannels have a lower energy level in relation to the surrounding speech

and are not used by someone who has or is attempting to gain the floor. Also,

backchannels are considered "background" speech.

The floor grabbers seen in Example 56 through Example 60 are shown merely to

illustrate how they appear in text. The surrounding context has been omitted for each

example, as it provides little to no information regarding how to identify floor grabbers.

mple 56: Bed004

1017.990-1018.180 c4 fg but uh ==

Example 57: Bed004

1052.310-1052.620 c2 fg okay .

Example 58: Bed004

2264.780-2265.060 c2 fg yeah but ==

Example 59: Bmr012

1814.65-1817.01 c2 fg|s.%- well | or also for you know - if people

are not ==

Example 60: Bmr012

1822.12-1824.17 c4 fg|qy^df well i mean - | is the - is the

handheld really any better ?

 Floor Holder <fh>

A floor holder occurs mid-speech by a speaker who has the floor. A floor holder is

usually an utterance such as "uh" or "so" and is used as a means to pause and continue

holding the floor. In some cases, a speaker will utter a floor holder at the end of his turn

as a means to relinquish the floor.

The duration of a floor holder is usually longer than that of the other words spoken by a

speaker. Also, the energy of a floor holder is often similar to that of the surrounding

speech by the same speaker. Common floor holders include, but are not limited to, the

following: "so," "and," "or," "um," "uh," "let's see," "well," "and what else," "anyway," "I

mean," "okay," and "yeah."

In terms of placement, floor holders do not occur at the beginning of a speaker's turn,

but rather occur throughout the middle and at the end5 of a speaker’s turn. Although

floor holders do not occur at the beginning of a speaker’s turn or speech, they may

occur at the beginning of a speaker's utterance. If a speaker begins his turn with a floor

grabber followed by a floor holder, it is permissible to label the suspected floor holder as

such.

Section 2 discusses the treatment of floor holders in succession.

Floor holders are often found mid-utterance. In such cases, if an utterance is complete

and splitting it to mark the floor holder would yield an incomplete utterance, the

utterance remains intact and the floor holder is not marked.

In some cases, an utterance will end with a typical floor-holding word such as "um" or

"uh" and, despite the presence of a common floor-holding word, a floor holder is not

actually present, since the floor-holding word lacks the duration or "pause" property

common to most floor holders. If such occurs, the utterance, while containing the floor-

holding word, is simply marked as incomplete and the floor-holding word is not marked

as an actual floor holder.

As previously mentioned, floor holders are not to be identified solely based upon the

vocabulary used, as floor holders, floor grabbers <fg>, holds <h>, backchannels ,

acknowledgements <bk>, and accepts <aa> share a very similar vocabulary. In order

to properly distinguish whether an utterance is performing as a floor holder, floor

grabber, hold, backchannel, acknowledgement, or accept, it is necessary to take into

account the details provided within the individual tag descriptions and to listen to the

audio portions corresponding to the examples within those tag descriptions. Utterances

labeled with these tags tend to appear very similar in text yet emerge exceedingly

different in sound.

5 As mentioned in Section 2, floor holders are not permitted to occur at the end of utterances. The

treatment of floor holders within the transcript is discussed in Section 2 and Section 3.3.

Example 61 through Example 65 present floor holders in context:

 Hold <h>

The <h> tag is used when a speaker who is given the floor and is expected to speak

"holds off" prior to making an utterance. The <h> tag is predominantly used when a

speaker is responding to a question that he in particular was asked, and that speaker

pauses or "holds off" prior to answering the question.

Common holds include, but are not limited to, the following: "so," "um," "uh," "let's see,"

"well," "I mean," "okay," and "yeah."

Holds are very similar to floor holders <fh> in the way that they sound, however holds

occur at the beginning of a speaker's turn, as opposed to floor holders which occur in

the middle or at the end of a speaker's turn.

xample 61: Bed003

2524.030-2526.510 c1 s so it's a - it's a rather huge huge thing .

2526.510-2531.970 c1 fh|s.%-- but um - um - | we can sort of ==

Example 62: Bed003

2579.930-2581.760 c4 s like all the different sort of general

schemas that they might be following .

2581.760-2583.600 c4 fh okay .

Example 63: Bed004

1336.010-1339.280 c2 s i think we got plenty of stuff to talk

about .

1340.180-1344.840 c2 fh|s and then um - | just see how a

discussion goes .

Example 64: Bed004

1596.700-1598.000 c2 s^arp no i understand that .

1598.000-1599.540 c2 fh but i- - but um ==

Example 65: Bed004

1672.310-1673.880 c2 fg okay so so ==

1673.880-1675.440 c2 fh uh ==

Although the primary distinction between holds and floor holders is location, holds are

not collapsed with floor holders as they provide explicit information regarding a

speaker’s turn. Utterances marked as holds explicitly indicate that a speaker is given

the floor, whereas utterances marked as holds indicate that a speaker merely has the

floor.

If a speaker's initial utterance is marked as a hold and his following utterances appear to

be either holds or floor holders, those following utterances are marked as holds. In

other words, if a speaker's initial utterance is a hold and his following utterances are

seemingly floor holders, those utterances appearing as floor holders are marked as

holds until an utterance is encountered that is to be marked with a question tag or with

the statement tag. After such a question or statement is encountered, any following

segment within that same speaker's speech that appears to be a floor holder is marked

as a floor holder and not as a hold.

As previously mentioned, holds are not to be identified solely based upon the

vocabulary used, as holds, floor grabbers <fg>, floor holders <fh>, backchannels ,

acknowledgements <bk>, and accepts <aa> share a very similar vocabulary. In order

to properly distinguish whether an utterance is performing as a hold, floor grabber, floor

holder, backchannel, acknowledgement, or accept, it is necessary to take into account

the details provided within the individual tag descriptions and to listen to the audio

portions corresponding to the examples within those tag descriptions.

Example 666 through Example 68 present instances of holds in context:

6 In Example 66, the word “uhhuh” is used as a floor holder <fh>. Although the word “uhhuh” is not

commonly used as a floor holder, this instance exemplifies the need to listen to corresponding audio

portions in order to correctly assess the function of an utterance and not to label utterances according

to the vocabulary used alone.

xample 66: Bro021

817.043-821.220 c1 qw i mean what was the rest of the

system ?

820.060-821.922 c2 h um ==

823.605-827.084 c2 s yeah it was - it was uh the same

system .

828.960-829.683 c2 fh uhhuh .

830.079-831.107 c2 s^r it was the same system .

838.050-839.197 c2 fh huh ==

Example 67: Bro021

3238.590-3243.580 c1 qy^d^rt so you estimated uh f- -

completely forgetting what you had

before ?

5.5 Group 4: Backchannels and Acknowledgments

This group contains the general tag for backchannels and the specific tags for

acknowledgments <bk>, assessments/appreciations <ba>, and rhetorical question

backchannels <bh>. The commonality among the tags of this group is that they are

most often used to mark utterances that are often responses, in the form of

acknowledgments or backchannels, to a speaker who has the floor as that speaker is

talking. Such responses generally do not elicit feedback. Also, utterances marked with

these tags generally do not serve the purpose of halting the speaker who has the floor.

It may seem as though the tags <bk> and <ba> could be grouped with the tags in Group

5, since they are responses of a sort, they are instead placed in Group 4 due to the

nature of the utterances they mark. The tags in Group 5 are limited to being

orthogonally categorized as positive, negative, or uncertain. Utterances marked with

<bk> are perceived as being neutral, whereas utterances marked with <ba> can be

either positive or negative. Thus the tag <ba> is not included within Group 5 as its

dynamic nature would prevent the preservation of the orthogonal categorization scheme

within Group 5. Additionally, utterances marked with the tag <ba> generally tend to

have more in common with utterances marked with the tag <bk> than with the tags in

Group 5. These similarities are discussed in the tag description for <ba>.

3244.200-3248.840 c4 h um ==

3248.840-3251.170 c4 s^ar|s^nd no no no | it's not completely noise .

Example 68: Bro018

1542.550-1546.120 c5 qy^rt does there some kind of a distance

metric that they use ?

1546.120-1549.520 c5 qw or how do they for cla- - what do they

do for classification ?

1550.050-1550.740 c0 h um ==

1550.740-1551.150 c0 h right .

1551.150-1559.900 c0 s so the - the simple idea behind a

support vector machine is um - you

have - you have this feature space .

 Backchannel

Utterances which function as backchannels are not made by the speaker who has the

floor. Instead, backchannels are utterances made in the background that simply

indicate that a listener is following along or at least is yielding the illusion that he is

paying attention. When uttering backchannels, a speaker is not speaking directly to

anyone in particular or even to anyone at all.

Common backchannels include the following: "uhhuh," "okay," "right," "oh," "yes,"

"yeah," "oh yeah," "uh yeah," "huh," "sure," and "hm."

The nature of backchannels does not usually permit utterances such as "uh," "um," and

"well" as being perceived as backchannels, since these utterances do not indicate that a

speaker is following along, but rather that a speaker has something to say or else is

attempting to say something.

As previously mentioned, backchannels are not to be identified solely based upon the

vocabulary used, as backchannels, floor grabbers <fg>, floor holders <fh>, holds <h>,

acknowledgements <bk>, and accepts <aa> share a very similar vocabulary. In order

to properly distinguish whether an utterance is performing as a floor grabber, floor

holder, hold, backchannel, acknowledgement, or accept, it is necessary to take into

account the details provided within the individual tag descriptions and to listen to the

audio portions corresponding to the examples within those tag descriptions. Utterances

labeled with these tags tend to appear very similar in text yet emerge exceedingly

different in sound.

Furthermore, backchannels are most often confused with acknowledgments and

accepts than with floor grabbers, floor holders, and holds. One method in distinguishing

if the , <bk> or <aa> tag is appropriate lies in the point at which the utterance occurs

with regard to the speaker who has the floor's utterance. Acknowledgments generally

appear after another speaker has completed a phrase or an utterance, as they are

acknowledging the semantic significance of what is said. Accepts usually occur at the

end of another speaker's utterances, as they are agreeing with what is said.

Backchannels, although they can occur in the same locations as acknowledgments and

accepts, can also be found in the middle of another speaker's phrase. Such mid-

phrasal placement is a strong indicator that an utterance is a backchannel, rather than

an acknowledgment or an accept, as the speaker uttering the backchannel lacks

adequate semantic information from the other speaker's utterance to acknowledge it or

agree to it. Additionally, backchannels are usually uttered with a significantly lower

energy level than the surrounding speech, while acknowledgments tend not to be quite

so low as backchannels and accepts are generally at the same level or else higher.

Additionally, the only specific tag that may be appended to a backchannel is the rising

tone tag <rt>.

Backchannels in context are seen in Example 69 through Example 71:

 Acknowledgment <bk>

The <bk> tag is used to express a speaker's acknowledgment of a previous speaker's

utterance or of a semantically significant portion of a previous speaker's utterance.

Acknowledgments are neither positive nor negative, as they only serve to acknowledge,

not to agree or disagree. In some cases, a speaker will acknowledge his own utterance

or a semantically significant portion of his own utterance.

Common acknowledgments, in addition to mimicked portions, include, but are not

limited to, the following: "I see," "okay," "oh," "oh okay," "yeah," "yes," "uhhuh," "huh,"

"ah," "all right," and "got it." If an utterance is suspected to be an acknowledgment

solely based upon the vocabulary used, yet does not sound as though it is an

acknowledgment, then it should not be marked as one.

As opposed to backchannels, acknowledgments encode a level of direct communication

between speakers. A speaker who acknowledges a previous speaker's utterance is

actually speaking directly to that previous speaker, yet is usually not seeking a response

from the previous speaker. As stated in the tag description for backchannels, the tags

<bk>, , and <aa> are often confused with one another. The tag description for

backchannels elucidates how to distinguish among the three tags.

Acknowledgements also tend to be confused with floor grabbers <fg>, floor holders

<fh>, and holds <h> due to their similar vocabularies. In order to properly distinguish

the function of an utterance, it is necessary to take into account the details provided

xample 69: Bro018

1821.160-1829.060 c2 s but i think that uh - this was a couple

years ago .

1821.510-1821.820 c5 b huh .

Example 70: Bro018

2005.020-2012.090 c5 qy^rt do you get out a - uh - a vector of

these ones and zeros and then try to

find the closest matching phoneme to

that vector ?

2006.210-2006.410 c0 b uhhuh .

Example 71: Bro007

837.018-838.648 c1 s^df well also just to know the numbers .

837.345-837.565 c3 b yeah .

838.648-838.828 c1 b right .

within the individual tag descriptions and to listen to the audio portions corresponding to

the examples within those tag descriptions. Utterances labeled with the <bk>, <fg>,

<fh>, and <h> tags, as well as with the and <aa> tags, tend to appear very similar

in text yet emerge exceedingly different in sound.

Restrictions apply to the usage of the <bk> tag with other specific tags. The <bk> tag is

only used when the primary function of an utterance is to acknowledge a portion of

another speaker's speech. The use of other tags to mark an utterance, such as those in

Group 5, indicates that an utterance serves a different primary purpose, such as

agreeing or disagreeing. So, when a tag from Group 5 is used to mark an utterance,

the <bk> tag may not be used in conjunction with that tag.

The <bk> tag also may not be used with <ba>, as the <ba> tag encodes the

acknowledging nature of <bk> within its definition and thus renders the <bk> tag

redundant when the two are used in conjunction. The use of the <ba> tag also

indicates that an utterance is either positive or negative, whereas an utterance marked

with the <bk> tag is neutral. The <bk> tag may not be used with <bh>, as <bh> is a

type of backchannel or acknowledgment, depending upon its usage, and may encode

the acknowledging nature of <bk> thus rendering the use of the <bk> tag redundant

when used in conjunction.

The specific tags with which <bk> is permitted to be used in conjunction are <m>, <r>,

<rt>, <fe>, <t1> and <t3>. When used in conjunction with the <bk> tag, a tag from this

list merely indicates a feature of the acknowledgment. In the case of the tag <fe>, when

used in conjunction with the tag <bk>, it indicates that an exclamatory acknowledgment

was uttered. When used with another functional tag, such as <aa> or <cs>, the tag

<fe> indicates that an exclamatory agreement or an exclamatory suggestion has been

made.

Acknowledgments in context are seen in Example 72 through Example 76:

xample 72: Bmr012

58.784-60.504 c3 qw^t3 so why didn't you get the same

results and the unadapted ?

62.153-64.053 c3 qw^r^t3 why didn't you get the same results

as the unadapted ?

64.235-68.995 c0 s^t3 oh because when it estimates the

transformer pro- - produces like

single matrix or something .

67.730-69.010 c3 s^bk^t3 o- - oh i see .

Example 73: Bed003

151.920-155.150 c1 s it opens the assistant that tells you that

 Assessment/Appreciation <ba>

Assessments/appreciations are acknowledgments directed at another speaker's

utterances and function to express slightly more emotional involvement than what is

seen in the utterances marked with the <bk> tag. The <ba> tag is similar to the <bk>

tag in that it acknowledges another speaker's utterance, however it lacks the neutral

nature of the <bk> tag. Utterances marked with <ba> can be either positive or negative.

When negative, utterances marked with the <ba> tag are often criticisms.

Utterances which function as acknowledgments in the senses discussed under the tag

descriptions for <bk>, <bh>, and <ba> may only be marked with one of these tags to

express the acknowledging nature of an utterance, not a combination of these tags.

