MANUAL

User Manual:

Open the PDF directly: View PDF .
Page Count: 39

Introduction
Online Web Application and Web Service
Release
- Download
  - Language Models
- License
UDPipe Installation
UDPipe User's Manual
UDPipe API Reference
Contact
Acknowledgements

UDPipe

Version 1.2.0

Contents

1 Introduction 3

2 Online Web Application and Web Service 3

3 Release 3

3.1 Download ................................................. 3

3.1.1 Language Models ......................................... 3

3.2 License ................................................... 3

4 UDPipe Installation 4

4.1 Requirements ............................................... 4

4.2 Compilation ................................................ 4

4.2.1 Platforms ............................................. 4

4.2.2 Further Details .......................................... 4

4.3 Other language bindings ......................................... 5

4.3.1 C# ................................................ 5

4.3.2 Java ................................................ 5

4.3.3 Perl ................................................ 5

4.3.4 Python .............................................. 5

5 UDPipe User’s Manual 5

5.1 Running UDPipe ............................................. 5

5.1.1 Immediate Mode ......................................... 6

5.1.2 Loading Model On Demand ................................... 6

5.1.3 Tokenizer ............................................. 6

5.1.4 Input Formats .......................................... 7

5.1.5 Tagger ............................................... 8

5.1.6 Dependency Parsing ....................................... 8

5.1.7 Output Formats ......................................... 8

5.2 Running the UDPipe REST Server ................................... 9

5.3 Training UDPipe Models ........................................ 9

5.3.1 Reusing Components from Existing Models .......................... 9

5.3.2 Random Hyperparameter Search ................................ 9

5.3.3 Tokenizer ............................................. 10

5.3.4 Tagger ............................................... 10

5.3.5 Parser ............................................... 11

5.3.6 Measuring Model Accuracy ................................... 13

5.4 Universal Dependencies 2.0 Models ................................... 13

5.4.1 Download ............................................. 13

5.4.2 Acknowledgements ........................................ 13

5.4.3 Model Description ........................................ 14

5.4.4 Model Performance ........................................ 14

5.5 CoNLL17 Shared Task Baseline UD 2.0 Models ............................ 16

5.5.1 Download ............................................. 16

5.5.2 Acknowledgements ........................................ 16

5.6 Universal Dependencies 1.2 Models ................................... 16

5.6.1 Download ............................................. 16

5.6.2 Acknowledgements ........................................ 16

5.6.3 Model Description ........................................ 17

5.6.4 Model Performance ........................................ 17

6 UDPipe API Reference 18

6.1 UDPipe Versioning ............................................ 18

6.2 Struct string piece ............................................ 18

6.3 Class token ................................................ 18

6.3.1 token::get space after() ..................................... 19

6.3.2 token::set space after() ...................................... 19

6.3.3 token::get spaces before() .................................... 19

6.3.4 token::set spaces before() .................................... 19

6.3.5 token::get spaces after() ..................................... 19

6.3.6 token::set spaces after() ..................................... 20

6.3.7 token::get spaces in token() ................................... 20

6.3.8 token::set spaces in token() ................................... 20

6.3.9 token::get token range() ..................................... 20

6.3.10 token::set token range() ..................................... 20

6.4 Class word ................................................ 20

6.5 Class multiword token .......................................... 21

6.6 Class empty node ............................................. 21

6.7 Class sentence ............................................... 21

6.7.1 sentence::empty() ......................................... 22

6.7.2 sentence::clear() ......................................... 22

6.7.3 sentence::add word() ....................................... 22

6.7.4 sentence:set head() ........................................ 22

6.7.5 sentence::unlink all words() ................................... 22

6.7.6 sentence::get new doc() ..................................... 22

6.7.7 sentence::set new doc() ..................................... 22

6.7.8 sentence::get new par() ..................................... 23

6.7.9 sentence::set new par() ...................................... 23

6.7.10 sentence::get sent id() ...................................... 23

6.7.11 sentence::set sent id() ...................................... 23

6.7.12 sentence::get text() ........................................ 23

6.7.13 sentence::set text() ........................................ 23

6.8 Class input format ............................................ 23

6.8.1 input format::read block() .................................... 24

6.8.2 input format::reset document() ................................. 24

6.8.3 input format::set text() ..................................... 24

6.8.4 input format::next sentence() .................................. 24

6.8.5 input format::new input format() ................................ 24

6.8.6 input format::new conllu input format() ............................ 25

6.8.7 input format::new generic tokenizer input format() ...................... 25

6.8.8 input format::new horizontal input format() .......................... 25

6.8.9 input format::new vertical input format() ........................... 25

6.8.10 input format::new presegmented tokenizer() .......................... 25

6.9 Class output format ........................................... 26

6.9.1 output format::write sentence() ................................. 26

6.9.2 output format::ﬁnish document() ................................ 26

6.9.3 output format::new output format() .............................. 26

6.9.4 output format::new conllu output format() .......................... 27

6.9.5 output format::new epe output format() ............................ 27

6.9.6 output format::new matxin output format() .......................... 27

6.9.7 output format::new plaintext output format() ......................... 27

6.9.8 output format::new horizontal output format() ........................ 27

6.9.9 output format::new vertical output format() ......................... 28

6.10 Class model ................................................ 28

6.10.1 model::load(const char*) ..................................... 28

6.10.2 model::load(istream&) ...................................... 28

6.10.3 model::new tokenizer() ...................................... 28

6.10.4 model::tag() ............................................ 29

6.10.5 model::parse() .......................................... 29

6.11 Class pipeline ............................................... 29

6.11.1 pipeline::set model() ....................................... 29

6.11.2 pipeline::set input() ....................................... 29

6.11.3 pipeline::set tagger() ....................................... 29

6.11.4 pipeline::set parser() ....................................... 30

6.11.5 pipeline::set output() ....................................... 30

6.11.6 pipeline::set immediate() .................................... 30

6.11.7 pipeline::set document id() ................................... 30

6.11.8 pipeline::process() ........................................ 30

6.12 Class trainer ............................................... 30

6.12.1 trainer::train() .......................................... 30

6.13 Class evaluator .............................................. 31

6.13.1 evaluator::set model() ...................................... 31

6.13.2 evaluator::set tokenizer() .................................... 31

6.13.3 evaluator::set tagger() ...................................... 31

6.13.4 evaluator::set parser() ...................................... 31

6.13.5 evaluator::evaluate() ....................................... 31

6.14 Class version ............................................... 31

6.14.1 version::current .......................................... 32

6.15 C++ Bindings API ............................................ 32

6.15.1 Helper Structures ......................................... 32

6.15.2 Main Classes ........................................... 34

6.16 C# Bindings ............................................... 36

6.17 Java Bindings ............................................... 36

6.18 Perl Bindings ............................................... 36

6.19 Python Bindings ............................................. 36

7 Contact 36

8 Acknowledgements 37

8.1 Publications ................................................ 37

8.2 Bibtex for Referencing .......................................... 37

8.3 Persistent Identiﬁer ............................................ 37

1 Introduction

UDPipe is a trainable pipeline for tokenization, tagging, lemmatization and dependency parsing of CoNLL-U

ﬁles. UDPipe is language-agnostic and can be trained given annotated data in CoNLL-U format. Trained

models are provided for nearly all UD treebanks. UDPipe is available as a binary for Linux/Windows/OS X,

as a library for C++, Python, Perl, Java, C#, and as a web service.

UDPipe is a free software distributed under the Mozilla Public License 2.0 and the linguistic models are free for

non-commercial use and distributed under the CC BY-NC-SA license, although for some models the original

data used to create the model may impose additional licensing conditions. UDPipe is versioned using Semantic

Versioning.

University, Czech Republic.

2 Online Web Application and Web Service

UDPipe Web Application is available at http://lindat.mﬀ.cuni.cz/services/udpipe/ using LINDAT/CLARIN

infrastructure.