As with the <bk> tag, the <ba> tag encodes a level of direct communication between

speakers. When appreciating or assessing the contents of a previous speaker's

utterance, a speaker is actually speaking directly to the previous speaker, yet usually is

not seeking a response from the previous speaker.

Although most utterances marked with the <ba> tag tend to be quite short, some

utterances tend to be somewhat lengthy. This is due to the very nature of the <ba> tag.

In briefly expressing appreciation or assessing a situation, which is usually the case, a

speaker's utterance may be something to the likes of "that's great," "that's terrible,"

"good enough," "wow," or "excellent." Brief utterances such as these are often uttered

as exclamations, thus requiring the <fe> tag.

the font type is too small .

155.780-156.120 c2 s^bk ah .

Example 74: Bed003

158.220-159.100 c2 s^nd i'd prefer not to .

159.140-159.500 c1 s^bk okay .

Example 75: Bed003

166.460-169.010 c2 s^rt because i'm going to switch to the

javabayes program .

167.820-168.400 c1 s^bk oh okay .

Example 76: Bed003

1615.540-1617.810 c2 s^rt so we can rel- open it up again .

1616.130-1616.410 c3 s^bk okay .

Longer appreciations tend to be akin to utterances such as "so I think that's a really

great way to approach it." Longer assessments tend to appear as criticisms, which take

many forms. Comments and opinions on an aspect a speaker has noticed within the

contents of another speaker's speech are often marked as assessments/appreciations

also.

In some cases, utterances which are assessments/appreciations are also affirmative

answers <na>, dispreferred answers <nd>, or negative answers <ng>. In these cases,

an utterance that is assessing or appreciating is also communicating that it is agreeing

or disagreeing. An utterance such as "I think that would be worth doing" would function

as an assessment/appreciation in that it embeds the speaker's own opinion. Assuming

the utterance is actually agreeing to another speaker's previous utterance, the utterance

also functions as an affirmative answer in that it accepts and agrees to what the

previous speaker said. An utterance such as "that's wonderful" is an

assessment/appreciation, yet is not an agreement since it only expresses an

assessment.

In determining whether an utterance is indeed an assessment/appreciation, it is

necessary to ensure that the assessment/appreciation is actually uttered in reference to

another speaker's utterance.

A variety of assessments/appreciations are seen in Example 77 through Example 89:

xample 77: Bed006

172.462-173.242 c3 s^ba it's very exciting .

Example 78: Bed006

257.526-257.916 c3 s^ba that's good .

Example 79: Bed006

266.653-267.043 c2 s^ba wonderful .

Example 80: Bed006

347.295-347.615 cA s^ba it's fine .

Example 77: Bed006

172.462-173.242 c3 s^ba it's very exciting .

Example 78: Bed006

257.526-257.916 c3 s^ba that's good .

xample 79: Bed006

266.653-267.043 c2 s^ba wonderful .

Example 80: Bed006

347.295-347.615 cA s^ba it's fine .

Example 81: Bmr021

261.000-262.000 c4 s^ba^fe wow !

Example 82: Bed006

1333.750-1337.640 c2 s^ba but it's - so this time we - we are at an

advantage .

Example 83: Bed008

1873.870-1876.850 c2 fg|s^ba uh - | anyway this is crude .

Example 84: Bed008

2035.000-2036.000 c2 s^ba but this is a good discussion .

Example 85: Bed008

3878.640-3880.450 c4 s^ba so this is slightly uh - more

complicated .

Example 86: Bed008

4997.490-5002.340 c0 s^ba that's uh - that's a whole lot of

constructions .

Example 87: Bed017

1462.890-1467.820 c2 s^ba so it's probably not that easy to simply

have a symbolic uh computational

model .

Example 88: Bmr002

1992.220-1996.800 c2 s^ba and i was very impressed by how well

you could hear separate speakers .

 Rhetorical Question Backchannel <bh>

Rhetorical question backchannels lack semantic content and are syntactically similar to

rhetorical questions, however they function as backchannels and acknowledgments.

Rhetorical question backchannels can be uttered as backchannels, which is often the

case, in that they can be made in the background and simply indicate that a listener is

following along or at least is yielding the illusion that he is paying attention. In these

cases, the use of a rhetorical question backchannel indicates that a speaker is not

speaking directly to anyone in particular or even to anyone at all. When uttered as an

acknowledgment, the rhetorical question backchannel expresses a speaker's

acknowledgment of a previous speaker's utterance or of a semantically significant

portion of a previous speaker's utterance. As acknowledgments, rhetorical question

backchannels encode a level of direct communication between speakers. A speaker

who acknowledges a previous speaker's utterance is actually speaking directly to that

previous speaker, yet is usually not seeking a response from the previous speaker.

However, when acknowledgments are uttered as rhetorical question backchannels, they

often receive answers such as "yeah." Additionally, when a rhetorical question

backchannel functions as an acknowledgment, it is unnecessary to mark the <bk> tag.

As stated in the tag descriptions for <bk> and <ba>, the default tag for

acknowledgments is the <bk> tag. If further descriptions apply to an acknowledgment

and a <ba> or <bh> tag is deemed necessary, than only one of these tags is used. The

<bk> tag cannot be used in conjunction with the <ba> or <bh> tags.

Common rhetorical question backchannels include, but are not limited to, the following:

"oh really?", "yeah?", "isn't that interesting?", and "you think so?".

Rhetorical question backchannels always receive the Y/N question general tag <qy>.

Example 90 through Example 99 present instances of rhetorical question backchannels:

xample 89: Bmr021

747.750-749.530 c0 fg|s^ba^cs well | it seems like just shortening them

is a good short term solution .

xample 90: Bed003

2136.810-2137.060 c1 qy^bh yeah ?

Example 91: Bed003

2319.660-2319.910 c2 qy^bh really ?

xample 92: Bed003

3493.590-3494.000 c3 qy^bh oh really ?

Example 93: Bmr005

1358.460-1358.690 c3 qy^bh^rt yeah ?

Example 94: Bmr012

671.580-672.090 c4 qy^bh^d^rt oh it did ?

Example 95: Bmr014

522.800-523.120 c8 qy^bh^m^rt no ?

Example 96: Bmr014

2357.840-2358.290 c8 qy^bh oh they won't ?

Example 97: Bmr021

193.000-194.000 c5 qy^bh isn't that something ?

Example 98: Bmr021

859.540-860.670 c5 qy^bh is that right ?

Example 99: Bro021

170.110-170.542 c5 qy^bh huh ?

5.6 Group 5: Responses

Group 5 is orthogonally divided into three subgroups: positive utterances, negative

utterances, and uncertain utterances. The tags in Group 5 are often used to

characterize responses to questions and suggestions.

POSITIVE

 Accept <aa>

The <aa> tag is used for utterances which exhibit agreement to or acceptance of a

previous speaker's question, proposal, or statement. Utterances marked with the <aa>

tag are quite short, as their lengthy counterparts are marked with the <na> tag.

Common utterances marked with the <aa> tag include, but are not limited to, the

following: "yeah," "yes," "okay," "sure," "uhhuh," "right," "I agree," "exactly," "definitely,"

and "that's true."

Additionally, the word "no" can be marked with the <aa> tag if it is used to agree to a

syntactically negative statement or question, as seen in Example 104.

Utterances marked with the <aa> tag may be confused with backchannels and

acknowledgments. Generally, utterances marked with the <aa> tag have much more

energy and are more assertive than backchannels and acknowledgments. The tag

descriptions for backchannels and acknowledgments further elucidate the distinctions

among the three tags.

Accepts are not to be identified solely based upon the vocabulary used, as accepts,

floor grabbers <fg>, floor holders <fh>, holds <h>, backchannels , and

acknowledgements <bk> share a very similar vocabulary. In order to properly

distinguish whether an utterance is performing as an accept, floor grabber, floor holder,

hold, backchannel, or acknowledgement, it is necessary to take into account the details

provided within the individual tag descriptions and listen to the audio portions

corresponding to the examples within those tag descriptions. Utterances labeled with

these tags tend to appear very similar in text yet emerge exceedingly different in sound.

Accepts in context are seen in Example 100 through Example 104:

xample 100: Bro017

2264.620-2271.560 c3 s.x if you want to decrease the importance

of a c- - parameter you have to

increase it's variance .

2267.450-2267.830 c1 s^aa yes .

2269.590-2269.840 c1 s^aa.x right .

2269.690-2269.980 c4 s.x multiply .

2270.470-2270.690 c1 s^aa yes .

2271.610-2272.050 c1 s^aa exactly .

Example 101: Bro022

1575.820-1579.190 c0 s^df because when you train up the aurora

system you're uh - you're also training

on all the data .

1579.190-1582.560 c0 s.%-- i mean it's ==

1580.350-1580.920 c2 s^aa that's right .

1580.920-1581.490 c2 s^aa yeah .

Example 102: Bro022

1475.950-1477.970 c4 s and it was about six point six percent .

1477.390-1477.780 c2 s^bk oh .

1477.790-1478.630 c1 s^aa right right right right .

1478.630-1479.470 c1 s^bk okay .

Example 103: Bro026

2416.730-2418.050 c2 s because that's what you're going to be

using .

2418.050-2418.210 c2 qy^d^g^rt right ?

2418.250-2418.740 c3 s^aa yeah .

2418.740-2419.220 c3 s^aa^r yeah .

Example 104: Bro026

854.850-858.060 c2 s^nd although you - you know you haven't

tested it actually on the german and

danish .

858.060-858.360 c2 qy^d^g^rt have you ?

858.850-859.520 c0 s^aa no we didn't .

 Partial Accept <aap>

The <aap> tag marks when a speaker explicitly accepts part of a previous speaker's

utterance. Partial accepts are often conditional responses that accept or agree to

another speaker's utterance.

Partial accepts are often confused with partial rejections <arp>. The distinction is that

an utterance marked with the <aap> tag focuses on agreeing with or accepting part of a

previous speaker's utterance. An utterance marked with the <arp> tag focuses on

disagreeing with or rejecting part of a previous speaker's utterance.

Partial accepts in context are seen in Example 105 through Example 108:

xample 105: Bed003

922.295-924.105 c1 s^bu^rt well the - the - sort of the landmark is

- is sort of the object .

924.105-925.915 c1 qy^d^g right ?

925.915-927.595 c1 qy^d^g^rt the argument in a sense ?

927.230-928.260 c4 s^aap usually .

Example 106: Bmr024

1147.330-1156.120 c3 fh|qy^bu^d um so | it's wizard in the sen- - usual

ense that the person who is asking the

questions doesn't know that it's uh a

machi- - not a machine ?

1155.600-1156.190 c5 s^aap at the beginning .

Example 107: Bmr006

944.455-949.460 c3 s but i think that - i'm raising that

because i think it's relevant exactly for

his idea up there that if you think about

well gee we have this really

complicated setup to do well maybe

you don't .

950.300-961.150 c3 s^cs maybe if - if - if really all you want is to

have a - a - a recording that's good

enough to get a - uh a transcription

from later you just need to grab a tape

recorder and go up and make a

recording .

950.660-951.260 c1 s^aap for some of it .

 Affirmative Answer <na>

The <na> tag marks an utterances that act as narrative affirmative responses to

questions, proposals, and statements. The <na> tag is much like the <aa> tag in that

they both exhibit agreement to or acceptance of a previous speaker's question,

proposal, or statement. The difference between the two tags is that, as the <aa> tag is

used for shorter utterances, the <na> tag is used for lengthy utterances.

In order to determine whether an utterance requires the <na> tag, the surrounding

context is generally required. Without surrounding context, an utterance requiring the

<na> tag may be considered merely as a statement <s> without any additional specific

tags representing agreement or acceptance.

Instances of the <na> tag in context are seen in Example 109 through Example 111:

xample 108: Bro007

1605.290-1612.800 c2 s^cs and - and perhaps i was thinking also a

fourth one with just - just a single k l t .

1612.800-1616.550 c2 s^df because we did not really test that .

1616.550-1620.300 c2 s^cs removing all these k l t's and putting

one single k l t at the end .

1622.760-1626.240 c1 s^na yeah i mean that would be pretty low

maintenance to try it .

1626.970-1628.480 c1 fh|s^aap uh - | if you can fit it in .

xample 109: Bed011

1528.600-1530.280 c2 s nobody's interested in that except for

the speech people .

1529.120-1529.290 c3 s^aa right .

1529.290-1530.300 c3 s^na no we don't care about that at all .

Example 110: Bmr001

374.134-377.954 c8 s a cabinet is probably going to cost a

hundred dollars two hundred dollars

something like that .

378.105-381.715 c0 s^na yeah i mean - you know - we - we can

spend under a thousand dollars or

something without - without worrying

about it .

NEGATIVE

 Reject <ar>

The <ar> tag marks negative words such as "no" and other semantic equivalents that

offer negative responses to questions, proposals, and statements. The <ar> tag marks

brief negative responses to questions, proposals, and statements in the same manner

that the <aa> tag marks brief affirmative answers.

Common utterances marked with the <ar> tag include, but are not limited to, the

following: "no," "nope," "no way," "nah," "not really," and "I don't think so."

When syntactically negative questions or statements arise, responses in the form of

"yes," "yeah," or the like can function as rejections. As discussed in the tag description

for <aa>, negative responses such as "no" can function as agreements in these cases.

Rejections in context are seen in Example 112 through Example 116:

xample 111: Bmr007

1656.590-1664.310 cA s if - if the goal were to just look at

overlap you would - you could serve

yourself -

save yourself a lot of time but

not even transcri- transcribe the

words .

1666.090-1668.990 c1 s well i was thinking you should be able

to do this from the acoustics on the

close talking mikes .

1668.990-1671.900 c1 qy^d^g right ?

1671.140-1674.800 cB s^na well that's - the - that was my - my

status report .

xample 112: Bed003

259.160-264.920 c4 qy.%- but are you saying that in this particular

domain it happens the - that

landmarkiness cor- - is correlated

with ?==

263.409-264.019 c3 s^ar no .

 Partial Reject <arp>

The <arp> tag marks when a speaker explicitly rejects part of a previous speaker's

utterance. Partial rejections are often responses posing exceptions when rejecting

another speaker's utterance.

Partial rejections are often confused with partial accepts <aap>. As stated in the tag

description for <aap>, the distinction between the two is that an utterance marked with

the <aap> tag focuses on agreeing with or accepting part of a previous speaker's

utterance. An utterance marked with the <arp> tag focuses on disagreeing with or

rejecting part of a previous speaker's utterance. An utterance marked with the <aap>

tag is formulated in a positive manner, whereas an utterance marked with the <arp> tag

is formulated in a negative manner.

Partial rejections in context are seen in Example 117 through Example 1197:

7 The tag <sj> is seen in Example 19. This tag was formerly part of the MRDA tagset eliminated in the

revision of the tagset. Appendix 4 details tags which are no longer a part of the MRDA tagset.

xample 113: Bed003

545.980-548.160 c4 qy and are those mutually exclusive sets ?

547.610-547.990 c3 s^ar not at all .

Example 114: Bed003

1758.350-1760.280 c2 qy^rt i didn't n- - is there an ampersand in

dos ?

1761.030-1761.370 c3 s^ar nope .

Example 115: Bed003

3022.070-3023.720 c2 qy^rt do you want to trade ?

3023.360-3024.610 c1 h|s^ar um - | no .

Example 116: Bed011

2776.460-2779.490 c1 qr.%- is that roughly the equivalent of - of

what i've seen in english or is it ?==

2779.390-2780.180 c2 s^ar no not at all .

 Dispreferred Answer <nd>

The <nd> tag marks statements which act explicit narrative forms of negative answers

to previous speakers' questions, proposals, and statements in the same manner in

which the <na> tag acts as an agreement with or acceptance of a previous speaker's

utterance. As with the <na> tag, the <nd> tag marks lengthier utterances than those

marked with the <ar> tag which exhibit rejection.

Surrounding context is generally required to determine whether an utterance requires

the <nd> tag. Without surrounding context, an utterance requiring the <nd> tag may be

considered merely as a statement <s> without any additional specific tags representing

rejection.

Dispreferred answers are often confused with negative answers <ng>. The main

distinction between the two tags is that the <nd> tag marks utterances that offer explicit

rejections and the <ng> tag marks utterances that offer implicit rejections through the

use of hedging.

xample 117: Bed003

1352.970-1355.790 c2 qy^bu^rt also - you know - didn't we have a size

as one ?

1357.120-1357.350 c3 qw^br what ?

1357.330-1358.250 c2 s^r^rt the size of the landmark .

1359.860-1361.550 c3 s^arp um - not when we were doing this .

Example 118: Bed003

1131.440-1132.880 c2 s it would actually slow that down

tremendously .

1136.540-1137.290 c3 s^arp not that much though .

Example 119: Bmr018

505.460-507.485 c4 s but you're listening to the mixed signal

and you're tightening the boundaries .

507.485-509.510 c4 s^bsc correcting the boundaries .

509.510-512.510 c4 s you shouldn't have to tighten them too

much because thilo's program does

that .

511.313-512.073 c0 sj.x should be pretty good .

512.550-515.710 c3 s^arp except for it doesn't do well on short

things remember .

Dispreferred answers in context are seen in Example 120 through Example 124:

 Negative Answer <ng>

As opposed to a dispreferred answer <nd> which explicitly offers a negative response to

a previous speaker's question, proposal, or statement, a negative answer <ng> implicitly

offers a negative response with the use of hedging.

The negative answer tag <ng> is often confused with the maybe tag <am> and the no

knowledge tag <no>. The maybe tag <am> marks utterances in which a speaker

asserts that his response is probable, yet not definite, and the no knowledge tag <no>

marks utterances in which a speaker does not know an answer. A negative answer

<ng> essentially offers an indirect negative response. In uttering an indirect negative

response, a speaker may employ responses similar to those marked with the maybe tag

xample 120: Bmr001

948.121-951.731 c8 s^bu^rt we figured out that it was t- - twelve

gig- - twelve gigabytes an hour .

949.056-949.806 c1 s^nd it was more than that .

Example 121: Bed003

156.910-157.510 c1 qy^rt do you want to try ?

158.220-159.100 c2 s^nd i'd prefer not to .

Example 122: Bed003

1163.060-1166.150 c4 s so i thought that was directly given by

the context switch .

1163.130-1166.160 c3 s^nd that's a different thing .

Example 123: Bmr005

781.990-783.000 c4 s probably de- - probably depends on

what the prepared writing was .

785.281-786.821 c1 s^bk|s^nd yeah | i don't think i would make that

leap .

Example 124: Bmr024

1987.890-1989.760 c1 s^bs he's saying get a whole different drive .

1989.680-1990.810 c5 s^nd but there's no reason to do that .

<am> and no knowledge tag <no> to hedge around uttering a direct refusal or negative

response.

Oftentimes, negative answers <ng> appear as alternative suggestions to a previous

speaker's question, proposal, or statement.

Negative answers <ng> in context are seen in Example 125 through Example 1338:

8 Regarding the use of the tag <sj> in Example 133, refer to footnote 7.

xample 125: Bed004

350.465-352.450 c4 qy^rt y- - you guys have plans for sunday ?

352.900-353.470 c4 s.%-- we're - we're not ==

353.470-360.645 c4 s

t's probably going to be this sunday but

um w- - we're sort of working with the

weather here .

360.645-367.820 c4 s^df because we also want to combine it

with some barbecue activity where we

just fire it up and what - whoever brings

whatever you know can throw it on

there .

368.787-371.447 c4 s so only the tiramisu is free nothing

else .

373.980-377.050 c1 s^ng well i'm going back to visit my parents

this weekend .

Example 126: Bmr005

4094.420-4099.430 c2 qw what if we give people you know - we

cater a lunch in exchange for them

having their meeting here or

something ?

4099.640-4103.350 c1 s^ng well you know - i - i do think eating

while you're doing a meeting is going to

be increasing the noise .

Example 127: Bmr007

14.467-15.967 cB qy^rt and uh shall i go ahead and do

some digits ?

16.724-17.504 c3 h|s^ng uh | we were going to do that at the

end .

xample 128: Bmr007

1750.790-1755.290 cA s we have - have in the past and i think

continue - will continue to have a fair

number of uh phone conference calls .

1756.380-1771.950 cA fh|s^cs and uh | and as a - to um as another

c- c- comparison condition we could

um see what - what what happens in

terms of overlap when you don't have

visual contact .

1774.140-1777.190 cB s^ng it just seems like that's a very different

thing than what we're doing .

Example 129: Bmr007

1773.730-1774.870 c1 qy^rt can we actually record ?

1775.870-1778.340 c3 fh|s^ng uh | well we'll have to set up for it .

Example 130: Bmr014

2637.240-2645.800 cB s i mean so it's like i- - in a way it's - it's

nice to have the responsibility still on

them to listen to the tape and - and

hear the transcript .

2645.800-2646.660 cB s.%-- to have that be the ==

2647.970-2652.800 c8 s^ng i mean most people will not want to

take the time to do that though .

Example 131: Bmr024

1237.760-1240.380 c9 s^cs maybe we can have him vary the

microphones too .

1241.190-1243.470 c5 fg|s so - so - so | for their usage they don't

need anything .

1243.880-1246.890 c4 s^ng but - but i'm not sure about the legal

aspect of - of that .

Example 132: Bmr024

2385.660-2389.950 cB s.%-- it might be that one more iteration

would - would help but it's sort of ==

2390.330-2390.650 cB fh you know .

2390.440-2392.350 c3 s^ng or maybe - or maybe you're doing one

too many .

UNCERTAIN

 Maybe <am>

The maybe tag <am> marks utterances in which a speaker's utterance conveys

probability or possibility by using the word "maybe" or other words denoting possibility

and probability. An utterance marked with the <am> tag is one which the speaker

asserts that his utterance is probable or possible, yet not definite.

The <am> tag is often confused with suggestions <cs> which have the form of "maybe

we should..."

Maybes <am> in context are seen in Example 134 through Example 138:

xample 133: Bmr024

818.269-825.296 c5 s sure there - there might be a place

where it's beep seven beep eight beep

eight beep .

826.056-829.156 c5 s but you know they - they're - they're

going to macros for inserting the beep

marks .

830.078-831.768 c5 sj and so i - i don't think it'll be a

problem .

831.768-832.708 c5 s^cs we'll have to see .

832.708-833.648 c5 sj^r but i don't think it's going to be a

problem .

834.643-834.903 c3 s^bk okay .

835.101-836.021 c3 fg|s^ng well | i - i - i don't know .

836.021-848.194 c3 s^cs i - i think that that's - if they are in fact

going to transcribe these things uh

certainly any process that we'd have to

correct them or whatever is - needs to

be much less elaborate for digits than

for other stuff .

xample 134: Bed003

1228.410-1231.250 c1 qw^rt we- - what set the - they set the

context to unknown ?

1232.500-1233.580 c3 s right now we haven't observed it .

 No Knowledge <no>

The no knowledge tag <no> marks utterances in which a speaker expresses a lack of

knowledge regarding some subject.

The most common expressions found within utterances marked with the no knowledge

tag are "I don't know" and "I'm not sure." However, in some cases, utterances

consisting of "I don't know" are actually floor holders <fh> and are not to be marked with

the no knowledge tag.

1233.580-1236.710 c3 s^am so i guess it's sort of averaging over all

those three possibilities .

Example 135: Bed003

2969.930-2971.610 c3 qy^rt is srini going to be at the meeting

tomorrow ?

2971.610-2971.870 c3 qy^rt do you know ?

2972.580-2972.910 c4 s^am maybe .

Example 136: Bed003

3206.200-3214.190 c1 s.%-- but you know - if we take a subject that

is completely unfamiliar with the task or

any of the set up we get a more

realistic ==

3212.060-3213.000 c3 s^am i guess that would be reasonable .

Example 137: Bmr009

1752.000-1754.000 c0 qw so - so what accent are we speaking ?

1756.500-1761.000 c3 s^am probably western yeah .

Example 138: Bmr018

1890.390-1893.760 c0 s^df because you have to uh - maneuver

around on the - on both windows then .

1895.010-1895.960 c4 qr^d to add or to delete ?

1896.110-1896.480 c0 s to delete .

1898.510-1898.860 c4 s^bk^rt okay .

1898.970-1900.440 c3 fg|%- anyways | so i - i guess ==

1900.380-1904.150 c4 s^am that - maybe that's an interface issue

that might be addressable .

Utterances marked with the no knowledge tag may be confused with utterances marked

with the negative answer tag <ng>. The tag description for the <ng> tag elucidates this

issue.

Instances of utterances labeled with the no knowledge tag, where some are shown in

context, are seen in Example 139 through Example 146:

xample 139: Bed003

142.790-146.410 c1 s but if you really want to find out what

it's about you have to click on the little

light bulb .

147.130-148.810 c2 s^no although i've - i've never - i don't know

what the light bulb is for .

Example 140: Bed003

1281.990-1284.650 c3 s^no but uh - i don't know y- what the right

thing is to do for that .

Example 141: Bed004

1417.360-1418.320 c2 s^no yeah i don't understand it .

Example 142: Bmr001

68.756-70.816 c0 fg|s^no um - | i have no idea which one i'm -

i'm on .

Example 143: Bmr001

354.108-359.588 c1 qy do we have any money at all that we

can go out and spend on things like

cabinets or a hard drive or things like

that ?

359.791-360.451 c0 h|s^no oh - i mean - | i don't know .

Example 144: Bmr001

366.306-368.646 c0 h|qw^rt uh | how much are we talking about

here ?

371.211-374.134 c8 h|s^no um - | i don't know .

Example 145: Bmr001

1365.460-1366.620 c0 qy didn't we already get that ?

5.7 Group 6: Action Motivators

This group contains specific tags pertaining to future action. Whether the future action

occurs immediately or after a long period of time is not relevant.

The tags in Group 6 either indicate that a command or a suggestion has been made

regarding some action to be taken at some point in the future or else indicate that a

speaker has committed himself to executing some action at some point in the future.

 Command <co>

The <co> tag marks commands. In terms of syntax, a command may arise in the form

of a question (e.g., "Do you want to go ahead?") or as a statement (e.g., "Give me the

microphone.").

Commands are often confused with suggestions <cs>. The distinction between the two

entails considering what sort of response such an utterance could receive as well as the

role of the speaker within the meeting. In terms of responses, commands are uttered as

orders, where a failure to comply (e.g., a "no" answer), in an extreme sense, is

perceived as a sign of indignation toward the speaker uttering the command. With

regard to a suggestion, rejecting a suggestion is not considered as impolite as rejecting

a command. If an utterance yields the illusion that it may be a command or a

suggestion, considering whether the utterance could receive a response that is a

rejection and whether that rejection is considered impolite is a helpful method to

determine if the utterance is a command or a suggestion. If a rejection is considered

impolite, the utterance is considered a command, otherwise it is considered a

suggestion.

In terms of the role of a speaker within a meeting, generally suggestions made by the

speaker running a meeting are perceived as commands. If the speaker running the

meeting says to another speaker, "let's try that one," such an utterance is considered a

command. Whereas, if the same utterance is made by another speaker who is not

running the meeting, then the utterance is considered a suggestion instead. However,

1365.650-1366.140 c8 s^no.% oh god knows .

Example 146: Bed003

2112.730-2113.480 c0 qw who was it trained on ?

2113.770-2114.510 cB h|s^no uh | i have no idea .

2114.740-2115.330 cB s^no i don't remember .

this is not to say that all suggestions made by the speaker running a meeting are to be

considered as commands. In distinguishing between commands and suggestions made

by a speaker running a meeting, it is helpful to consider the method regarding whether a

rejection is impolite as discussed in the previous paragraph.

Commands are seen in Example 147 through Example 162. Note that commands that

appear to be suggestions within these examples are actually commands made by the

speaker running the meeting.

xample 147: Bed003

160.020-160.440 c1 s^co continue .

Example 148: Bed003

177.840-178.190 c4 s^co proceed .

Example 149: Bed003

581.856-582.226 c3 s^co wait .

Example 150: Bed003

1440.550-1441.820 c1 s^co let's get this uh - b- - clearer .

Example 151: Bed003

1467.230-1473.090 c2 s^co explain to me why it's necessary to

istinguish between whether something

has a door and is not public .

Example 152: Bed003

1670.450-1675.190 c1 s^co close it and - and load up the old state

so it doesn't screw - screw that up .

Example 152: Bed003

1761.440-1762.790 c3 s^co just s- - l- - start up a new d o s .

Example 153: Bmr001

127.000-127.450 c1 s^co fill it out .

ample 154: Bmr001

131.458-131.988 c8 s^co just write it down .

Example 155: Bmr001

2016.020-2017.270 c0 s^co well - let's do some more while we got

them here .

Example 156: Bmr005

4248.000-4250.020 c8 fh|s^co so | we should think about trying to

wrap up here .

Example 157: Bmr007

3080.090-3082.130 c3 qw^co so why don't you explain it quickly ?

Example 158: Bro026

236.320-247.993 c2 s^co^t but i guess maybe the thing - since you

weren't - yo- - you guys weren't at

that - that meeting might be just - just

to um - sort of recap uh - the - the

conclusions of the meeting .

Example 159: Bro026

311.870-317.825 c2 fh|s^co^t uh - | maybe describe roughly what -

what we are keeping constant for now .

Example 160: Bro026

2068.470-2071.780 c2 s^co yeah so maybe just c c hari and say

that you've just been asked to handle

the large vocabulary part here .

Example 161: Bro021

2611.590-2618.090 c1 s^bk|s^co okay | so now once you get that - that

one then you - then you do a first- - or

econd order or something taylor series

expansion of this .

 Suggestion <cs>

The suggestion tag marks proposals, offers, advice, and, most obviously, suggestions.

Suggestions are often found in constructions such as "maybe we should..."

Suggestions containing the word "maybe" are not to be confused with the maybe tag

<am>. Additionally, if the phrase "excuse me" precedes something for which a speaker

is negotiating permission (Jurafsky 35), then it is marked as a suggestion rather than an

apology <fa>.

Suggestions are also often confused with commands <co>. The tag description for

<co> clarifies how such might occur.

Suggestions are seen in Example 163 through Example 173:

xample 162: Bro026

614.735-617.130 c2 s^co^t and then uh - maybe you should just

continue telling what - what else is in

the - the form we have .

xample 163: Bro018

948.67-950.165 c5 fg|s^cs yeah | i was just going to say maybe it

has something to do with hardware .

Example 164: Bro021

28.107-28.938 c5 qy^cs^rt should we take turns ?

Example 165: Bro021

28.938-29.768 c5 qy^cs^d^rt you want me to run it today ?

Example 166: Bro021

33.052-36.270 c5 s^cs let's see maybe we should just get a list

of items .