UDPipe REST Web Service is also available, with the API documentation available at

http://lindat.mﬀ.cuni.cz/services/udpipe/api-reference.php.

3 Release

3.1 Download

UDPipe releases are available on GitHub, both as source code and as a pre-compiled binary package. The

binary package contains Linux, Windows and OS X binaries, Java bindings binary, C# bindings binary, and

source code of UDPipe and all language bindings). While the binary packages do not contain compiled Python

or Perl bindings, packages for those languages are available in standard package repositories, i.e. on PyPI and

CPAN.

•Latest release

•All releases,Changelog

3.1.1 Language Models

To use UDPipe, a language model is needed. The language models are available from LINDAT/CLARIN

infrastructure and described further in the UDPipe User’s Manual. Currently, the following language models

are available:

•Universal Dependencies 2.0 Models: udpipe-ud2.0-170801 (documentation)

•CoNLL17 Shared Task Baseline UD 2.0 Models: udpipe-ud2.0-conll17-170315 (documentation)

•Universal Dependencies 1.2 Models: udpipe-ud1.2-160523 (documentation)

3.2 License

UDPipe is an open-source project and is freely available for non-commercial purposes. The library is distributed

under Mozilla Public License 2.0 and the associated models and data under CC BY-NC-SA, although for some

models the original data used to create the model may impose additional licensing conditions.

If you use this tool for scientiﬁc work, please give credit to us by referencing Straka et al. 2016 and the UDPipe

website.

4 UDPipe Installation

UDPipe releases are available on GitHub, either as a pre-compiled binary package, or source code only. The

binary package contains Linux, Windows and OS X binaries, Java bindings binary, C# bindings binary, and

source code of UDPipe and all language bindings. While the binary packages do not contain compiled Python

or Perl bindings, packages for those languages are available in standard package repositories, i.e. on PyPI and

CPAN.

To use UDPipe, a language model is needed. Here is a list of available language models.

If you want to compile UDPipe manually, sources are available on on GitHub, both in the pre-compiled binary

package releases and in the repository itself.

4.1 Requirements

•g++ 4.7 or newer, clang 3.2 or newer, Visual C++ 2015 or newer

•make

•SWIG 3.0.8 or newer for language bindings other than C++

4.2 Compilation

To compile UDPipe, run make in the src directory.

Make targets and options:

•exe: compile the binaries (default)

•server: compile the REST server

•lib: compile the static library

•BITS=32 or BITS=64: compile for speciﬁed 32-bit or 64-bit architecture instead of the default one

•MODE=release: create release build which statically links the C++ runtime and uses LTO

•MODE=debug: create debug build

•MODE=profile: create proﬁle build

4.2.1 Platforms

Platform can be selected using one of the following options:

•PLATFORM=linux,PLATFORM=linux-gcc: gcc compiler on Linux operating system, default on Linux

•PLATFORM=linux-clang: clang compiler on Linux, must be selected manually

•PLATFORM=osx,PLATFORM=osx-clang: clang compiler on OS X, default on OS X; BITS=32+64 enables

multiarch build

•PLATFORM=win,PLATFORM=win-gcc: gcc compiler on Windows (TDM-GCC is well tested), default on

Windows

•PLATFORM=win-vs: Visual C++ 2015 compiler on Windows, must be selected manually; note that the

cl.exe compiler must be already present in PATH and corresponding BITS=32 or BITS=64 must be speciﬁed

Either POSIX shell or Windows CMD can be used as shell, it is detected automatically.

4.2.2 Further Details

UDPipe uses C++ BuilTem system, please refer to its manual if interested in all supported options.

4.3 Other language bindings

4.3.1 C#

Binary C# bindings are available in UDPipe binary packages.

To compile C# bindings manually, run make in the bindings/csharp directory, optionally with the options

described in UDPipe Installation.

4.3.2 Java

Binary Java bindings are available in UDPipe binary packages.

To compile Java bindings manually, run make in the bindings/java directory, optionally with the options

described in UDPipe Installation. Java 6 and newer is supported.

The Java installation speciﬁed in the environment variable JAVA HOME is used. If the environment variable does

not exist, the JAVA HOME can be speciﬁed using

make JAVA_HOME=path_to_Java_installation

4.3.3 Perl

The Perl bindings are available as Ufal-UDPipe package on CPAN.

To compile Perl bindings manually, run make in the bindings/perl directory, optionally with the options

described in UDPipe Installation. Perl 5.10 and later is supported.

Path to the include headers of the required Perl version must be speciﬁed in the PERL INCLUDE variable using

make PERL_INCLUDE=path_to_Perl_includes

4.3.4 Python

The Python bindings are available as ufal.udpipe package on PyPI.

To compile Python bindings manually, run make in the bindings/python directory, optionally with options

described in UDPipe Installation. Both Python 2.6+ and Python 3+ are supported.

Path to the include headers of the required Python version must be speciﬁed in the PYTHON INCLUDE variable

using

make PYTHON_INCLUDE=path_to_Python_includes

5 UDPipe User’s Manual

Like any supervised machine-learning tool, UDPipe needs a trained linguistic model. This section describes the

available language models and also the command line tools and interfaces.

5.1 Running UDPipe

Probably the most common usage of UDPipe is to tokenize, tag and parse the input using

udpipe --tokenize --tag --parse udpipe_model

The input is assumed to be in UTF-8 encoding and can be either already tokenized and segmented, or it can

be a plain text which will be tokenized and segmented automatically.

Any number of input ﬁles can be speciﬁed after the udpipe model and if no ﬁle is given, the standard input is

used. The output is by default saved to the standard output, but if --outfile=name is used, it is saved to the

given ﬁle name. The output ﬁle name can contain a {}, which is replaced by a base name of the processed ﬁle

(i.e., without directories and an extension).

The full command syntax of running UDPipe is

Usage: udpipe [running_opts] udpipe_model [input_files]

udpipe --train [training_opts] udpipe_model [input_files]

udpipe --detokenize [detokenize_opts] raw_text_file [input_files]

Running opts: --accuracy (measure accuracy only)

--input=[conllu|generic_tokenizer|horizontal|vertical]

--immediate (process sentences immediately during loading)

--outfile=output file template

--output=[conllu|matxin|horizontal|plaintext|vertical]

--tokenize (perform tokenization)

--tokenizer=tokenizer options, implies --tokenize

--tag (perform tagging)

--tagger=tagger options, implies --tag

--parse (perform parsing)

--parser=parser options, implies --parse

Training opts: --method=[morphodita_parsito] which method to use

--heldout=heldout data file name

--tokenizer=tokenizer options

--tagger=tagger options

--parser=parser options

Detokenize opts: --outfile=output file template

Generic opts: --version

--help

5.1.1 Immediate Mode

By default UDPipe loads the whole input ﬁle into memory before starting to process it. That allows to store

the space markup (see the following Tokenizer section) in most consistent way, i.e., store all spaces following a

sentence in the last token of that sentence.

However, sometimes it is desirable to process the input as soon as possible, which can be achieved by specifying

the --immediate option. In immediate mode, the input is processed and printed as soon as a block of input

guaranteed to contain whole sentences is loaded. Speciﬁcally, for most input formats the input is processed after

loading an empty line (with the exception of horizontal input format and presegmented tokenizer, where the

input is processed after each line).

5.1.2 Loading Model On Demand

Although a model for UDPipe always has to be speciﬁed, the model is loaded only if really needed. It is therefore

possible to use for example none as the model in case it is not required for performing the requested operation

(e.g., converting between formats or using a generic tokenizer).

5.1.3 Tokenizer

If the --tokenize option is supplied, the input is assumed to be plain text and is tokenized using model

tokenizer. Additional arguments to the tokenizer might be speciﬁed using --tokenizer=data option (which

implies --tokenize), where data is a semicolon-separated list of the following options:

•normalized spaces: by default, UDPipe uses custom MISC ﬁelds to exactly encode spaces in the original

document (as described below). If the normalized spaces option is given, only the standard CoNLL-U

v2 markup (SpaceAfter=No and # newpar) is used.