Example 167: Bro021

414.758-419.812 c1 s^cs i- - i really would like to suggest

 Commitment <cc>

The commitment tag <cc> is used to mark utterances in which a speaker explicitly

commits himself to some future course of action. Commitments are not to be confused

with suggestions in which a speaker suggests that he, the speaker himself, execute

some action. With commitments, a speaker mentions what he will do in the future, not

what he might do.

looking um a little bit at the kinds of

errors .

Example 168: Bro021

1967.920-1969.610 c2 s^cs maybe you have to standardize this

thing also .

Example 169: Bro021

1987.380-2000.980 c1 qw^cs um given that we're going to have for

this test at least of - uh boundaries

what if initially we start off by using

known sections of nonspeech for the

estimation ?

Example 170: Bro021

2054.740-2058.370 c4 s^cs if you want you c- - i can say

something about the method .

Example 171: Bro021

2340.390-2341.720 c1 s^cs maybe we can take it off line .

Example 172: Bro021

2564.920-2566.410 c1 s^cs i think these things are a lot clearer

when you can use fonts - different

fonts there .

Example 173: Bro021

711.142-715.021 c1 s^cs and maybe you'd want to have

something that was a little more

adaptive .

Commitments are seen in Example 174 through Example 181:

xample 174: Bmr018

278.930-281.910 c0 s^cc i'll - i'll - i'll um - get - make that

available .

Example 175: Bmr018

526.910-527.560 c4 s^cc^j i'll work on that .

Example 176: Bmr024

1972.600-1974.890 c5 s^cc my intention is to do a script that'll do

everything .

Example 177: Bmr026

196.510-198.560 c5 s^cc i'll send it out to the list telling people to

look at it .

Example 178: Bmr026

202.562-203.282 c0 s^cc i'll try to get to that .

Example 179: Bmr026

211.838-212.668 c0 s^cc i'm just going to do it .

Example 180: Bmr026

218.868-227.628 c0 s^cc i'm going to send out to the participants

uh - with links to web pages which

contain the transcripts and allow them

to suggest edits .

Example 181: Bmr026

271.030-271.440 c5 s^cc i'll wait .

5.8 Group 7: Checks

This group contains specific tags pertaining to understanding or being understood.

 "Follow Me" <f>

The <f> tag marks utterances made by a speaker who wants to verify that what he is

saying is being understood. Utterances marked with the <f> tag explicitly communicate

or else implicitly communicate the questions "do you follow me?" or "do you

understand?" In implicitly communicating those questions, a speaker's utterance may

be a tag question <g>, such as "right?" or "okay?", where a sense of "do you

understand?" is being conveyed.

Tag questions marked with the "follow me" <f> tag often occur in instances in which a

speaker is attempting to be instructional or else is offering an explanation. After an

instruction or explanation, a speaker may utter a tag question <g> that is also a "follow

me" in order to gauge whether what he is saying is understood.

Instances of the "follow me" tag, some of which are shown with their surrounding

context, are seen in Example 182 through Example 187:

xample 182: Bed008

589.304-590.304 c5 qy^d^f^rt this is understandable ?

Example 183: Bmr006

23970.340-3971.190 c1 qy^f^rt do you know what i'm saying ?

Example 184: Bmr007

2821.400-2823.070 c3 qy^d^f^rt you know what i mean ?

Example 185: Bmr008

670.000-676.000 c4 qy^d^f well - i guess i was thinking maybe you

know how you were taking information

off of the digits and putting it onto that ?

Example 186: Bro021

1267.930-1268.770 c0 s.%-- i - i - i was thinking ==

1268.770-1272.600 c0 s^bk|s okay | so just set to - set to some really

 Repetition Request

An utterance marked as a repetition request indicates that a speaker wishes for another

speaker to repeat all or part of his previous utterance. Repetition requests are usually

used when a speaker could not decipher another speaker's previous utterance and

wishes to hear that portion again.

Common repetition requests include, but are not limited to, the following: "what?",

"sorry?", "huh?", "pardon?", "excuse me?", and "say that again." The tag description for

wh-questions <qw> proves to be quite useful in determining the general tag for some

repetition requests.

Instances of repetition requests, some of which are shown with their surrounding

context, are seen in Example 188 through Example 195:

low number the - the nonvoiced um

phones .

1272.600-1274.520 c0 qy^d^f^g^rt right ?

1274.520-1276.440 c0 s and then renormalize .

Example 187: Bro016

264.902-267.287 c4 s i mean y- - don't want to do this over a

hundred different things that they've

tried .

267.287-268.822 c4 s but you know for some version that you

say is a good one .

268.822-270.356 c4 qy^d^f^g you know ?

273.619-279.864 c4 qw how - how much uh does it improve if

you actually adjust that ?

284.961-288.832 c4 s but it is interesting .

xample 188: Bed003

1291.740-1300.550 c1 fh|qw^rt um | how long would it take to - to add

another node on the observatory and

um - play around with it ?

1301.430-1302.290 c3 qw^br^rt another node on what ?

Example 189: Bed003

1352.970-1355.790 c2 qy^bu^rt also - you know - didn't we have a size

as one ?

 Understanding Check <bu>

The understanding check tag marks when a speaker checks to see if he understands

what a previous speaker said or else to see if he understands some sort of information.

With understanding checks, a speaker usually states what he is trying verify as correct

and follows that with a tag question <g>. Only the utterance, or portion of the utterance

if a pipe bar is used, containing the information to be verified is marked with the <bu>

tag. Tag questions <g> are not marked with the <bu> tag as they do not contain the

information that is to be verified.

1357.120-1357.350 c3 qw^br what ?

Example 190: Bed003

3146.860-3148.940 c3 qw so who would be the subject of this trial

run ?

3149.670-3149.910 c1 qw^br pardon me ?

Example 191: Bmr018

2495.240-2495.770 c0 qw^br what did you say ?

Example 192: Bro015

365.840-366.470 c3 qw^br what was that again ?

Example 193: Bmr008

3114.260-3116.010 c8 qw what about doing it with just the single

channels ?

3117.010-3117.270 c2 qw^br^rt sorry ?

Example 194: Bmr005

2687.890-2688.970 c2 qw^rt how many meetings is that ?

2689.200-2689.640 c8 qw^br^rt what's that ?

Example 195: Bmr030

243.000-244.000 c1 qw how much memory does he have ?

244.000-245.000 c0 qy^br^d^rt i'm sorry ?

Understanding checks are often confused with repetition requests and summaries

<bs>. With a repetition request, a speaker is seeking to hear what another speaker said

again, whereas, with an understanding check, a speaker is seeking to verify if what he

is saying is indeed correct. With a summary, a speaker summarizes something that

was previously said and is not seeking any sort of verification of correctness.

Understanding checks in context are seen in Example 196 through Example 199:

xample 196: Bed003

1907.630-1909.300 c2 s there's a bayes net spec for - in x m l .

1909.400-1910.680 c3 qy^bu^rt he's - like this guy has ?

1910.780-1911.550 c3 qy^bu^d^g^rt the javabayes guy ?

Example 197: Bed011

1988.840-1994.600 c2 s i e uh - it's either uh - for sightseeing

for meeting people for running errands

or doing business .

2006.120-2010.250 c1 qy^bu^d so business is supposed to uh - be sort

of - it - like professional type stuff ?

2010.250-2012.320 c1 qy^d^g right ?

Example 198: Bed011

1504.790-1525.140 c2 s the reading task is a lot shorter .

1511.580-1516.010 c3 s.%-- and other than that yeah i guess we'll

just have to uh - listen ==

1516.010-1520.440 c3 s^bu although i guess it's only ten minutes

each .

1520.440-1520.670 c3 qy^d^g^rt right ?

Example 199: Bmr012

231.944-233.704 c2 qw^t3 i guess - what time do we have to

leave ?

234.144-234.774 c2 qy^bu^d^rt^t3 three thirty ?

5.9 Group 8: Restated Information

This group, as the name states, contains specific tags pertaining to information that has

been restated. The group is further divided into two subgroups: repetition and

correction.

REPETITION

 Repeat <r>

The repeat tag <r> is used when a speaker repeats himself. This often occurs in

response to repetition requests or else to place emphasis on a certain point.

In repeating himself, a speaker repeats all or part of one of his previous utterances.

However, in order for an utterance to be considered a repeat, it must be a repeat of an

utterance made at most a few seconds prior to the repeat. Also, the guidelines

regarding segmentation, as discussed in Section 2, are to be taken into consideration

so that utterances in which a speaker begins speaking and then starts over using the

same words are within the same utterance are not segmented and the pipe bar is not

employed so that the repeated portions are labeled as repeats.

It is not required that a speaker repeat himself verbatim in order for a utterance to be

marked with the repeat tag <r>. If a speaker repeats himself and the repeated

utterance differs by a small number of words yet approximates the original utterance,

the <r> tag may be used. However, the <r> tag is not to be used if a speaker alters an

utterance so much so that no obvious structural likeness can be seen. For instance, if a

speaker says, "my pen has run out of ink" and then says "my pen's run out," the second

statement can be considered a repeat of the first. However, if the speaker's second

utterance was instead "there's no ink in my pen," that utterance would not be

considered a repeat of the first.

Additionally, in repeating himself, a speaker's utterance marked as a repeat may contain

more speech in addition to what was repeated. For instance, if a speaker says, "I have

to leave at one," and then follows that utterance with "I have to go at one and make

some phone calls," the latter utterance is still considered a repeat despite the additional

information.

Repeats <r> are not to be confused with mimics <m>. As previously stated, a repeat

occurs when a speaker repeats his own utterance. A mimic occurs when a speaker

repeats another speaker's utterance. Repeats are also not to be confused with

summaries <bs> where a speaker summarizes his own utterances as many structural

differences occur between the summary and the information being summarized.

Repeats in context are seen in Example 200 through Example 202:

 Mimic <m>

The mimic tag marks when a speaker mimics another speaker's utterance, or portion of

another speaker's utterance.

As with repeats <r>, mimics do not have to be repeated verbatim in order to be

considered mimics. This condition is discussed in the tag description for repeats <r>.

Also, if a speaker's utterance is marked as a mimic, it may contain more speech in

addition to what is mimicked. For instance, if one speaker says, "there's a problem with

the phone system," and then another speaker follows that utterance with "there's a

problem with the phone system concerning what aspect?," the latter utterance would

still be considered a mimic despite the additional speech.

Mimics are often forms of acknowledgments <bk> and, when such is the case, are

labeled in conjunction with the <bk> tag. The most common scenario when a mimic is a

form of acknowledgment occurs as a speaker who has the floor is talking and another

speaker acknowledges the speaker who has the floor by mimicking part of what he

says.

xample 200: Bro017

1821.640-1822.990 c1 s and hev- - everything is fixed .

1822.990-1823.950 c1 s^r everything is fixed .

Example 201: Bro017

1827.470-1828.860 c1 s for both - you would have to do .

1829.110-1829.720 c5 s^bu^m you would do it on both .

1829.560-1829.720 c1 s^aa yeah .

1829.720-1830.390 c5 s.%- so you'd actually ==

1829.830-1830.870 c1 s^r you have to do bo- - both .

Example 202: Bro025

870.243-872.737 c1 qy^bu^d^rt and there didn't seem to be any uh

penalty for that ?

873.030-873.386 c2 qy^br^rt pardon ?

873.390-876.620 c1 qy^bu^d^r^rt there didn't seem to be any penalty for

making it causal ?

In other cases, a speaker will mimic another speaker and phrase the mimic in the form

of a declarative question as a request for more information about what they mimicked.

For instance, if a speaker's utterance is "I went to the restaurant" and another speaker's

utterance in response is "the restaurant?", the response is a mimic of the first utterance

and acts as a request for more information about the restaurant.

Mimics <m> are not to be confused with repeats <r>. As previously stated, A mimic

occurs when a speaker repeats another speaker's utterance. A repeat occurs when a

speaker repeats his own utterance.

Also, mimics are not to be confused with summaries <bs> where a speaker summarizes

another speaker's utterances as many structural differences occur between the

summary and the information being summarized.

Mimics in context are seen in Example 203 through Example 211:

xample 203: Bed003

1875.040-1875.550 c3 s^co^rt go up one .

1875.700-1876.410 c2 s^bk^m up one .

Example 204: Bed004

1567.700-1568.320 c4 qw what's tourbook ?

1569.180-1570.630 c1 s^m.%-- tourbook ==

Example 205: Bmr001

1700.790-1704.110 c8 s so - so they - they're going to - they're

going to have to make speaker

assignments or something like this .

1704.030-1705.880 c1 s^bk^m they're going to have to make speaker

assignments .

Example 206: Bmr001

878.126-878.426 c8 s^bc nine .

878.352-878.672 c1 s^bk^m nine .

Example 207: Bmr001

1043.710-1044.080 c8 s it's a pain .

1044.500-1044.810 c1 s^bk^m it's a pain .

 Summary <bs>

The <bs> tag marks when a speaker summarizes a previous utterance or discussion,

regardless of whose speech he is summarizing.

Summaries are not to be confused with understanding checks <bs>. Understanding

checks restate information for validation while summaries do not require validation.

Furthermore, a DA may not contain both the <bs> and <bu> tags.

Summaries are also not to be confused with repeats <r> and mimics <m>. The tag

descriptions for repeats and mimics detail how such might occur.

Summaries in context are seen in Example 212 and Example 213:

xam

le 208: Bmr005

1492.390-1495.610 c3 s i - i - i - i consider - i consider

acoustic events uh - the silent too .

1497.240-1497.860 c1 s^m silent .

Example 209: Bmr005

2785.520-2786.340 c8 s^na it's what we're aiming for .

2786.060-2786.970 c2 s^bk^m that we're aiming for .

Example 210: Bmr009

1963.930-1966.420 c3 s well you have a like techno speak

accent i think .

1965.700-1967.180 c0 qy^bu^d^m^rt a techno speak accent ?

Example 211: Bmr012

123.504-124.024 c3 s^cs california .

124.251-124.871 c4 s^bk^m california .

xample 212: Bro011

75.120-82.956 c3 fh|s^rt well - uh | first we discussed about

some of the points that i was

addressing in the mail i sent last week .

87.253-90.293 c3 s^rt about the um - well - the

downsampling problem .

91.763-94.322 c3 s uh - and about the fit- - uh the length

of the filters .

98.530-100.610 c1 qw^rt so what's the - w- - what was the

downsampling problem again ?

98.609-98.929 c3 %- so we had ==

100.610-101.180 c1 s i forget .

100.813-105.273 c3 s so the fact that there - there is no uh -

low pass filtering before the

downsampling .

107.394-113.682 c3 s there is because there is l d a filtering

but that's perhaps not uh - the best .

114.640-117.470 c1 s|s^aa depends what it's frequency

characteristic is | yeah .

117.680-119.610 c1 s^cs so you could do a - you could do a

stricter one .

118.240-118.580 c4 qy^rt^t3 is the system on ?

120.255-120.545 c1 s^am maybe .

122.143-125.083 c3 s.%-- so we discussed about this about the

um ==

125.550-126.740 c1 qy^rt was there any conclusion about that ?

128.482-129.032 c3 h|s^co^na^rt uh - | try it .

130.300-130.640 c1 s^bk i see .

135.230-140.890 c1 s^bs so again this is th- - this is the

downsampling uh - of the uh - the

feature vector stream .

Example 213: Bro017

539.307-543.396 c1 s so i mean uh - uh - add moderate

amount of noise to all data .

544.447-549.417 c1 s so that makes uh - th- - any additive

noise less addi- - less a- - a- -

effective .

549.417-549.737 c1 qy^d^g^rt right ?

549.550-549.870 c5 s^aa right .

549.957-552.487 c1 s.%-- because you already uh - had the

noise uh - in a ==

552.487-555.017 c1 s and it was working at the time .

555.017-557.032 c1 s.%-- it was kind of like one of these things

you know but ==

559.870-566.410 c1 s so well you know just take a - take a

spectrum and - and - and add of the

constant c to every - every value .

560.570-561.820 c5 s.%- well you're - you're basically y- ==

567.550-569.560 c5 s^bs so you're making all your training data

more uniform .

CORRECTION

 Correct Misspeaking <bc>

The <bc> tag is used when a speaker corrects another speaker's utterance.

Corrections are based upon whether the word choice of a speaker is corrected or the

pronunciation of a word is corrected.

Instances in which the correct misspeaking tag <bc> are used are shown in context in

Example 214 through Example 217:

 Self-Correct Misspeaking <bsc>

The <bsc> tag marks when a speaker corrects his own error, with regard to either

pronunciation or word choice.

Segmentation is an issue regarding the <bsc> tag. As with repeats, a speaker may

begin an utterance and correct himself within the same utterance. In such cases, the

utterance is not segmented and the pipe bar is not employed to mark the <bsc> tag.

Section 2 details the guidelines surrounding how and why utterances are segmented.

xample 214: Bro012

1221.540-1225.420 c5 s^ar|s^rt oh no | i've ninety four .

1218.660-1219.640 c1 s^bc ninety three point six four .

Example 215: Bed012

2122.730-2124.280 c2 s^j^2 killing machines !

2125.890-2126.880 c1 s^bc reasoning machines .

Example 216: Bmr011

3098.000-3100.000 c6 s native speaking native speaking

english .

3100.000-3102.000 c7 s^bc i bet he meant native speaking

american .

Example 217: Bmr011

1308.000-1309.000 c1 s^rt and there we're already using fourteen .

1309.000-1311.000 c7 s^bc and we actually only have fifteen .

Instances in which the self-correct misspeaking tag <bsc> are used are shown in

context in Example 218 through Example 223:

xample 218: Bed003

567.066-574.026 c3 s^bk|s okay | so - yeah so note the four nodes

down there the - sort of the things that

are not directly extracted .

574.316-575.176 c3 s^bsc actually the five things .

Example 219: Bed003

1013.070-1013.210 c3 s^aa yeah .

1013.260-1013.420 c3 s^ar^bsc no .

Example 220: Bmr009

301.025-303.500 c2 fh|s um and uh | they don't look very

separate .

303.750-305.600 c2 fh|s^bsc uh | separated .

Example 221: Bmr013

1632.080-1632.920 c8 s^rt.%-- well we did the hand ==

1632.920-1633.760 c8 s^bsc the one by hand .

Example 222: Bmr024

653.072-659.242 c5 h|s.%--

h so | we have a whole bunch of digits

that we've read and we have the forms

and so on um but only a small number

of that ha- ==

659.384-660.524 c5 s^bsc well not a small number .

Example 223: Bmr018

507.485-508.498 c4 s^e and you're tightening the boundaries .

508.498-509.51 c4 s^bsc correcting the boundaries .

5.10 Group 9: Supportive Functions

This group contains tags that apply to utterances in which a speaker supports his own

argument by defending himself, offering an explanation, or else offering additional

details and utterances in which a speaker attempts to support another speaker by

finishing the other speaker's utterance.

 Defending/Explanation <df>

The <df> tag marks cases in which a speaker defends his own point or offers an

explanation. Often, the word "because" signals an explanation.

The <df> tag is often confused with the elaboration tag <e>. The two tags differ in that,

as the <df> tag marks utterances in which a speaker defends a point or offers an

explanation, the <e> tag marks utterances in which a speaker offers further details.

Example 224 through Example 229 present instances of the <df> tag in context:

xample 224: Bmr005

949.459-951.044 c4 s^ar no no it isn't sensitive at all .

951.044-951.837 c4 s^df i was just - i was jus- - i was

overreacting just because we've been

talking about it .

Example 225: Bmr005

1012.960-1019.350 c4 s^arp but i - i mean - i think also to some

extent its just educating the human

subjects people in a way .

1019.350-1022.540 c4 s^df because there's if uh - you know -

there's court transcripts there's -

there's transcripts of radio shows .

Example 226: Bmr007

14.467-15.967 cB qy^rt and uh shall i go ahead and do some

digits ?

16.724-17.504 c3 h|s^ng uh | we were going to do that at the

end .

17.504-18.284 c3 qy^d^rt remember ?

18.700-19.840 cB s^bk|s okay | whatever you want .

 Elaboration <e>

The elaboration tag marks when a current speaker elaborates on a previous utterance

of his by adding further details as opposed to simply continuing to speak on the same

topic. When a speaker describes something using an example, the example is

regarded as an elaboration.

The elaboration tag is often confused with the defending/explanation tag <df> which

marks utterances in which a speaker defends a point or offers an explanation. As the

defending/explanation tag revolves around reasons, the elaboration tag revolves around

details.

A convention has been established in handling instances when a question is followed by

an elaboration <e> which requires its own line. In such cases, the following elaboration

could be considered a declarative form of the question. Instead, the elaboration

receives a DA of <s^e>, along with any other necessary specific tags. The reasoning

behind labeling an elaboration following a question as a statement <s> rather than a

20.396-23.856 c3 s^co^df just - just to be consistent from here on

in at least that - that we'll do it at the

end .

Example 227: Bmr009

459.997-463.620 c2 s but i had maybe made it too

omplicated by suggesting early on that

you look at scatter plots .

463.620-467.244 c2 s^df because that's looking at a distribution

in two dimensions .

Example 228: Bro008

1356.660-1357.940 c4 s^na yeah because a lot of time that's true .

1357.940-1366.720 c4 s^df there were a lot of times when we

would try something and it didn't work

right away even though we had an

ntuition that there should be something

there .

Example 229: Bro015

449.830-450.490 c0 s^nd this week i haven't .

450.490-453.980 c0 s^df^ng i've been - my whole time's been taken

up with uh meeting recorder stuff .

question is that, if the elaboration were to be considered a question, then the

elaboration itself would be asking something. For instance, if a speaker were to ask,

"have you gone to that restaurant I suggested?", and then followed that question with an

elaboration such as "the one on Sixth Street," labeling the elaboration as a type of

question would indicate that the elaboration, "the one on Sixth Street," was actually

eliciting some sort of answer. Instead, the question, "have you gone to that restaurant I

suggested?", seeks an answer and the elaboration, "the one on Sixth Street," merely

adds a detail to the question without actually asking something.

Elaborations are shown in context in Example 230 through Example 237:

xample 230: Bed011

1516.010-1520.440 c3 s^bu although i guess it's only ten minutes

each .

1520.440-1520.670 c3 qy^d^g^rt right ?

1521.030-1521.480 c3 s^e roughly .

Example 231: Bro004

1179.080-1185.130 c1 qw well what was - is that i- - what was it

that you had done last week when you

showed - do you remember ?

1185.310-1188.230 c1 s^e^rt wh- - when you showed me the - your

table last week .

Example 232: Bmr024

1424.290-1427.230 c5 fg|s^df well but - but | i put it under the same

directory tree .

1427.230-1429.620 c5 fh|s^e you know | it's in user doctor speech

data m r .

Example 233: Bro004

2028.080-2038.300 c3 s^cs so uh - we were thinking about is

perhaps um - one way to solve this

problem is increase the number of

outputs of the neural networks .

2040.010-2044.450 c3 s^e.%-- doing something like um - um -

phonemes within context and ==

Example 234: Bro004

2170.080-2175.840 c3 s and basically the net- - network is

trained almost to give binary decisions .

 Collaborative Completion <2>

The collaborative completion tag <2> tag marks utterances in which a speaker attempts

to complete a portion of another speaker's utterance. Whether the speaker whose

utterance is completed by another speaker agrees with the content of the completion is

inconsequential. If a speaker does agree with the completion, then the agreement is

marked with the appropriate tag.

2177.730-2181.920 c3 s^e and uh - binary decisions about

phonemes .

Example 235: Bro004

2261.170-2264.060 c3 s so you - you have more information in

your features .

2264.060-2272.160 c3 s^e so um - you have more information in

the uh - posterior spectrum .

Example 236: Bro011

546.896-555.660 c1 fh|s^co^t^tc so um - | i suggest actually now we -

we - we sort of move on and - and

hear what's - what's - what's

happening in - in other areas .

555.660-562.490 c1 s^e^t like what's - what's happening with

your investigations about echos and so

on .

Example 237: Bro011

1471.250-1476.140 c1 fh|s and uh - | because in the ideal case we

would be going for posterior

probabilities .

1476.140-1481.030 c1 s^e if we had uh - enough data to really get

posterior probabilities .

1481.430-1486.460 c1 s^e and if the - if we also had enough data

so that it was representative of the test

data .

1486.460-1491.500 c1 s^e then we would in fact be doing the right

thing to train everything as hard as we

can .

In some cases, a speaker attempts to complete another speaker's utterance and, in

doing so, interrupts and stops the speaker whose utterance he is trying to complete.

The interrupted speaker then resumes speaking, usually having either accepted or

rejected the collaborative completion. If the collaborative completion is accepted, the

tags <aa>, <na>, and <aap> are used to characterize the acceptance. Acceptance of a

collaborative completion usually arises in the form of a "yes" word, as those labeled with

the <aa> tag, or else by mimicking the completion, and such is marked with the <na>

tag. If the collaborative completion is rejected, the tags <ar>, <nd>, <ng>, and <arp>

are used to characterize the rejection. Rejection of a collaborative completion usually

arises in the form of a "no" word, as those labeled with the <ar> tag, or else by a

speaker completing his utterance in a manner which differs from the collaborative

completion, and such is marked with either the <nd> or <ng> tag.

Collaborative completions in context are seen in Example 238 through Example 245:

xample 238: Bed003

463.416-469.753 c2 s.%- because we were thinking uh - if they

were in a hurry there'd be less likely to

- like - or th- ==

469.220-469.780 c3 s^2 want to do vista .

Example 239: Bed003

593.810-599.330 c3 s that kind of thing is all uh - sort of -

you know - probabilistically depends on

the other things .

598.030-599.260 c4 qy^bu^d^rt^2 inferred from the other ones ?

Example 240: Bmr007

1652.350-1654.960 cB s well but from the acoustic point of view

it's all good .

1655.120-1655.620 c4 s^aa^2 is the same .

Example 241: Bmr009

1937.990-1941.720 c3 s.%-- i think originally it was north -

northwest but ==

1941.420-1941.930 c0 s^2 northwest .

Example 242: Bmr012

435.384-437.674 c2 s.%- but there's a significant amount of ==

436.608-437.368 c5 qy^d^rt^2 non zero ?

5.11 Group 10: Politeness Mechanisms

This group contains tags that apply to utterances in which speakers exhibit

courteousness.

 Downplayer <bd>

The downplayer tag <bd> marks cases in which a speaker downplays or de-

emphasizes another utterance. The utterance that is downplayed may be uttered by the

same speaker or a different speaker.

Apologies, compliments, and other courteous utterances are often downplayed. In

other cases, a speaker makes a strong assertion and then downplays it.

Downplayers vary in form. Some may be long utterances and others may be quite

short. The following is a list of common short downplayers: "that's okay," "that's all

right," "it's okay," "I'm kidding," "it's just a thought," and "never mind."

Downplayers in context are presented in Example 246 through Example 252:

xample 243: Bmr012

1825.930-1828.470 c2 s but i d- - i know the lapel is really

suboptimal .

1827.450-1827.910 c4 qy^rt^2 is awful ?

Example 244: Bro004

1462.620-1472.340 c3 s^e the uh - the um - networks are trained

with noise from aurora - t i digits .

1471.470-1471.880 c4 s^2 aurora two .

Example 245: Bmr008

177.000-180.000 c1 qw how fine a resolution do you need on

that for this ?

181.000-182.000 c2 s^2 is the question .

xample 246: Bmr012

960.050-960.790 c8 s^ba congratulations .

961.254-964.724 c2 s^bd well it was i mean - i really didn't do

this myself .

Example 247: Bmr005

954.368-958.498 c1 s^t i - i came up with something from the

human subjects people that i wanted to

mention .

958.498-959.743 c1 s^bd i mean it fits into the area of the

mundane .

Example 248: Bed006

1953.730-1954.170 c1 s^fa sorry .

1955.080-1955.380 cA s^bd it's okay .

Example 249: Bro018

501.447-503.797 c2 s but suppose you don't really know what

the right thing is .

504.377-508.497 c2 s and that's what these sort of dumb

machine learning methods are good at .

510.950-511.540 c2 s^bd it's just a thought .

Example 250: Bmr011

2778.000-2779.000 c0 s.%-- and then the other thing is ==

2780.000-2781.500 c0 s^bd i don't know if this is at all useful .

Example 251: Bmr029

1232.580-1238.270 c2 s.%-- the - the other difference that we'd

have to take care of is that ==

1238.270-1242.430 c2 fh|s uh - | yeah we - we don't have a mike

that uh is particular to a person .

1242.430-1244.510 c2 s and so we'll have to do some

clustering .

1244.510-1249.770 c2 s and that'll be another another uh issue

too .

1252.160-1253.810 c2 s^bd but it - it - i could be wrong .

 Sympathy <by>

The <by> tag marks utterances in which a speaker exhibits sympathy. Oftentimes, the

phrase "I'm sorry" is used sympathetically. However, that very phrase also has the

potential to be marked as a repetition request or as an apology <fa>, depending

upon its function.

Instances of the <by> tag in context are displayed in Example 253 through Example

255:

 Apology <fa>

An utterance is marked as an apology <fa> when a speaker apologizes for something

he did (e.g., after coughing, sneezing, interrupting another speaker, etc.).

The phrase "I'm sorry," depending upon its usage, may be interpreted as a repetition

request or as sympathy <by>.

xample 252:

Bro010

631.950-633.005 c2 s so you would think as long as it's under

half a second or something .

633.005-633.533 c2 s^bd uh i'm not an expert on that .

xample 253: Bed003

3033.120-3034.070 c1 s^rt so i had to reboot .

3033.440-3034.140 c4 s^by^fe^rt oh no .

Example 254: Bmr027

1972.740-1977.040 c0 s and then you can see here g p s was

misinterpreted .

1977.450-1978.850 c0 s^by.%-- it's just totally understanda- ==

Example 255: Bmr027

2186.760-2189.800 c3 s.%-- without thinking about it when i offered

up my hard drive last week ==

2189.260-2190.040 c5 s^by^fe oh no !

Additionally, the phrase "excuse me" can be used as an apology <fa> or else can be

found within a suggestion <cs>. The phrase is found within a suggestion when it

precedes something for which a speaker is negotiating permission (Jurafsky 35).

Apologies <fa>, some of which are in context, are shown in Example 256 through

Example 261:

xample 256: Bmr001

876.821-877.541 c1 s so we could have eight .

876.899-877.029 c8 s^aa yeah .

878.126-878.426 c8 s^bc nine .

878.352-878.672 c1 s^bk^m nine .

878.672-879.432 c1 s^fa|s^r excuse me | nine .

Example 257: Bmr005

832.753-837.990 c5 s^fa sorry to interrupt .

Example 258: Bmr009

1563.000-1566.500 c0 s.%-- because the date is when you actually

read the digits and the time and ==

1566.500-1568.250 c0 s^fa excuse me .