•presegmented: the input ﬁle is assumed to be already segmented, with each sentence on a separate line,

and is only tokenized (respecting sentence breaks)

•ranges: for each token, a range in the original document is stored in the format described below.

•joint with parsing: an experimental mode performing sentence segmentation jointly using the tokenizer

and the parser (see Milan Straka and Jana Strakov´a: Tokenizing, POS Tagging, Lemmatizing and Parsing

UD 2.0 with UDPipe paper for details). The following options are utilized:

–joint max sentence len (default 20): maximum sentence length

–joint change boundary logprob (default -0.5): logprob of using sentence boundary not generated

by the tokenizer

–joint sentence logprob (default -0.5): additional logprob of every sentence

The logprob of a sentence is computed using logprob of its best dependency parsing tree, together with

joint sentence logprob and also joint change boundary logprob for every sentence boundary not

returned by the tokenizer (i.e., either 0, 1 or 2 times). The joint sentence segmentation chooses such a

segmentation, where every sentence has length at most joint max sentence len and the sum of logprobs

of all sentences is as large as possible.

Preserving Original Spaces

By default, UDPipe uses custom MISC ﬁelds to store all spaces in the original document. This markup is

backward compatible with CoNLL-U v2 SpaceAfter=No feature. This markup can be utilized by the plaintext

output format, which allows reconstructing the original document.

Note that in theory not only spaces, but also other original content can be saved in this way (for example XML

tags if the input was encoded in a XML ﬁle).

The markup uses the following MISC ﬁelds on tokens (not words in multi-word tokens):

•SpacesBefore=content (by default empty): spaces/other content preceding the token

•SpacesAfter=content (by default a space if SpaceAfter=No feature is not present, empty otherwise):

spaces/other content following the token

•SpacesInToken=content (by default equal to the FORM of the token): FORM of the token including

original spaces (this is needed only if tokens are allowed to contain spaces and a token contains a tab or

newline characters)

The content of all the three ﬁelds must be escaped to allow storing tabs and newlines. The following C-like

schema is used:

• \s: space

• \t: tab

• \r: CR character

• \n: LF character

• \p:|(pipe character)

• \\:\(backslash character)

Preserving Token Ranges

When the ranges tokenizer option is used, the range of each token in the original document is stored in the

TokenRange MISC ﬁeld.

The format of the TokenRange ﬁeld (inspired by Python) is TokenRange=start:end, where start is a zero-

based document-level index of the start of the token (counted in Unicode characters) and end is a zero-based

document-level index of the ﬁrst character following the token (i.e., the length of the token is end-start).

5.1.4 Input Formats

If the tokenizer is not used, the input format can be speciﬁed using the --input option. The individual input

formats can be parametrized in the same way a tokenizer is, by using format=data syntax. Currently supported

input formats are:

•conllu (default): the CoNLL-U format. Supported options:

–v2 (default): use CoNLL-U v2

–v1: allow loading only CoNLL-U v1 (i.e., no empty nodes and no spaces in forms and lemmas)

•generic tokenizer: generic tokenizer for English-like languages (with spaces separating tokens and

English-like punctuation). The tokenizer is rule-based and needs no trained model. It supports the

same options as a model tokenizer, i.e., normalized spaces,presegmented and ranges.

•horizontal: each sentence on a separate line, with tokens separated by spaces. In order to allow spaces

in tokens, Unicode character ’NO-BREAK SPACE’ (U+00A0) is considered part of token and converted

to a space during loading.

•vertical: each token on a separate line, with an empty line denoting end of sentence; only the ﬁrst

tab-separated word is used as a token, the rest of the line is ignored.

Note that a model tokenizer can be speciﬁed using the --input option too, by using the tokenizer input

format, for example using --input tokenizer=ranges.

5.1.5 Tagger

If the --tag option is supplied, the input is POS tagged and lemmatized using the model tagger. Additional

arguments to the tagger might be speciﬁed using the --tagger=data option (which implies --tag).

5.1.6 Dependency Parsing

If the --parse option is supplied, the input is parsed using the model dependency parser. Additional arguments

to the parser might be speciﬁed using the --parser=data option (which implies --parse).

5.1.7 Output Formats

The output format is speciﬁed using the --output option. The individual output formats can be parametrized

in the same way as input formats, by using the format=data syntax. Currently supported output formats are:

•conllu (default): the CoNLL-U format Supported options:

–v2 (default): use CoNLL-U v2

–v1: produce output in CoNLL-U v1 format. Note that this is a lossy process, as empty nodes are

ignored and spaces in forms and lemmas are converted to underscores.

•matxin: the Matxin format

•horizontal: writes the words (in the UD sense) in horizontal format, that is, each sentence is on a

separate line, with words separated by a single space. Because words can contain spaces in CoNLL-U

v2, the spaces in words are converted to Unicode character ’NO-BREAK SPACE’ (U+00A0). Supported

options:

–paragraphs: an empty line is printed after the end of a paragraph or a document (recognized by #

newpar or # newdoc comments)

•plaintext: writes the tokens (in the UD sense) using original spacing. By default, UDPipe’s custom

MISC features (SpacesBefore,SpacesAfter and SpacesInToken, see the description in the Tokenizer

section) are used to reconstruct the exact original spaces. However, if the document does not contain

these features or if you want only normalized spacing, you can use the following option:

–normalized spaces: write one sentence on a line, and either one or no space between tokens accord-

ing to the SpaceAfter=No feature

•vertical: each word on a separate line, with an empty line denoting the end of sentence. Supported

options:

–paragraphs: an empty line is printed after the end of a paragraph or a document (recognized by #

newpar or # newdoc comments)

5.2 Running the UDPipe REST Server

UDPipe also provides a REST server binary called udpipe server. The binary uses MicroRestD as a REST

server implementation and provides UDPipe REST API.

The full command syntax of udpipe server is

udpipe_server [options] port default_model (rest_id model_file acknowledgements)*

Options: --concurrent_models=maximum concurrently loaded models (default 10)

--daemon (daemonize after start)

--no_check_models_loadable (do not check models are loadable)

--no_preload_default (do not preload default model)

The udpipe server can run either in foreground or in background (when --daemon is used).

Since UDPipe 1.1.1, the models are loaded on demand, so that at most concurrent models (default 10) are

kept in memory at the same time. The model ﬁles are opened during start and never closed until the server

stops. Unless no check models loadable is speciﬁed, the model ﬁles are also checked to be loadable during

start. Note that the default model is preloaded and never released, unless no preload default is given. (Before

UDPipe 1.1.1, speciﬁed model ﬁles were loaded during start and kept in memory all the time.)

5.3 Training UDPipe Models

Custom UDPipe models can be trained using the following syntax:

udpipe --train model.output [--heldout=heldout_data] training_file ...

The training data should be in the CoNLL-U format.

By default, three model components are trained – tokenizer, tagger and parser. Any subset of the model

components can be trained and a model component may be copied from an existing model.

The training options are speciﬁed for each model component separately using the --tokenizer,--tagger

and --parser options. If a model component should not be trained, value none should be used (e.g.,

--tagger=none).

The options are name=value pairs separated by a semicolon. The value can be either a simple string value

(ending by a semicolon), ﬁle content speciﬁed as name=file:filename, or an arbitrary string value speciﬁed as

name=data:length:value, where the value is exactly length bytes long.

5.3.1 Reusing Components from Existing Models

The model components (tagger, parser or tagger) can be reused from existing models, by specifying the

from model=file:filename option.

5.3.2 Random Hyperparameter Search

The default values of hyperparameters are set to the values which were used the most during UD 1.2 models

training, but if you want to reach best performance, the hyperparameters must be tuned.