1568.250-1570.000 c0 s^bsc the time is when you actually read the

digits but i'm filling out the date

beforehand .

Example 259: Bmr018

217.760-219.630 c1 s^fa he's - i - i'm sorry i should have

forwarded that along .

Example 260: Bmr026

1202.170-1203.530 c3 s^fa oh i'm sorry i misunderstood .

Example 261: Bmr006

1202.100-1205.320 c9 s^fa sorry i- have to - sorry i have to leave .

 Thanks <ft>

The <ft> tag marks utterances in which a speaker thanks another speaker.

Instances of the <ft> tag, one of which with surrounding context, are shown in Example

2629 through Example 264:

 Welcome <fw>

The <fw> tag marks utterances which function as responses to utterances marked with

the thanks tag <ft>. Phrases such as "you're welcome" and "my pleasure" are marked

with the welcome tag <fw>.

No instances of the <fw> tag exist within the Meeting Recorder data.

5.12 Group 11: Further Descriptions

This group contains various tags that do not fit into any of the pre-established groups.

The tags within this group characterize meeting agendas, changes in topic, exclamatory

material, humorous matter, self talk, third party talk, as well as syntactic and prosodic

features of utterances.

9 Regarding the use of the tag <sj> in Example 262, refer to footnote 7.

xample 262: Bed003

216.310-217.340 c4 sj^ba nice coinage .

219.833-220.463 c2 s^ft thank you .

Example 263: Bmr007

3266.710-3267.720 c8 s^ft thanks .

3267.810-3268.270 c8 s^ft appreciate that .

Example 264: Bmr024

2928.220-2929.450 c3 s^ft thank you for the box .

 Exclamation <fe>

The <fe> tag marks utterances in which a speaker expresses excitement, surprise, or

enthusiasm. Utterances marked with the <fe> tag, excluding quotes, are punctuated

with an exclamation mark < ! > within the transcript.

Utterances marked with the <fe> tag can range from consisting of one word to a lengthy

string of words. The most salient factor in determining if an utterance is an exclamation

is the level of energy. Exclamations usually have a much higher energy than that of the

surrounding utterances.

Instances of the <fe> tag are seen in Example 265 through Example 279:

xample 265: Bed003

47.760-47.920 c3 s^fe wow !

Example 266: Bed003

119.945-120.205 c2 s^fe aha !

Example 267: Bed003

1626.000-1626.240 c4 s^fe whew !

Example 268: Bed003

1676.950-1677.070 c2 s^fe oops !

Example 269: Bed003

1761.080-1761.190 c4 s^fe god !

Example 270: Bed003

1794.550-1794.750 c2 s^fe oh !

Example 271: Bed003

2004.230-2004.480 c3 s^fe ha !

Example 272: Bed004

3200.900-3201.260 c2 s^fe oh yeah !

 About-Task <t>

The about-task tag marks utterances that are in reference to meeting agendas or else

address the direction of meeting conversations with regard to meeting agendas.

The about-task tag is not to be confused with the topic change tag <tc>. The topic

change tag marks utterances which either end or begin a topic regardless of a meeting

agenda. The about-task tag marks utterances which regard previously established

items to be discussed or managed within a meeting. However, this is not to say that an

utterance can only be marked by either the about-task tag or the topic change tag.

Rather, both tags may be used to label an utterance so long as an utterance is

changing a topic in reference to a meeting agenda. For instance, if a speaker is talking

about a topic that is not part of the meeting agenda and then he or another speaker

changes the topic and mentions the agenda, then the utterance in which the change in

xample 273: Bmr009

2394.570-2396.130 c0 s^fe oh no !

Example 274: Bed003

133.711-134.431 c4 s^fe^j i can read !

Example 275: Bmr005

1956.430-1962.910 c4 s^fe^m twelve minutes !

Example 276: Bmr008

3293.420-3294.600 c3 s^fe^t3 oh it's seventy five per cent !

Example 277: Bed006

2876.320-2877.010 cA s^fe^j damn this project !

Example 278: Bro012

3213.110-3215.050 c0 s^fe^rt then do some more spectral

subtraction !

Example 279: Bmr015

525.983-527.896 c0 s^ba^fe so that's amazing you showed up at

this meeting !

topic and reference to the agenda occurred would be marked with the tags <t> and

<tc>.

Additionally, a restriction applies to the usage of the about-task tag. The about-task tag

is used to mark utterances which mention agendas and agenda items. In essence, the

about-task tag marks utterances which revolve around what tasks are to be completed

within the course of a meeting. So what is marked with the about-task tag is what is to

be accomplished within a meeting, but when an agenda item is in the process of being

"accomplished," it is not marked by the about-task tag. For instance, if a speaker

mentions that an agenda item is to discuss a certain subject and then other speakers

begin to discuss that subject, then the utterance mentioning that the agenda item to

discuss a subject is marked with the about-task tag. However, the actual discussion

about the subject is not marked with the about-task tag.

Example 280 through Example 289 display instances in which the about-task tag is

used:

xample 280: Bmr005

381.017-383.717 c4 s^t um - so i - i do have a - an agenda

suggestion .

Example 281: Bmr006

1224.410-1229.080 c3 fh|s^t^tc and | then um i guess another topic

would be where are we in the whole

disk resources question .

Example 282: Bmr006

4464.590-4466.090 c3 s^co^t^tc let's do digits .

Example 283: Bmr007

1938.400-1941.590 c3 s^t^tc speaking of taking control you said you

had some research to talk about .

Example 284: Bmr008

15.000-18.000 c1 s^co^rt^t let's discuss agenda items .

Example 285: Bmr010

239.005-242.305 c6 qh^t^tc so yeah why don't we do the speech

nonspeech discussion ?

100

 Topic Change <tc>

The <tc> tag marks utterances which either begin or end a topic. As the <tc> tag marks

when a topic changes, once the topic has indeed changed and a new topic is in the

course of discussion, the discussion of the new topic is not marked with the <tc> tag.

Oftentimes, a speaker will utter a floor grabber <fg> and then introduce a new topic. As

the floor grabber appears as though it is used as a mechanism to gain the floor and

introduce a new topic, and in effect signals a change in topic, it is not marked with the

<tc> tag. Rather, only utterances which convey a change in topic are marked with the

<tc> tag. In which case, a speaker must specify in his utterance that he wishes to end a

topic or else he must state that he wishes to begin a new topic either by initiating and

specifying a new topic or else by merely stating that he wishes to talk about something

else.

The <tc> tag may be used in conjunction with the about-task tag <t>. The tag

description for the about-task tag details the rules governing such usage.

Topic changes, some of which with surrounding context, are shown in Example 290

through Example 296:

xample 286: Bmr012

209.361-211.781 c4 qy^cs^rt^t okay so should we do agenda items ?

Example 287: Bmr012

219.415-223.365 c4 s^t uh - well i have - i want to talk about

new microphones and wireless stuff .

Example 288: Bmr014

51.589-52.929 c8 qo^t any agenda items today ?

53.672-61.382 c4 s^t i want to talk a little bit about getting -

how we're going to to get people to edit

bleeps parts of the meeting that they

don't want to include .

Example 289: Bro022

35.044-41.771 c0 qy^cs^rt^t^tc so should we just do the same kind of

deal where we go around and do uh

status report kind of things ?

101

xample 290: Bro0

713.450-713.910 c3 fg let's see .

715.580-725.090 c3 fh|s^cs^t^tc um | why don't - why don't we uh - if

there aren't any other major things why

don't we do the digits and then - then

uh - turn the mikes off .

Example 291: Bro007

1770.390-1776.060 c1 s^co^t^tc k uh - if nobody has anything else

maybe we should go around do - do

our digits - do our digits duty .

Example 292: Bmr008

2697.000-2698.000 c3 s^t^tc okay enough on forms .

Example 293: Bro004

3756.280-3766.420 c1 s^co^t^tc so with that maybe we should uh - go

to our digit recitation task .

Example 294: Bro013

1899.320-1899.750 c0 fg okay .

1902.920-1905.180 c0 fh|s^tc um | i think we're sort of done .

Example 295: Bro013

691.240-691.550 c0 fg okay .

691.680-692.500 c0 s^tc that was that topic .

692.500-693.140 c0 qw^t^tc what else we got ?

Example 296: Bro015

96.560-99.450 c3 s anyway hynek will be here next week

and maybe he'll know more about it .

105.440-105.990 c2 fg oh yeah .

106.680-111.530 c2 s^tc well the news more specifically t- - for

aurora .

111.530-112.450 c2 fh um ==

113.880-121.622 c2 s so i guess there was again a

conference call but uh they are not

decide on everything yet .

102

 Joke <j>

The <j> tag marks utterances of humorous or sarcastic nature. If a speaker is

attempting to be humorous, then the utterances containing humorous material are

marked with the <j> tag, regardless of how those utterances received by other

speakers.

Utterances marked with the <j> tag are often context dependent, in that jokes are often

made with regard to the current topic at hand. A majority of jokes require the

surrounding context in order to be perceived as jokes, as when jokes are seen without

surrounding context, they usually tend not to appear as being humorous or sarcastic.

Example 297 through Example 301 display jokes with surrounding context:

xample 297: Bro021

1877.030-1878.270 c5 qw^rt what - what is v t s again ?

1878.070-1881.140 c4 s uh vectorial taylor series .

1880.420-1881.070 c5 s^bk oh yes .

1881.070-1881.710 c5 s^aa right right .

1882.530-1885.350 c5 s i think i ask you that every single

meeting .

1885.350-1886.750 c5 qy^g don't i ?

1884.860-1885.590 c4 qw^br what ?

1886.750-1888.160 c5 s i ask you that question every meeting .

1887.310-1888.120 c4 s^aa yeah .

1888.080-1890.790 c1 s^j so that'd be good from - for analysis .

1890.790-1892.140 c1 s^df^j it's good to have some uh cases of the

same utterance at different - different

times .

1891.680-1893.200 c5 s^bk yeah .

1893.200-1894.720 c5 qw^j what is v t s ?

Example 298: Bro017

2173.380-2175.970 c1 s^cs.%-- but what you can do - i'm confident we

ca- ==

2175.970-2178.550 c1 s well i'm reasonably confident and i

putting it on the record .

2178.550-2178.730 c1 qy^d^f^rt right ?

2178.730-2183.790 c1 s^j i mean y- - people will listen to it for -

for centuries now .

Example 299: Bro016

1386.190-1388.280 c5 qy do you have speaker information ?

103

 Self Talk <t1>

The <t1> tag is used when a speaker talks to himself. Often, utterances marked as self

talk are quieter and softer than the surrounding speech.

A case in which the self talk tag is used occurs when a speaker is writing something

down and consequently repeats what he writes to himself. In other instances, a

speaker may be attempting to make some sort of a calculation or solve a problem and

talks to himself in the process of figuring out the answer.

Although it has been mentioned that certain types of utterances, such as backchannels

and floor holders <fh>, are not forms of direct communication between speakers,

these utterances are not considered self talk either.

Example 302 through Example 305 display instances of the self talk tag, most of which

are shown with surrounding context.

1388.930-1393.370 c4 s^j social security number .

1389.800-1392.410 c5 s^ba that would be good .

1391.980-1395.370 c1 s like we have male female .

1392.410-1394.130 c5 s^j bank pin .

Example 300: Bro014

8.347-9.712 c1 fg okay .

9.712-11.077 c1 qy^j^rt did you solve speech recognition last

week ?

Example 301: Bro014

40.831-41.701 c2 qy^rt is he going to come here ?

42.154-44.306 c1 h uh ==

44.306-45.382 c1 s^j^na well we'll drag him here .

45.382-46.458 c1 s^j i know where he is .

xample 302: Bmr007

787.674-792.891 c8 s.%-- in that case um my c- the coding that i

was using - since we haven't uh

incorporated adam's uh coding of

overlap yets the coding of ==

792.891-798.109 c8 s^t1 yeah yets is not a word .

104

 Third Party Talk <t3>

The third party talk tag marks utterances of side conversations. Side conversations are

conversations which are not directed toward the main conversation and may only

consist of a handful of utterances or may be quite lengthy.

Instances of third party talk are shown in Example 306 through Example 309 with

surrounding context.

xample 303: Bro018

2987.260-2989.580 c2 s.%-- i - i - i th- - i think he ==

2991.360-2992.210 c2 qo^t1 what am i saying here ?

Example 304: Bro014

50.154-51.928 c4 s^t1 doo doo doo .

53.633-54.207 c4 s^t1 doo doo .

Example 305: Bro021

2230.830-2235.540 c1 fh|s.%-- uh - | so that's log of x plus log of one

plus uh ==

2236.170-2236.760 c1 fh well .

2237.360-2238.270 c1 qy^rt^t1 is that right ?

2238.270-2239.180 c1 s^e^t1.%-- log of ==

2238.710-2240.560 c3 s^t1 one plus n by x .

xample 306: Bmr007

1389.340-1394.230 cA s so so - actually um that's in part

because the nodding - if you have

visual contact the nodding has the

same function .

1394.230-1399.120 cA s but on the phone in switchboard you -

you - that wouldn't work .

1398.900-1399.680 cB s^na yeah you don't have it .

1399.120-1401.260 cA s so so you need to use the

backchannel .

1401.140-1405.880 cB s^t3.%-- your mike is ==

1403.000-1410.570 c0 qy^r^rt so in the two person conversations

when there's backchannel is there a

105

great deal of overlap in the speech ?

1405.880-1410.630 cB s^co^t3 that is an earphone so if you just put it

so it's on your ear .

1410.570-1411.000 c0 qrr.%-- or ?==

1411.000-1417.160 c0 s because my impression is sometimes it

happens when there's a pause .

1411.170-1411.450 c1 s^aa yes .

1411.250-1411.660 cB s^t3 there you go .

1412.160-1412.380 c1 b yeah .

1412.630-1412.940 cB s^ft^t3 thank you .

Example 307: Bro004

1109.570-1111.640 c2 qy^d^rt^t3 these numbers are uh - ratio to

baseline ?

1110.650-1111.840 c1 qw.%- so i mean - wha- - what's the ?==

1111.840-1121.980 c1 qy^bu^d this - this chart - this table that we're

looking at is um - sho- - is all testing

for t i digits ?

1123.260-1126.910 c3 s^rt so you have uh - basically two uh -

parts .

1123.610-1123.880 c9 s^t3 bigger is worse .

1123.880-1125.290 c9 s^t3 this is error rate i think .

1125.570-1125.690 c9 s^ar^t3.% no no .

1125.640-1126.040 c2 s^t3 ratio .

1126.910-1130.580 c3 s^rt the upper part is for t i digits .

1130.580-1134.240 c3 s^rt and it's divided in three rows of four -

four rows each .

1128.380-1128.640 c9 s^aa^t3 yeah yeah yeah .

Example 308: Bro003

2159.050-2161.170 c0 qy^rt is that - was that distributed with

aurora ?

2161.170-2162.230 c0 qrr.%-- or ?==

2161.490-2161.730 c8 s.% italian .

2161.960-2163.020 c2 qr^bu^d^rt^t3 one l or two l's ?

Example 309: Bed012

998.980-1001.180 c1 s^rt and we get a certain - we have a

situation vector and a user vector and

everything is fine .