Apart from manual grid search, UDPipe can perform a simple random search. You can perform the random

search by repeatedly training UDPipe (preferably in parallel, most likely on diﬀerent computers) while specifying

diﬀerent training run number – some of the hyperparameters (chosen by us; you can of course override their

value by specifying it on the command line) change their values in diﬀerent training runs. The pseudorandom

sequences of hyperparameters are of course deterministic.

The training run can be speciﬁed by providing the run=number option to a model component. The run number

1 is the default one (with the best hyperparameters for the UD 1.2 models), run numbers 2 and more randomize

the hyperparameters.

5.3.3 Tokenizer

The tokenizer is trained using the SpaceAfter=No features in the CoNLL-U ﬁles. If the feature is not present, a

detokenizer can be used to guess the SpaceAfter=No features according to a supplied plain text (which typically

does not overlap with the texts in the CoNLL-U ﬁles).

In order to use the detokenizer, use the detokenizer=file:filename with plaintext option. In UD 1.2

models, the optimal performance is achieved with very small plain texts – only 500kB.

The tokenizer recognizes the following options:

•tokenize url (default 1): tokenize URLs and emails using a manually implemented recognizer

•allow spaces (default 1 if any token contains a space, 0 otherwise): allow tokens to contain spaces

•dimension (default 24): dimension of character embeddings and of the per-character bidirectional GRU.

Note that inference time is quadratic in this parameter. Supported values are only 16, 24 and 64, with 64

needed only for languages with complicated tokenization like Japanese, Chinese or Vietnamese.

•epochs (default 100): the number of epochs to train the tokenizer for

•batch size (default 50): batch size used during tokenizer training

•learning rate (default 0.005): the learning rate used during tokenizer training

•dropout (default 0.1): dropout used during tokenizer training

•early stopping (default 1 if heldout is given, 0 otherwise): perform early stopping, choosing training

iteration maximizing sentences F1 score plus tokens F1 score on heldout data

During random hyperparameter search, batch size is chosen uniformly from {50,100}and learning rate

logarithmically from <0.0005, 0.01).

Detokenizing CoNLL-U Files

The --detokenizer option allows generating the SpaceAfter=No features automatically from a given plain

text. Even if the current algorithm is very simple and makes quite a lot of mistakes, the tokenizer trained on

generated features is very close to a tokenizer trained on gold SpaceAfter=No features (the diﬀerence in token

F1 score is usually one or two tenths of percent).

The generated SpaceAfter=No features are only used during tokenizer training, not printed. However, if you

would like to obtain the CoNLL-U ﬁles with automatic detokenization (generated SpaceAfter=No features),

you can run UDPipe with the --detokenize option. In this case, you have to supply plain text in the given

language (usually the best results are achieved with just 500kB or 1MB of text) and UDPipe then detokenizes

all the given CoNLL-U ﬁles.

The complete usage of the --detokenize option is:

udpipe --detokenize [detokenize_opts] raw_text_file [input_files]

Detokenize opts: --outfile=output file template

5.3.4 Tagger

The tagging is currently performed using MorphoDiTa. The UDPipe tagger consists of possibly several Mor-

phoDiTa models, each tagging some of the POS tags and/or lemmas.

By default, only one model is constructed, which generates all available tags (UPOS, XPOS, Feats and Lemma).

However, we found out during the UD 1.2 models training that performance improves if one model tags the

UPOS, XPOS and Feats tags, while the other is performing lemmatization. Therefore, if you utilize two

MorphoDiTa models, by default the ﬁrst one generates all tags (except lemmas) and the second one performs

lemmatization.

The number of MorphoDiTa models can be speciﬁed using the models=number parameter. All other pa-

rameters may be either generic for all models (guesser suffix rules=5), or speciﬁc for a given model

(guesser suffix rules 2=6), including the from model option (therefore, MorphoDiTa models can be trained

separately and then combined together into one UDPipe model).

Every model utilizes UPOS for disambiguation and the ﬁrst model is the one producing the UPOS tags on

output.

The tagger recognizes the following options:

•use lemma (default for the second model and also if there is only one model): use the lemma ﬁeld internally

to perform disambiguation; the lemma may be not outputted

•provide lemma (default for the second model and also if there is only one model): produce the disam-

biguated lemma on output

•use xpostag (default for the ﬁrst model): use the XPOS tags internally to perform disambiguation; it

may not be outputted

•provide xpostag (default for the ﬁrst model): produce the disambiguated XPOS tag on output

•use feats (default for the ﬁrst model): use the Feats internally to perform disambiguation; it may not

be outputted

•provide feats (default for the ﬁrst model): produce the disambiguated Feats ﬁeld on output

•dictionary max form analyses (default 0 - unlimited): the maximum number of (most frequent) form

analyses from UD training data that are to be kept in the morphological dictionary

•dictionary file (default empty): use a given custom morphological dictionary, where each line contains

5 tab-separated ﬁelds FORM, LEMMA, UPOSTAG, XPOSTAG and FEATS. Note that this dictionary

data is appended to the dictionary created from the UD training data, not replacing it.

•guesser suffix rules (default 8): number of rules generated for every suﬃx

•guesser prefixes max (default 4 if “provide lemma‘, 0 otherwise): maximum number of form-generating

preﬁxes to use in the guesser

•guesser prefix min count (default 10): minimum number of occurrences of form-generating preﬁx to

consider using it in the guesser

•guesser enrich dictionary (default 6 if no dictionary file is passed, 0 otherwise): number of rules

generated for forms present in training data (assuming that the analyses from the training data may not

be all)

•iterations (default 20): number of training iterations to perform

•early stopping (default 1 if heldout is given, 0 otherwise): perform early stopping, choosing training

iteration maximizing tagging accuracy on the heldout data

•templates (default lemmatizer for second model, tagger otherwise): MorphoDiTa feature templates

to use, either lemmatizer which focuses more on lemmas, or tagger which focuses more on UPOS/X-

POS/FEATS

During random hyperparameter search, guesser suffix rules is chosen uniformly from {5,6,7,8,9,10,11,12}

and guesser enrich dictionary is chosen uniformly from {3,4,5,6,7,8,9,10}.

5.3.5 Parser

The parsing is performed using Parsito, which is a transition-based parser using a neural-network classiﬁer.

The transition-based systems can be conﬁgured by the following options:

•transition system (default projective): which transition system to use for parsing (language dependent,

you can choose according to language properties or try all and choose the best one)

–projective: projective stack-based arc standard system with shift,left arc and right arc tran-

sitions

–swap: fully non-projective system which extends projective system by adding the swap transition

–link2: partially non-projective system which extends projective system by adding left arc2 and

right arc2 transitions

•transition oracle (default dynamic/static lazy static whichever ﬁrst is applicable): which transition

oracle to use for the chosen transition system:

–transition system=projective: available oracles are static and dynamic (dynamic usually gives

better results, but training time is slower)

–transition system=swap: available oracles are static eager and static lazy (static lazy al-

most always gives better results)

–transition system=link2: only available oracle is static

•structured interval (default 8): use search-based oracle in addition to the translation oracle speci-

ﬁed. This almost always gives better results, but makes training 2-3 times slower. For details, see the paper

Straka et al. 2015: Parsing Universal Dependency Treebanks using Neural Networks and Search-Based

Oracle

•single root (default 1): allow only single root when parsing, and make sure only the root node has the

root deprel (note that training data are checked to be in this format)

The Lemmas/UPOS/XPOS/FEATS used by the parser are conﬁgured by:

•use gold tags (default 0): if false and a tagger exists, the Lemmas/UPOS/XPOS/FEATS for both the

training and heldout data are generated by the tagger, otherwise they are taken from the gold data

The embeddings used by the parser can be speciﬁed as follows:

•embedding upostag (default 20): the dimension of the UPos embedding used in the parser

•embedding feats (default 20): the dimension of the Feats embedding used in the parser

•embedding xpostag (default 0): the dimension of the XPos embedding used in the parser

•embedding form (default 50): the dimension of the Form embedding used in the parser

•embedding lemma (default 0): the dimension of the Lemma embedding used in the parser

•embedding deprel (default 20): the dimension of the Deprel embedding used in the parser

•embedding form file: pre-trained word embeddings in word2vec textual format

•embedding lemma file: pre-trained lemma embeddings in word2vec textual format

•embedding form mincount (default 2): for forms not present in the pre-trained embeddings, generate

random embeddings if the form appears at least this number of times in the trainig data (forms not

present in the pre-trained embeddings and appearing less number of times are considered OOV)

•embedding lemma mincount (default 2): for lemmas not present in the pre-trained embeddings, generate

random embeddings if the lemma appears at least this number of times in the trainig data (lemmas not

present in the pre-trained embeddings and appearing less number of times are considered OOV)

The neural-network training options:

•iterations (default 10): number of training iterations to use

•hidden layer (default 200): the size of the hidden layer

•batch size (default 10): batch size used during neural-network training

•learning rate (default 0.02): the learning rate used during neural-network training

•learning rate final (0.001): the ﬁnal learning rate used during neural-network training

•l2 (0.5): the L2 regularization used during neural-network training

•early stopping (default 1 if heldout is given, 0 otherwise): perform early stopping, choosing training

iteration maximizing LAS on heldout data

During random hyperparameter search, structured interval is chosen uniformly from {0,8,10},

learning rate is chosen logarithmically from <0.005,0.04) and l2 is chosen uniformly from <0.2,0.6).

Pre-trained Word Embeddings

The pre-trained word embeddings for forms and lemmas can be speciﬁed in the word2vec textual format using

the embedding form file and embedding lemma file options.

Note that pre-training word embeddings even on the UD data itself improves the accuracy (we use

word2vec with -cbow 0 -size 50 -window 10 -negative 5 -hs 0 -sample 1e-1 -threads 12 -binary

0 -iter 15 -min-count 2 options to pre-train on the UD data after converting it to the horizontal format

using udpipe --output=horizontal).

Forms and lemmas can contain spaces in CoNLL-U v2, so these spaces are converted to a Unicode character

’NO-BREAK SPACE’ (U+00A0) before performing the embedding lookup, because spaces are usually used

to delimit tokens in word embedding generating software (both word2vec and glove use spaces to separate

words on input and on output). When using UDPipe to generate plain texts from CoNLL-U format using

--output=horizontal, this space replacing happens automatically.

When looking up an embedding for a given word, the following possibilities are tried in the following order until

a match is found (or an embedding for unknown word is returned):

•original word

•all but the ﬁrst character lowercased

•all characters lowercased

•if the word contains only digits, just the ﬁrst digit is tried

5.3.6 Measuring Model Accuracy

Measuring custom model accuracy can be performed by running:

udpipe --accuracy [udpipe_options] udpipe_model file ...

The command syntax is similar to the regular UDPipe operation, only the input must be always in the CoNLL-U

format and the --input and --output options are ignored.

Three diﬀerent settings (depending on --tokenize(r),--tag(ger) and --parse(r)) can be evaluated:

•--tokenize(r) [--tag(ger) [--parse(r)]]: Tokenizer is used to segment and tokenize plain text (ob-

tained by SpaceAfter=No features and # newdoc and # newpar comments in the input ﬁle). Optionally,

a tagger is used on the resulting data to obtain Lemma/UPOS/XPOS/Feats columns and eventually a

parser can be used to parse the results.

The tokenizer is evaluated using F1-score on tokens, multi-word tokens, sentences and words. The words

are aligned using a word-alignment algorithm described in the CoNLL 2017 Shared Task in UD Pars-

ing. The tagger and parser are evaluated on aligned words, resulting in F1 scores of Lemmas/UPOS/X-

POS/Feats/UAS/LAS.

•--tag(ger) [--parse(r)]: The gold segmented and tokenized input is tagged (and then optionally

parsed using the tagger outputs) and then evaluated.

•--parse(r): The gold segmented and tokenized input is parsed using gold morphology (Lemmas/U-

POS/XPOS/Feats) and evaluated.

5.4 Universal Dependencies 2.0 Models

Universal Dependencies 2.0 Models are distributed under the CC BY-NC-SA licence. The models are based

solely on Universal Dependencies 2.0 treebanks. The models work in UDPipe version 1.2 and later.

Universal Dependencies 2.0 Models are versioned according to the date released in the format YYMMDD, where

YY,MM and DD are two-digit representation of year, month and day, respectively. The latest version is 170801.

5.4.1 Download

The latest version 170801 of the Universal Dependencies 2.0 models can be downloaded from LINDAT/CLARIN

repository.

5.4.2 Acknowledgements

This work has been partially supported and has been using language resources and tools developed, stored

and distributed by the LINDAT/CLARIN project of the Ministry of Education, Youth and Sports of

the Czech Republic (project LM2015071 ). The wark was also partially supported by OP VVV projects

CZ.02.1.01/0.0/0.0/16\013/0001781 and CZ.02.2.69/0.0/0.0/16\018/0002373, and by SVV project number

260 453.

The models were trained on Universal Dependencies 2.0 treebanks.

For the UD treebanks which do not contain original plain text version, raw text is used to train the tokenizer

instead. The plain texts were taken from the W2C – Web to Corpus.

Publications

•(Straka et al. 2017) Milan Straka and Jana Strakov´a. Tokenizing, POS Tagging, Lemmatizing and Parsing

UD 2.0 with UDPipe. In Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw

Text to Universal Dependencies, Vancouver, Canada, August 2017.

•(Straka et al. 2016) Straka Milan, Hajiˇc Jan, Strakov´a Jana. UDPipe: Trainable Pipeline for Processing

CoNLL-U Files Performing Tokenization, Morphological Analysis, POS Tagging and Parsing. In Pro-

ceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016),

Portoroˇz, Slovenia, May 2016.

5.4.3 Model Description

The Universal Dependencies 2.0 models contain 68 models of 50 languages, each consisting of a tokenizer, tagger,

lemmatizer and dependency parser, all trained using the UD data. Note that we use custom train-dev split, by

moving sentences from the beginning of dev data to the end of train data, until the training data is at least 9

times the dev data.

The tokenizer is trained using the SpaceAfter=No features. If the features are not present in the data, they can

be ﬁlled in using raw text in the language in question.

The tagger, lemmatizer and parser are trained using gold UD data.

Details about model architecture and training process can be found in the (Straka et al. 2017) paper.

Reproducible Training

In case you want to train the same models, scripts for downloading and resplitting UD 2.0 data, precomputed

word embedding, raw texts for tokenizers, all hyperparameter values and training scripts are available in the

second archive on the model download page.

5.4.4 Model Performance

We present the tagger, lemmatizer and parser performance, measured on the testing portion of the data, eval-

uated in three diﬀerent settings: using raw text only, using gold tokenization only, and using gold tokenization

plus gold morphology (UPOS, XPOS, FEATS and Lemma).