1001.540-1004.130 c1 %- an- - an- - and - and our - and

our ==

1002.750-1005.980 c2 qy^rt^t3 did you just sti- - did you just stick the

106

 Declarative Question <d>

The declarative question tag marks questions which have the syntactic appearance of a

statement. In declarative questions, the subject precedes the verb and subject-auxiliary

inversion and wh-movement do not occur. It is not uncommon for a rising tone <rt> to

be found on a declarative question, however a rising tone does not always function as

an indicator that a question is being asked.

Additionally, tag questions <g> are often declarative questions. This is only the case

when subject-auxiliary inversion does not occur (e.g., "you do?" rather than "do you?")

or if the question consists of only one word (e.g., "right?") or does not contain a verb

(e.g., "the tenth of July?"). However, if a question consists of one word and that word is

a "wh" word, such as those mentioned in the tag description for wh-questions <wh>,

then neither the tags <d> or <g> are used.

Declarative questions are seen in Example 310 through Example 324:

m- - the - the - the microphone

actually in the tea ?

1005.790-1008.320 c0 s^ar^t3 no .

1008.500-1009.530 c1 fh and um ==

1009.480-1010.290 c0 s^ng^t3 i'm not drinking tea .

1010.290-1011.100 c0 qw^t3 what are you talking about ?

1011.770-1012.260 c2 s^bk^t3 oh yeah .

1012.260-1012.750 c2 s^fa^t3 sorry .

1013.580-1017.780 c1 s^co^rt

et's just assume our bayes net just has

three decision nodes for the time

being .

xample 310: Bro021

979.242-980.846 c1 qy^d^g^rt right ?

Example 311: Bro013

2020.370-2020.610 c0 qy^d^f^g you know ?

Example 312: Bro021

2493.820-2495.190 c4 qy^d^g^rt no ?

Example 313: Bmr007

92.862-98.798 c3 fh|qo^d^rt

m | and anything else anyone wants to

talk about ?

107

xample 314: Bmr007

112.365-116.868 c3 fh|qo^d^rt um | and anything else ?

Example 315: Bmr007

117.088-118.018 c3 qo^d nothing else ?

Example 316: Bmr007

171.144-171.704 c0 qy^d^rt^2 same idea ?

Example 317: Bmr007

628.021-630.973 c3 qy^bu^d oh so the bottom three did have s- stuff

going on ?

Example 318: Bmr007

653.124-653.594 c3 qy^d you don't know ?

Example 319: Bmr021

342.000-343.000 c4 qy^bu^d^rt a wired one ?

Example 320: Bed006

2804.550-2807.290 c4 qy^bu^d^rt or you'd like - so you're saying you

could practically turn this structure

inside out ?

Example 321: Bmr024

929.052-930.972 c4 qy^d the references for - for those

segments ?

Example 322: Bmr024

1075.910-1081.850 c3 fg|qy^d^t^tc um | another one that we had on

adam's agenda that definitely involved

you was s- - something about

smartkom ?

Example 323: Bro017

2117.620-2122.540 c5 qy^d^rt

o that effectively the c one never really

108

 Tag Question <g>

A tag question follows a statement and is a short question seeking confirmation of that

statement. Tag questions receive a general tag of <qy> and are often used in

conjunction with the "follow me" tag and the declarative question tag <d>. The tag

description for declarative questions <d> discusses the instances in which it may be

used in conjunction with the tag <g>. Utterances preceding tag questions are labeled

as statements <s> rather than declarative yes/no questions <qy^d>.

Tag questions are often found following statements marked with the understanding

check tag <bu>.

Common utterances marked with the <g> tag include, but are not limited to, the

following: "right?", "yes?", "yeah?", "no?", "okay?", "isn't it?", "correct?", "won't it?",

"doesn't it?", and "you know?".

Tag questions in context are seen in Example 325 through Example 334:

contributes to the score ?

Example 324: Bro017

2487.900-2489.260 c5 qy^d^rt see how many cycles we used ?

xample 325: Bed011

2073.940-2074.690 c1 s^bu exchange money is an errand .

2074.690-2075.440 c1 qy^d^g right ?

Example 326: Bed003

407.887-409.477 c2 s so then our next idea was to add a

middle layer .

409.477-409.777 c2 qy^d^f^g right ?

Example 327: Bed003

1391.100-1398.880 c1 s in the sense that you know - if it's tom

- the house of tom cruise you know -

it's enterable but you may not enter it .

1399.230-1399.520 c1 qy^d^f^g^rt you know ?

109

xample 328: Bed003

2298.190-2301.170 c1 s:s and then the persons says um - yeah i

want to see it .

2302.210-2302.320 c1 qy^d^g yeah ?

Example 329: Bed004

3059.570-3065.040 c2 s there - the - the land- - the

construction implies the there's a con-

- this thing is being viewed as a

container .

3065.920-3066.250 c2 qy^d^f^g okay ?

Example 330: Bmr001

95.697-98.097 c8 s and this - this one is right at the end of

the table .

98.477-98.757 c8 qy^d^f^g okay ?

Example 331: Bmr005

1473.790-1474.370 c8 s^m that's a lot of overlap .

1474.370-1474.940 c8 qy^d^g^rt yeah ?

Example 332: Bmr001

1237.390-1238.960 c1 fg|s^bu yeah | so we don't store any of our

audio formats compressed in any way .

1238.960-1240.530 c1 qy^d^g do we ?

Example 333: Bmr005

1257.220-1260.490 c8 fg|s^bu well | you weren't talking about just

overlaps .

1260.490-1260.740 c8 qy^d^g^rt were you ?

Example 334: Bmr005

1763.010-1764.720 c2 fh|s i mean - | the normalization you do is

over the whole conversation .

1764.720-1766.490 c2 qy^g^rt isn't it ?

110

 Rising Tone <rt>

The rising tone tag is used to mark utterances in which a speaker's tone rises at the end

of his utterance. Rising tones at the end of utterances occur in both questions and

statements. Although intonation does not constitute a dialog act, the use of the <rt> tag

provides useful information for automatic speech recognition.

5.13 Group 12: Disruption Forms

As stated in Section 3.4, disruption forms are used to mark utterances that are

indecipherable, abandoned, or interrupted. Only one disruption form may be used per

utterance. Guidelines and restrictions surrounding the format and use of disruption

forms that are not mentioned in the tag descriptions for the indecipherable, interrupted,

abandoned, and nonspeech tags are found in Section 3.4.

Examples are not provided within the tag descriptions for the indecipherable,

interrupted, and nonspeech tags, as they require the corresponding audio portion in

order to convey why it is that an utterance is indecipherable, interrupted, abandoned, or

is considered nonspeech.

Additionally, Section 2 discusses segmentation and proves to be of much assistance in

using disruption forms.

 Indecipherable <%>

The indecipherable tag marks indecipherable speech such as mumbled or muffled

words or utterances that are too difficult to hear on account of the microphone picking

up sounds from breathing.

The indecipherable tag <%> is not to be confused with the nonspeech tag <x>. The

nonspeech tag <x> is used for sound segments which are silent or otherwise contain

non-vocal sounds such as doors slamming and phones ringing. The nonspeech tag

<x> does not apply to sounds such as breathing and sighs, as these are vocal sounds.

However, sounds such as coughing and sneezing may be considered vocal sounds but

are instead categorized with the nonspeech variety.

111

 Interrupted <%->

The interrupted tag marks incomplete utterances in which a speaker stops talking on

account of being interrupted by another speaker. This tag is not to be confused with the

abandoned tag <%--> which is used to mark instances in which a speaker intentionally

abandons an utterance.

As the most salient examples of the interrupted tag involve speakers giving up the floor

immediately, the interrupted tag is even used in cases in which a speaker has the floor

and is interrupted but does not immediately relinquish the floor. The reasoning behind

using the interrupted tag rather than the abandoned tag <%--> in such instances is

because the speaker gives up the floor on account of being interrupted.

 Abandoned <%-->

The abandoned tag marks utterances which are abandoned by a speaker. Abandoned

utterances occur when a speaker trails off or else chooses to either reformulate an

utterance or change the topic by abandoning his current utterance and beginning a new

one.

The issues mentioned in Section 2 regarding segmentation are of crucial importance

when using the abandoned tag. For instance, if a speaker begins an utterance and

restarts it in a different manner, and the prosody and pauses are such that the original

utterance and the restarted version constitute a single utterance, the entire utterance

remains intact and is labeled in a way that reflects its completeness. The utterance is

not split at the point between the beginning and the restarted portion, and the beginning

portion is not marked as being abandoned. In Example 335, an utterance is shown that

is restarted and remains intact, rather than being split at the region where it is restarted:

Abandoned utterances are seen with surrounding context in Example 336 through

Example 339:

xample 335: Bro021

1730.970-1733.270 c3 s and it - it - it gave like - i just got the

signal out .

xample 336: Bro021

186.057-194.998 c2 s well uh there is one thing that we can

observe is that the mean are more

different for - for c zero and c one than

for the other coefficients .

112

195.634-196.920 c2 fh and ==

198.663-199.323 c2 fh yeah .

200.819-203.469 c2 s.%-- and - yeah it - the c one is ==

203.469-215.256 c2 s there are strange - strange thing

happening with c one is that when you

have different kind of noises the mean

for the - the silence portion is - can be

different .

Example 337: Bro021

261.708-276.050 c2 fh|s^rt um | a third thing is um that instead of

t- - having a fixed time constant i try to

have a time constant that's smaller at

the beginning of the utterances .

276.050-279.990 c2 s^e to adapt more quickly to the r- -

something that's closer to the right

mean .

280.273-282.108 c2 fh t- - t- - um ==

283.723-286.491 c2 s^bk yeah .

286.491-287.875 c2 s and then this time constant increases .

287.875-289.259 c2 s.%-- and i have a threshold that ==

289.855-298.584 c2 s well if it's higher than a certain

threshold i keep it to this threshold to

still uh adapt um the mean when - if

the utterance is uh long enough to - to

continue to adapt after like one

second .

Example 338: Bro026

1235.390-1237.000 c3 qy^rt would - would that set on the handset ?

1237.000-1237.420 c3 qrr.%-- or ?==

Example 339: Bro025

118.800-127.061 c1 s^na yeah i mean it's - it's actually uh very

similar .

127.061-128.844 c1 s.%-- i mean if you look at databases ==

129.611-130.740 c1 fh uh ==

132.232-141.440 c1 s the uh one that has the smallest -

maller overall number is actually better

on the finnish and spanish .

142.317-147.387 c1 fh|s uh | but it is uh worse on the uh aurora .

145.334-146.817 c4 s^2.%-- it's worse on ==

147.387-151.000 c1 s^bsc i mean on the uh t i- - t i digits .

113

 Nonspeech <x>

The nonspeech tag marks any utterance that is unintelligible on account of non-vocal

noises such as doors slamming, phones ringing, and problems with a recording. The

nonspeech tag also marks coughing and sneezing sounds, as well as utterances filled

with silence.

The nonspeech tag is not to be confused with the indecipherable tag <%> which marks

utterances that are unintelligible on account of muffled speech, mumbling, breathing

sounds, and sighing.

5.14 Group 13: Nonlabeled

Group 13 solely contains the nonlabeled tag <z>. As stated in Section 3.2, the tag <z>

does not provide any information regarding the characteristics and functions of

utterances as the tags of the other groups do, and for this reason it is separated from

those groups.

 Nonlabeled <z>

The nonlabeled tag marks utterances that are not to be labeled with a DA. Types of

utterances that are not to be labeled are those containing to pre- or post-meeting

chatter, those pertaining to "bleeped" portions in the corresponding audio file, and those

pertaining to the reading of digits. The tag <z> marks utterances which otherwise would

be labeled with DAs but instead are intentionally not to be labeled.

An additional, but rare, instance in which the tag <z> is used arises when one speaker

wears multiple microphones, thus causing his utterances to be recorded on multiple

channels. In such a case, the speaker’s utterance on his original microphone (i.e. the

microphone he has been using throughout the meeting) receives the appropriate DA.

Subsequent channels with the same utterance are labeled with the tag <z> and receive

a note of “DUPLICATED-MICROPHONE” in the comment field.

As a side note, the convention of marking pre- and post-meeting chatter with the <z>

tag was a fairly recent development. In which case, a number of utterances which are

now marked with the <z> tag were originally marked with DAs consisting of the tags

found in Groups 1 through 12 along with adjacency pairs. As these original DAs have

been replaced with the <z> tag, the APs, however, have been preserved per chance

they are of use for future research. As the information derived from APs is optimized

with the use of corresponding DAs, APs corresponding to utterances marked with the

114

<z> tag can only provide optimal information upon being relabeled with DAs consisting

of the tags found in Groups 1 through 12.

115

APPENDIX 1: LABELED MEETING SAMPLE

A labeled five-minute portion of Bro021 is shown below. Included are start and end

times, channel numbers, DAs, adjacency pairs, and the corresponding portions of the

transcript.

1828.250-1832.820 c3 s i like plugged some groupings

for computing this eigen- - uh uh

uh s- -

values and eigenvectors .

1832.820-1839.250 c3 s so just - i just some small block

of things which i needed to put

together for the subspace

approach .

1839.250-1845.680 c3 s and i'm in the process of like

building up that stuff .

1846.670-1849.080 c3 fh and um ==

1850.400-1852.790 c3 fh uh - yeah .

1854.120-1856.580 c3 s i guess - yep i guess that's it .

1856.580-1859.040 c3 s and uh th- - th- - that's where i

am right now .

1859.620-1860.630 c3 fh so .

1861.560-1863.000 c5 qo^tc 1a oh how about you carmen ?

1862.830-1865.740 c4 s 1b huh i'm working with v t s .

1866.330-1869.160 c4 fh|s

m | i do several experiment with

the spanish database first .

1869.150-1873.400 c4 s^e 2a only with v t s and nothing more .

276.050-279.990 c2 s^e to adapt more quickly to the r- -

something that's closer to the

right mean .

1875.520-1876.580 c4 s^e no l d a .

1873.400-1875.520 c4 s^e not v a d .

1876.580-1877.640 c4 s^e nothing more .

1877.030-1878.270 c5 qw^rt 2b.3a what - what is v t s again ?

1878.070-1881.140 c4 s 3b.4a uh vectorial taylor series .

1878.320-1879.090 c3 %- new ==

1880.420-1881.070 c5 s^bk 4b oh yes .

1881.070-1881.710 c5 s^aa 4b+ right right .

1881.350-1883.060 c4 s to remove the noise too .

1882.530-1885.350 c5 s 5a i think i ask you that every single

meeting .

1885.350-1886.750 c5 qy^g 5a+ don't i ?

116

1884.860-1885.590 c4 qw^br 5b.6a what ?

1886.750-1888.160 c5 s 6b.7a i ask you that question every

meeting .

1887.310-1888.120 c4 s^aa 7b-1 yeah .

1888.120-1888.930 c4 %- if - well ==

1888.080-1890.790 c1 s^j 7b-2.8a so that'd be good from - for

analysis .

1890.790-1892.140 c1 s^df^j 7b-2+.8a+ it's good to have some uh cases

f the same utterance at different

- different times .

1892.140-1893.490 c1 fh yeah .

1891.680-1893.200 c5 s^bk 8b yeah .

1893.200-1894.720 c5 qw^j 8b+.9a what is v t s ?

1895.100-1896.260 c4 s^m 9b v t s .

1896.260-1897.410 c4 s.%-- i'm sor- ==

1897.410-1898.980 c4 s.%-- well um the question is that ==

1898.980-1900.540 c4 fh well .

1900.540-1903.300 c4 s remove some noise but not too

much .

1903.700-1909.290 c4 fh|s and | when we put the m- - m- -

the them - v a d the result is

better .