Treebank Mode Words Sents UPOS XPOS Feats AllTags Lemma UAS LAS

Ancient Greek Raw text 100.0% 98.7% 82.4% 72.3% 85.8% 72.3% 82.6% 64.4% 57.8%

Ancient Greek Gold tok - - 82.4% 72.4% 85.8% 72.3% 82.7% 64.6% 57.9%

Ancient Greek Gold tok+morph - - - - - - - 69.2% 64.4%

Ancient Greek-PROIEL Raw text 100.0% 47.2% 95.8% 96.0% 88.6% 87.2% 92.6% 71.8% 67.1%

Ancient Greek-PROIEL Gold tok - - 95.8% 96.1% 88.7% 87.2% 92.8% 77.2% 72.3%

Ancient Greek-PROIEL Gold tok+morph - - - - - - - 79.7% 76.1%

Arabic Raw text 93.8% 83.1% 88.4% 83.4% 83.5% 82.3% 87.5% 71.7% 65.8%

Arabic Gold tok - - 94.4% 89.5% 89.6% 88.3% 92.6% 81.3% 74.3%

Arabic Gold tok+morph - - - - - - - 82.9% 77.9%

Basque Raw text 100.0% 99.5% 93.2% - 87.6% - 93.8% 75.8% 70.7%

Basque Gold tok - - 93.3% - 87.7% - 93.9% 75.9% 70.8%

Basque Gold tok+morph - - - - - - - 82.3% 78.4%

Belarusian Raw text 99.4% 76.8% 88.2% 85.6% 71.7% 68.6% 81.3% 68.0% 60.6%

Belarusian Gold tok - - 88.7% 85.7% 72.4% 69.2% 81.5% 69.4% 61.9%

Belarusian Gold tok+morph - - - - - - - 76.8% 74.0%

Bulgarian Raw text 99.9% 93.9% 97.6% 94.6% 95.6% 94.0% 94.6% 88.8% 84.8%

Bulgarian Gold tok - - 97.7% 94.7% 95.6% 94.1% 94.7% 89.5% 85.5%

Bulgarian Gold tok+morph - - - - - - - 92.6% 89.1%

Catalan Raw text 100.0% 99.2% 98.0% 98.0% 97.1% 96.5% 97.9% 88.8% 85.7%

Catalan Gold tok - - 98.0% 98.0% 97.2% 96.5% 97.9% 88.8% 85.8%

Catalan Gold tok+morph - - - - - - - 91.1% 88.7%

Chinese Raw text 90.2% 98.8% 84.0% 83.8% 89.0% 82.7% 90.2% 62.9% 58.7%

Chinese Gold tok - - 92.2% 92.0% 98.7% 90.8% 100.0% 75.6% 70.1%

Chinese Gold tok+morph - - - - - - - 84.1% 81.4%

Coptic Raw text 65.8% 35.7% 62.6% 62.1% 65.7% 62.1% 64.6% 41.1% 39.3%

Coptic Gold tok - - 95.1% 94.3% 99.7% 94.2% 96.2% 83.2% 79.2%

Coptic Gold tok+morph - - - - - - - 88.1% 84.9%

Croatian Raw text 99.9% 97.0% 95.9% - 84.3% - 94.4% 83.6% 77.9%

Croatian Gold tok - - 96.0% - 84.4% - 94.4% 83.9% 78.1%

Croatian Gold tok+morph - - - - - - - 87.1% 83.2%

Czech Raw text 99.9% 91.6% 98.3% 92.8% 92.1% 91.7% 97.8% 86.8% 83.2%

Czech Gold tok - - 98.4% 92.9% 92.2% 91.9% 97.9% 87.7% 84.1%

Czech Gold tok+morph - - - - - - - 90.2% 87.5%

Czech-CAC Raw text 100.0% 99.8% 98.1% 90.6% 89.4% 89.1% 97.0% 86.9% 82.7%

Czech-CAC Gold tok - - 98.1% 90.7% 89.5% 89.1% 97.1% 87.0% 82.8%

Czech-CAC Gold tok+morph - - - - - - - 89.7% 86.6%

Czech-CLTT Raw text 99.5% 92.3% 96.5% 87.5% 87.8% 87.3% 96.8% 80.2% 76.6%

Czech-CLTT Gold tok - - 97.0% 87.9% 88.3% 87.7% 97.2% 81.0% 77.6%

Czech-CLTT Gold tok+morph - - - - - - - 83.8% 80.8%

Danish Raw text 99.8% 77.9% 95.2% - 94.2% - 94.9% 78.4% 74.7%

Danish Gold tok - - 95.5% - 94.5% - 95.0% 80.4% 76.6%

Danish Gold tok+morph - - - - - - - 85.6% 82.7%

Dutch Raw text 99.8% 77.6% 91.4% 88.1% 89.3% 87.0% 89.9% 75.4% 69.6%

Dutch Gold tok - - 91.8% 88.8% 89.9% 87.7% 90.1% 77.0% 71.2%

Dutch Gold tok+morph - - - - - - - 82.9% 79.4%

Dutch-LassySmall Raw text 100.0% 80.4% 97.6% - 97.2% - 98.1% 84.4% 82.0%

Dutch-LassySmall Gold tok - - 97.7% - 97.4% - 98.2% 87.5% 85.0%

Dutch-LassySmall Gold tok+morph - - - - - - - 89.7% 87.4%

English Raw text 99.0% 76.6% 93.5% 92.9% 94.4% 91.5% 96.0% 80.2% 77.2%

English Gold tok - - 94.5% 93.9% 95.4% 92.5% 96.9% 84.3% 81.2%

English Gold tok+morph - - - - - - - 87.8% 86.0%

English-LinES Raw text 99.9% 86.2% 95.0% 92.7% - - - 78.6% 74.4%

English-LinES Gold tok - - 95.1% 92.8% - - - 79.5% 75.3%

English-LinES Gold tok+morph - - - - - - - 84.1% 81.1%

English-ParTUT Raw text 99.6% 97.5% 94.2% 94.0% 93.3% 92.0% 96.9% 81.6% 77.9%

English-ParTUT Gold tok - - 94.6% 94.4% 93.6% 92.3% 97.3% 82.1% 78.4%

English-ParTUT Gold tok+morph - - - - - - - 86.4% 84.5%

Estonian Raw text 99.9% 94.2% 91.2% 93.2% 85.0% 83.2% 84.5% 72.4% 65.6%

Estonian Gold tok - - 91.3% 93.2% 85.0% 83.3% 84.5% 72.8% 66.0%

Estonian Gold tok+morph - - - - - - - 83.1% 79.6%

Finnish Raw text 99.7% 86.7% 94.5% 95.7% 91.5% 90.3% 86.5% 80.5% 76.9%

Finnish Gold tok - - 94.9% 96.0% 91.8% 90.7% 86.8% 82.0% 78.4%

Finnish Gold tok+morph - - - - - - - 86.9% 84.7%

Finnish-FTB Raw text 100.0% 86.4% 92.0% 91.0% 92.5% 89.2% 88.9% 80.1% 75.7%

Finnish-FTB Gold tok - - 92.2% 91.3% 92.7% 89.5% 88.9% 81.7% 77.3%

Finnish-FTB Gold tok+morph - - - - - - - 88.8% 86.5%

French Raw text 98.9% 94.6% 95.4% - 95.5% - 96.6% 84.2% 80.7%

French Gold tok - - 96.5% - 96.5% - 97.6% 85.4% 82.0%

French Gold tok+morph - - - - - - - 88.4% 86.0%

French-ParTUT Raw text 99.0% 97.8% 94.5% 94.2% 91.9% 90.8% 94.3% 82.9% 78.7%

French-ParTUT Gold tok - - 95.6% 95.3% 92.7% 91.6% 95.2% 84.1% 80.2%

French-ParTUT Gold tok+morph - - - - - - - 88.1% 85.3%

French-Sequoia Raw text 99.1% 84.0% 95.9% - 95.1% - 96.8% 83.2% 80.6%

French-Sequoia Gold tok - - 96.8% - 96.0% - 97.7% 85.1% 82.7%

French-Sequoia Gold tok+morph - - - - - - - 88.7% 87.4%

Galician Raw text 99.9% 95.8% 97.2% 96.7% 99.7% 96.4% 97.1% 81.0% 77.8%

Galician Gold tok - - 97.2% 96.8% 99.8% 96.4% 97.1% 81.2% 77.9%

Galician Gold tok+morph - - - - - - - 83.1% 80.5%

Galician-TreeGal Raw text 98.7% 86.7% 91.1% 87.8% 89.9% 87.0% 92.6% 71.5% 66.3%

Galician-TreeGal Gold tok - - 92.4% 88.8% 91.0% 88.0% 93.7% 74.4% 68.7%

Galician-TreeGal Gold tok+morph - - - - - - - 81.5% 77.1%

German Raw text 99.7% 79.3% 90.7% 94.7% 80.5% 76.3% 95.4% 74.0% 68.6%

German Gold tok - - 91.2% 95.0% 80.9% 76.7% 95.6% 76.5% 70.7%

German Gold tok+morph - - - - - - - 84.7% 82.2%

Gothic Raw text 100.0% 29.5% 94.2% 94.8% 87.6% 85.6% 92.9% 69.7% 63.5%

Gothic Gold tok - - 94.8% 95.3% 88.0% 86.5% 92.9% 78.8% 72.6%

Gothic Gold tok+morph - - - - - - - 82.2% 78.3%

Greek Raw text 99.9% 88.2% 95.8% 95.8% 90.3% 89.1% 94.5% 84.2% 80.4%

Greek Gold tok - - 96.0% 96.0% 90.5% 89.3% 94.6% 85.0% 81.1%

Greek Gold tok+morph - - - - - - - 87.9% 85.9%

Hebrew Raw text 85.2% 100.0% 80.9% 80.9% 77.6% 76.8% 79.6% 62.2% 57.9%

Hebrew Gold tok - - 95.1% 95.1% 91.3% 90.5% 93.2% 84.5% 78.9%

Hebrew Gold tok+morph - - - - - - - 87.8% 84.3%

Hindi Raw text 100.0% 99.1% 95.8% 94.9% 90.3% 87.7% 98.0% 91.3% 87.3%

Hindi Gold tok - - 95.8% 94.9% 90.3% 87.7% 98.0% 91.4% 87.3%

Hindi Gold tok+morph - - - - - - - 93.9% 91.0%

Hungarian Raw text 99.8% 96.2% 91.6% - 70.5% - 89.3% 74.1% 68.1%

Hungarian Gold tok - - 91.8% - 70.6% - 89.5% 74.5% 68.5%

Hungarian Gold tok+morph - - - - - - - 81.2% 78.5%

Indonesian Raw text 100.0% 92.0% 93.5% - 99.5% - - 80.6% 74.3%

Indonesian Gold tok - - 93.5% - 99.6% - - 80.8% 74.5%

Indonesian Gold tok+morph - - - - - - - 83.1% 79.1%

Irish Raw text 99.4% 94.3% 88.0% 86.9% 75.1% 72.7% 85.5% 72.5% 62.4%

Irish Gold tok - - 88.5% 87.4% 75.5% 73.1% 86.0% 73.3% 63.1%

Irish Gold tok+morph - - - - - - - 78.1% 71.4%

Italian Raw text 99.8% 97.1% 97.2% 97.0% 97.0% 96.1% 97.3% 88.8% 86.1%

Italian Gold tok - - 97.4% 97.2% 97.2% 96.3% 97.5% 89.3% 86.6%

Italian Gold tok+morph - - - - - - - 91.3% 89.7%

Japanese Raw text 91.9% 95.1% 89.1% - 91.8% - 91.1% 78.0% 76.6%

Japanese Gold tok - - 96.6% - 100.0% - 99.0% 93.4% 91.5%

Japanese Gold tok+morph - - - - - - - 95.6% 95.0%

Kazakh Raw text 94.0% 84.9% 52.0% 52.1% 47.2% 40.0% 59.2% 40.2% 23.9%

Kazakh Gold tok - - 55.4% 55.4% 50.1% 42.2% 63.1% 45.2% 27.0%

Kazakh Gold tok+morph - - - - - - - 60.5% 42.5%

Korean Raw text 99.7% 92.7% 94.4% 89.7% 99.3% 89.7% 99.4% 67.4% 60.5%

Korean Gold tok - - 94.7% 90.0% 99.6% 90.0% 99.7% 68.4% 61.5%

Korean Gold tok+morph - - - - - - - 71.7% 65.8%

Latin Raw text 100.0% 98.0% 83.4% 67.6% 72.5% 67.6% 51.2% 56.5% 46.0%

Latin Gold tok - - 83.4% 67.6% 72.5% 67.6% 51.2% 56.6% 46.1%

Latin Gold tok+morph - - - - - - - 67.8% 61.5%

Latin-ITTB Raw text 99.9% 82.5% 97.2% 92.7% 93.5% 91.3% 97.8% 79.7% 76.0%

Latin-ITTB Gold tok - - 97.3% 92.8% 93.6% 91.4% 97.9% 81.8% 78.1%

Latin-ITTB Gold tok+morph - - - - - - - 87.6% 85.2%

Latin-PROIEL Raw text 99.9% 31.0% 94.9% 95.0% 87.7% 86.7% 94.8% 66.1% 60.7%

Latin-PROIEL Gold tok - - 95.2% 95.2% 88.4% 87.4% 95.0% 75.3% 69.4%

Latin-PROIEL Gold tok+morph - - - - - - - 79.0% 75.0%

Latvian Raw text 99.2% 97.1% 89.6% 76.2% 83.2% 75.7% 87.6% 69.2% 62.8%

Latvian Gold tok - - 90.2% 76.8% 84.0% 76.3% 88.3% 70.3% 63.9%

Latvian Gold tok+morph - - - - - - - 78.7% 74.9%

Lithuanian Raw text 98.2% 92.0% 74.0% 73.0% 68.9% 63.7% 73.5% 44.0% 32.4%

Lithuanian Gold tok - - 74.6% 73.5% 69.7% 64.2% 74.2% 44.6% 33.0%

Lithuanian Gold tok+morph - - - - - - - 55.6% 46.5%

Norwegian-Bokmaal Raw text 99.8% 96.5% 96.9% - 95.3% - 96.6% 86.9% 84.1%

Norwegian-Bokmaal Gold tok - - 97.1% - 95.5% - 96.8% 87.5% 84.7%

Norwegian-Bokmaal Gold tok+morph - - - - - - - 91.7% 89.6%

Norwegian-Nynorsk Raw text 99.9% 92.2% 96.5% - 94.9% - 96.4% 85.6% 82.5%

Norwegian-Nynorsk Gold tok - - 96.6% - 95.0% - 96.5% 86.5% 83.3%

Norwegian-Nynorsk Gold tok+morph - - - - - - - 91.0% 88.6%

Old Church Slavonic Raw text 100.0% 40.5% 93.8% 93.8% 86.9% 85.7% 91.2% 73.6% 66.9%