1909.290-1915.030 c4 s and we put everything the result

is better .

1915.030-1920.770 c4 s 10a but it's not better than the result

that we have without v t s .

1921.110-1921.780 c4 s^ar no no .

1923.210-1924.060 c1 s^bk 10b i see .

1924.060-1930.290 c1 s.%-- 11a so that given that you're using

the v a d also the effect of the

v t s is not so far ==

1929.630-1930.270 c4 s^na 11b is not .

1930.780-1934.640 c1 qw^rt 12a do you - how much of that do

you think is due to just the

particular implementation and

how much you're adjusting it ?

1934.640-1938.490 c1 qw.%-- 12a+ or how much do you think is

intrinsic to ?==

1936.770-1937.830 c4 s^no 12b pfft i don't know .

1937.830-1938.880 c4 s^df.%-- 12b+ because ==

1938.880-1940.500 c4 fh hhh ==

1939.210-1941.350 c2 qy 13a are you still using only the ten

first frame for noise estimation ?

1941.350-1943.490 c2 qrr.%-- or ?==

1944.260-1953.610 c4 h|s^rt 13b uh | i do the experiment using

117

only the f- - onl- - uh to use on- -

only one fair estimation of the

noise .

1944.890-1946.040 c2 qrr.%-- or i- ?==

1948.290-1948.820 c2 b yeah .

1949.670-1950.580 c2 b huh .

1953.610-1961.850 c4 s 13b+ and also i did some experiment

h doing um a lying estimation of

the noise .

1962.430-1965.860 c4 s.%-- and well it's a little bit better but

not ==

1966.550-1967.100 c4 x n- ==

1967.920-1969.610 c2 s^cs maybe you have to standardize

this thing also .

1970.450-1974.600 c2 s^df.%--

ecause all the thing that you are

testing use a different ==

1969.610-1970.450 c2 s^e noise estimation .

1975.430-1975.930 c4 b huh .

1975.490-1976.000 c3 b huh .

1975.780-1978.860 c2 s^df they all need some - some noise

- noise spectra .

1978.860-1981.940 c2 s^df but they use - every - all use a

different one .

1976.720-1979.030 c4 s^ar|s no | i do that two - t- - did two

time .

1982.310-1983.860 c1 s i have an idea .

1983.860-1985.620 c1 s.%-- if - if uh uh ==

1985.620-1986.500 c1 s^aa y- - you're right .

1986.500-1987.380 c1 s i mean each of these require

this .

1987.380-2000.980 c1 qw^cs um given that we're going to

have for this test at least of - uh

boundaries what if initially we

tart off by using known sections

of nonspeech for the

estimation ?

1999.540-2000.350 c4 b uhhuh .

1999.630-2000.020 c2 b uhhuh .

2003.140-2003.740 c1 qy^d^g^rt right ?

2003.740-2005.860 c1 fh s- - so e- - um ==

2003.760-2004.160 c2 b yeah .

2004.160-2004.570 c2 b uhhuh .

2005.860-2010.710 c1 s^df first place i mean even if

ultimately we wouldn't be given

the boundaries uh this would be

a good initial experiment to

118

separate out the effects of

things .

2010.710-2015.930 c1 qw

mean how much is the poor you

know relatively uh unhelpful

result that you're getting in t

is or

this or this ?

2015.930-2021.370 c1 qy is due to some inherent

imitation to the method for these

tasks ?

2021.370-2031.420 c1 qw and how much of it is just due to

he fact that you're not accurately

finding enough regions that - t

are really n- - noise ?

2028.600-2029.070 c3 b huh .

2030.230-2030.880 c4 b uhhuh .

2030.780-2031.490 c2 b uhhuh .

2032.080-2033.070 c1 fh um ==

2033.070-2037.980 c1 s^df 14a so maybe if you tested it using

that you'd have more reliable

tretches of nonspeech to do the

estimation from .

2037.980-2042.900 c1 s 14a+ and see if that helps .

2042.880-2045.120 c4 s^bk 14b yeah .

2045.120-2046.250 c4 s^tc another thing is the them - the

codebook .

2046.250-2047.370 c4 s^bsc the initial codebook .

2047.370-2049.380 c4 s.%-- that maybe ==

2049.380-2050.380 c4 s well it's too clean .

2050.380-2051.380 c4 fh and ==

2051.240-2051.980 c1 b uhhuh .

2051.380-2052.560 c4 s^df.%-- because it's a ==

2052.560-2053.150 c4 fh i don't know .

2053.150-2053.740 c4 s.%-- the methods ==

2054.740-2058.370 c4 s^cs 15a if you want you c- - i can say

something about the method .

2058.420-2059.090 c1 s^aa 15b uhhuh .

2059.380-2060.780 c4 s.%-- yeah in the ==

2065.040-2070.080 c4 s^df

ecause it's a little bit different of

the other method .

2071.310-2072.790 c4 s.%-- well we have ==

2073.710-2088.990 c4 s if this - if this is the noise signal

uh in the log domain we have

something like this .

2102.010-2103.390 c4 s now we have something like

this .

2103.390-2107.640 c4 s.%-- and the idea of these methods is

119

to n- - given a um ==

2107.640-2111.900 c4 qw how do you say ?

2108.620-2110.040 c1 b huh huh .

2111.900-2115.240 c4 s i will read because it's better for

my english .

2116.130-2117.780 c4 %-- i- - i- - given ==

2117.780-2120.610 c4 s is the estimate of the p d f of the

noise signal .

2120.610-2131.340 c4 s when we have a - um a statistic

of the clean speech and an

statistic of the noisy speech .

120

APPENDIX 2: UNUSED/MERGED SWBD-DAMSL TAGS

As indicated in Section 1.2, certain SWBD-DAMSL tags are not found in the MRDA

tagset. Of these tags, some have been merged with other tags and others are not

included in the MRDA tagset entirely. Below is a list of these tags. Each SWBD-

DAMSL tag listed below is followed by a brief description indicating whether it has been

merged or why it is not included in the MRDA tagset.

 About-communication <c>

Utterances such as "pardon me?" and "I can't hear you" that are marked with <c> in the

SWBD-DAMSL tagset are considered Repetition Requests in the MRDA tagset.

The tag is more specific in characterizing these utterances. Also, the <c> tag

marks utterances such as "I heard a laugh in the background" and "I think a train went

by" (Jurafsky et al. 1997). Such utterances generally do not tend to occur in the MRDA

meetings. Rather than generally address communication with the <c> tag, the tag

is implemented for specificity.

 Statement-non-opinion <sd> and Statement-opinion <sv>

The <sd> and <sv> tags were quite difficult to use with the MRDA data, as their use

resulted in a lack of agreement among annotators. They were eventually eliminated

from the MRDA tagset and replaced with the <s> tag, which marks statements in

general, without having to distinguish between "non-opinion" and "opinion." (For overt

opinions, the <ba> tag is used).

 Open-option <oo>

This tag is no longer included in the MRDA tagset due to its redundancy with

suggestions <cs>. Refer to Appendix 4 for more information.

 Conventional-opening <fp>

This tag is not included in MRDA tagset due to lack of use. Utterances that would be

marked with this tag usually occur in pre-meeting chatter, which is marked with the <z>

tag.

121

 Conventional-closing <fc>

This tag is not included in MRDA tagset due to lack of use. Utterances that would be

marked with this tag usually occur in post-meeting chatter, which is marked with the <z>

tag.

 Explicit-performative <fx>

This tag is no longer included in the MRDA tagset due to its lack of use. Refer to

Appendix 4 for more information.

 Other-forward-function <fo>

This tag is not included in MRDA tagset due to lack of use.

 Yes Answers <ny>

This tag has been merged with the SWBD-DAMSL tag <aa> to form the MRDA tag

<aa>.

 No Answers <nn>

This tag has been merged with the SWBD-DAMSL tag <ar> to form the MRDA tag <ar>.

 Quoted Material <q>

Due to the various DA tags quoted material within the MRDA data had the potential to

receive, the use of the SWBD-DAMSL tag <q> was replaced with a convention that

actually used DAs to characterize the quoted material. In doing so, more information

regarding the character and function of quoted material is gained than through using a

tag such as <q> to merely indicate that quoted material is present. Section 3.5 details

the treatment of quoted material.

 Hedge <h>

This tag is not included in the MRDA tagset due to lack of use and ambiguity as to what

sort of utterance would be labeled as a hedge as opposed to another label.

122

 Continued from Previous Line <+>

This tag is not included in the MRDA tagset because utterances continued from a

previous line by the same speaker are given a new DA to depict the function of the

continuation.

123

APPENDIX 3: UNIQUE MRDA TAGS

Due to the nature of the MRDA data, the SWBD-DAMSL tagset proved to be inefficient

in accurately characterizing all facets of the MRDA data. Consequently, tags were

created to account for areas where the SWBD-DAMSL tagset was insufficient. Below is

a list of the tags that were created specifically for the MRDA data. Each tag listed below

is followed by a brief description indicating why it entered the MRDA tagset.

 Interrupted <%->

Throughout the meetings, incomplete utterances arose on account of speakers

abandoning their utterances or being interrupted. To characterize why an incomplete

utterance arose, the interrupted tag was added (as the abandoned tag <%--> was

already present).

 Topic Change <tc>

Within the MRDA data, many instances arose in which speakers attempted to change

the topic. No other mechanism was present to mark such occurrences, so the <tc> tag

entered the MRDA tagset to mark changes in topic.

 Floor Holder <fh>

The SWBD-DAMSL tagset contained the tag <h> (hold), which was also incorporated

into the MRDA tagset. Utterances similar to those marked with <h> appeared mid-

speech within the MRDA data. The <fh> tag was implemented to distinguish between a

hold, which marks utterances in which a speaker "holds off" prior to answering a

question or prior to speaking when he is expected to speak, and these mid-speech

"holds.

 Floor Grabber <fg>

This tag entered the tagset as there were significant similarities among the means by

which speakers “gained” the floor and also due to the lack of a tack to mark such

instances. Speakers’ utterances often contained specific lexical items and higher

124

energy during these attempts to “gain” the floor. The <fg> tag entered the MRDA tagset

as a means to mark such utterances.

 Repeat <r>

This tag entered the MRDA tagset in order to mark possible subtle changes in the

manner in which a speaker repeats an utterance, whether for purposes of emphasis or

in response to a repetition request.

 Self-Correct Misspeaking <bsc>

This tag was added to differentiate cases in which the primary speaker alone corrected

his speech rather than being corrected by another speaker, which is indicated by the

<bc> tag.

 Understanding Check <bu>

This tag entered the MRDA tagset as there seemed to be a large number of distinct

cases in which a speaker wanted to check if his information was correct.

 Defending/Explanation <df>

This tag was added as speakers tended to defend their suggestions either immediately

prior to making a suggestion or immediately after. Its usage was later expanded to

include when speakers generally defended their points or offered explanations.

 “Follow Me” <f>

This tag was added as speakers tended to occasionally seek verification from their

listeners that their utterances were understood or agreed upon.

 Joke <j>

This tag was added to mark utterances of humorous content and jokes, as there was

previously no other means to mark such utterances.

125

 Rising Tone <rt>

Although this tag is not an actual dialog act, it was implemented to mark whether an

utterance ended with a rising tone for the purpose of providing information for automatic

speech recognition.

 Nonlabeled <z>

Certain utterances arose in the data that were intentionally not to be labeled. The <z>

tag entered the MRDA tagset specifically for this purpose.

126

APPENDIX 4: FINAL MRDA TAGSET REVISIONS

As work on dialog act labeling progressed, the original tagset used underwent many

changes and eventually evolved to the form that is presented within this guide. As most

changes to the tagset occurred early on, in its final stages, the tagset underwent a scant

number of changes prior to being finalized. During its final stages, a number of

meetings were labeled and consequently do not reflect a few of the minute changes

present within the current tagset. Those changes include the elimination of the <sj>,

<fx>, and <oo> tags. Instances in which the <sj> tag was used are preserved within the

data, however instances in which the <fx> and <oo> tags were used are not preserved

and the data has subsequently been updated to reflect the current tagset.

 Subjective Statement <sj>

Originally, a distinction existed where the statement tag <s> marked objective and

factual statements and the <sj> tag marked opinions and other subjective statements.

The <sj> tag eventually merged with the <s> tag, as there was a lack of agreement

among annotators regarding the use of the <sj> tag. The twenty-six meetings listed

below currently contain the <sj> tag:

Bed003 Bmr008 Bro004

Bed004 Bmr009 Bro005

Bed009 Bmr010 Bro007

Bed010 Bmr012 Bro008

Bed011 Bmr013 Bro012

Bmr001 Bmr014 Bro017

Bmr005 Bmr018 Bro018

Bmr006 Bmr024 Bro026

Bmr007 Bmr026

 Explicit Performative <fx>

This tag marked utterances in which a speaker made a declaration or performed some

sort of act, such as the act of "firing" in saying "you're fired" and the act of

"recommending" in saying "I recommend you try the other one." This tag was removed

from the tagset completely due to its lack of use.

127

Although no examples exist in the data of the welcome tag <fw>, the welcome tag is

complementary to the thanks tag <ft> and persists as a result of this relationship. The

explicit performative tag lacks a complementary relationship of this sort.

 Open Option <oo>

This tag marked utterances in which a speaker posed multiple options. It was removed

from the tagset completely due to its redundancy with suggestions <cs>.

128

BIBLIOGRAPHY

Jurafsky, Dan, Shriberg, Liz, and Biasca, Debra. 1997. “Switchboard SWBD-DAMSL

Shallow-Discourse-Function Annotation Coders Manual, Draft 13.” Technical Report

97-02, University of Colorado, Boulder, Institute of Cognitive Science.

Levinson, Stephen C. 1983. Pragmatics. Cambridge: Cambridge University Press.

129

INDEX OF TAGS

aa Accept, 57

aap Partial Accept, 59

am Maybe, 67

ar Reject, 61

arp Partial Reject, 62

b Backchannel, 49

ba Assessment/Appreciation, 52

bc Correct Misspeaking, 85

bd Downplayer, 92

bh Rhetorical Question

Backchannel, 55

bk Acknowledgement, 50

br Repetition Request, 77

bs Summary, 83

bsc Self-Correct Misspeaking, 85

bu Understanding Check, 78

by Sympathy, 94

cc Commitment, 74

co Command, 70

cs Suggestion, 73

d Declarative Question, 106

df Defending/Explanation, 87

e Elaboration, 88

f "Follow Me", 76

fa Apology, 94

fe Exclamation, 97

fg Floor Grabber, 43

fh Floor Holder, 45

ft Thanks, 96

fw Welcome, 96

g Tag Question, 108

h Hold, 46

j Joke, 102

m Mimic, 81

na Affirmative Answer, 60

nd Dispreferred Answer, 63

ng Negative Answer, 64

no No Knowledge, 68

qh Rhetorical Question, 42

qo Open-ended Question, 41

qr Or Question, 37

qrr Or Clause After Y/N Question, 40

qw Wh-Question, 35

qy Y/N Question, 33

r Repeat, 80

rt Rising Tone, 110

s Statement, 32

t About-Task, 98

tc Topic Change, 100

t1 Self Talk, 103

t3 Third Party Talk, 104

x Nonspeech, 113

z Nonlabeled, 113

2 Collaborative Completion, 90

% Indecipherable, 110

%- Interrupted, 111

%-- Abandoned, 111

Version4 Mrda Manual

Navigation menu

Versions of this User Manual:

Views

Navigation