Old Church Slavonic Gold tok - - 94.1% 94.1% 87.6% 86.5% 91.2% 81.6% 74.7%

Old Church Slavonic Gold tok+morph - - - - - - - 86.7% 82.2%

Persian Raw text 99.7% 98.2% 96.0% 96.0% 96.1% 95.4% 93.5% 83.3% 79.4%

Persian Gold tok - - 96.4% 96.3% 96.4% 95.7% 93.8% 83.8% 80.0%

Persian Gold tok+morph - - - - - - - 87.7% 84.9%

Polish Raw text 99.9% 99.7% 95.6% 84.0% 84.1% 83.1% 93.4% 86.7% 80.7%

Polish Gold tok - - 95.7% 84.1% 84.2% 83.3% 93.6% 87.0% 81.0%

Polish Gold tok+morph - - - - - - - 92.9% 89.5%

Portuguese Raw text 99.6% 89.4% 96.4% 72.7% 93.3% 71.6% 96.8% 86.0% 82.6%

Portuguese Gold tok - - 96.8% 73.0% 93.7% 71.9% 97.2% 87.2% 83.6%

Portuguese Gold tok+morph - - - - - - - 89.6% 87.5%

Portuguese-BR Raw text 99.9% 96.8% 97.0% 97.0% 99.7% 97.0% 98.8% 88.5% 86.3%

Portuguese-BR Gold tok - - 97.2% 97.2% 99.9% 97.2% 98.9% 88.8% 86.6%

Portuguese-BR Gold tok+morph - - - - - - - 90.5% 89.1%

Romanian Raw text 99.7% 93.9% 96.6% 95.9% 96.0% 95.7% 96.5% 85.6% 80.2%

Romanian Gold tok - - 96.9% 96.2% 96.3% 96.0% 96.8% 86.2% 80.8%

Romanian Gold tok+morph - - - - - - - 87.8% 83.0%

Russian Raw text 99.9% 96.9% 94.7% 94.4% 84.4% 82.8% 75.0% 80.3% 75.5%

Russian Gold tok - - 94.8% 94.5% 84.5% 82.9% 75.1% 80.8% 76.0%

Russian Gold tok+morph - - - - - - - 84.8% 81.9%

Russian-SynTagRus Raw text 99.6% 98.0% 98.0% - 93.6% - 95.6% 89.8% 87.2%

Russian-SynTagRus Gold tok - - 98.4% - 93.9% - 95.9% 90.4% 87.9%

Russian-SynTagRus Gold tok+morph - - - - - - - 91.8% 90.5%

Sanskrit Raw text 88.1% 29.0% 52.0% - 35.2% - 50.2% 38.8% 22.5%

Sanskrit Gold tok - - 57.6% - 43.6% - 60.6% 58.5% 34.3%

Sanskrit Gold tok+morph - - - - - - - 72.9% 58.5%

Slovak Raw text 100.0% 83.5% 93.2% 77.5% 79.7% 77.1% 85.9% 80.4% 75.2%

Slovak Gold tok - - 93.3% 77.6% 79.9% 77.2% 86.0% 82.0% 76.9%

Slovak Gold tok+morph - - - - - - - 88.2% 85.5%

Slovenian Raw text 99.9% 98.9% 96.2% 88.2% 88.5% 87.7% 95.3% 84.9% 81.6%

Slovenian Gold tok - - 96.2% 88.2% 88.6% 87.7% 95.4% 85.0% 81.7%

Slovenian Gold tok+morph - - - - - - - 91.8% 90.5%

Slovenian-SST Raw text 99.9% 17.8% 89.0% 81.1% 81.3% 78.6% 91.6% 53.0% 46.6%

Slovenian-SST Gold tok - - 89.4% 81.6% 81.8% 79.3% 91.7% 63.4% 56.0%

Slovenian-SST Gold tok+morph - - - - - - - 75.5% 70.6%

Spanish Raw text 99.7% 95.3% 95.5% - 96.1% - 95.9% 84.9% 81.4%

Spanish Gold tok - - 95.8% - 96.3% - 96.1% 85.5% 81.9%

Spanish Gold tok+morph - - - - - - - 88.0% 85.3%

Spanish-AnCora Raw text 99.9% 98.0% 98.1% 98.1% 97.5% 96.8% 98.1% 87.7% 84.5%

Spanish-AnCora Gold tok - - 98.2% 98.2% 97.5% 96.9% 98.1% 87.8% 84.7%

Spanish-AnCora Gold tok+morph - - - - - - - 90.2% 87.6%

Swedish Raw text 99.8% 94.6% 95.6% 93.9% 94.4% 92.8% 95.5% 81.4% 77.8%

Swedish Gold tok - - 95.8% 94.1% 94.6% 93.1% 95.7% 82.1% 78.4%

Swedish Gold tok+morph - - - - - - - 88.0% 85.0%

Swedish-LinES Raw text 100.0% 85.7% 94.8% 92.2% - - - 80.4% 75.7%

Swedish-LinES Gold tok - - 94.8% 92.3% - - - 81.3% 76.6%

Swedish-LinES Gold tok+morph - - - - - - - 86.0% 82.6%

Tamil Raw text 95.3% 89.2% 82.2% 77.7% 80.9% 77.2% 85.3% 59.5% 52.0%

Tamil Gold tok - - 85.8% 81.0% 84.2% 80.3% 89.1% 64.9% 56.5%

Tamil Gold tok+morph - - - - - - - 78.9% 71.8%

Turkish Raw text 98.1% 96.8% 92.4% 91.5% 87.3% 85.5% 90.2% 62.9% 55.8%

Turkish Gold tok - - 94.0% 93.0% 88.9% 87.0% 91.7% 65.5% 58.0%

Turkish Gold tok+morph - - - - - - - 66.8% 61.1%

Ukrainian Raw text 99.8% 95.1% 88.5% 70.7% 70.9% 67.6% 86.7% 69.9% 61.5%

Ukrainian Gold tok - - 88.6% 70.8% 71.0% 67.7% 86.9% 70.2% 61.8%

Ukrainian Gold tok+morph - - - - - - - 79.0% 74.5%

Urdu Raw text 100.0% 98.3% 92.4% 90.5% 80.6% 76.3% 93.0% 84.6% 77.6%

Urdu Gold tok - - 92.4% 90.5% 80.7% 76.3% 93.0% 84.7% 77.7%

Urdu Gold tok+morph - - - - - - - 88.2% 83.0%

Uyghur Raw text 99.8% 67.2% 74.7% 79.1% - - - 55.1% 35.0%

Uyghur Gold tok - - 75.1% 79.3% - - - 56.5% 35.8%

Uyghur Gold tok+morph - - - - - - - 62.3% 42.0%

Vietnamese Raw text 85.3% 92.9% 77.4% 75.4% 85.1% 75.4% 84.5% 46.9% 42.5%

Vietnamese Gold tok - - 89.3% 86.8% 99.6% 86.8% 99.0% 64.4% 57.2%

Vietnamese Gold tok+morph - - - - - - - 70.7% 67.9%

5.5 CoNLL17 Shared Task Baseline UD 2.0 Models

As part of CoNLL 2017 Shared Task in UD Parsing, baseline models for UDPipe were released. The CoNLL

2017 Shared Task models were trained on most of UD 2.0 treebanks (64 of them) and are distributed under the

CC BY-NC-SA licence.

Note that the models were released when the test set of UD 2.0 was unknown. Therefore, the models were trained

on a subset of training data only, to allow fair comparison on the development data (which were unused during

training and hyperparameter settings). Consequently, the performance of the models is not directly comparable

to other models. Details about the concrete data split, hyperparameter values and model performance are

available in the model archive.

5.5.1 Download

The CoNLL17 Shared Task Baseline UD 2.0 Models can be downloaded from LINDAT/CLARIN repository.

5.5.2 Acknowledgements

This work has been partially supported and has been using language resources and tools developed, stored and

distributed by the LINDAT/CLARIN project of the Ministry of Education, Youth and Sports of the Czech

Republic (project LM2015071 ).

The models were trained on a Universal Dependencies 2.0 treebanks.

5.6 Universal Dependencies 1.2 Models

Universal Dependencies 1.2 Models are distributed under the CC BY-NC-SA licence. The models are based

solely on Universal Dependencies 1.2 treebanks. The models work in UDPipe version 1.0.

Universal Dependencies 1.2 Models are versioned according to the date released in the format YYMMDD, where

YY,MM and DD are two-digit representation of year, month and day, respectively. The latest version is 160523.

5.6.1 Download

The latest version 160523 of the Universal Dependencies 1.2 models can be downloaded from LINDAT/CLARIN

repository.

5.6.2 Acknowledgements

This work has been partially supported and has been using language resources and tools developed, stored and

distributed by the LINDAT/CLARIN project of the Ministry of Education, Youth and Sports of the Czech

Republic (project LM2015071 ).

The models were trained on Universal Dependencies 1.2 treebanks.

For the UD treebanks which do not contain original plain text version, raw text is used to train the tokenizer

instead. The plain texts were taken from the W2C – Web to Corpus.

Publications

•(Straka et al. 2016) Straka Milan, Hajiˇc Jan, Strakov´a Jana. UDPipe: Trainable Pipeline for Processing

CoNLL-U Files Performing Tokenization, Morphological Analysis, POS Tagging and Parsing. LREC

2016, Portoroˇz, Slovenia, May 2016.

5.6.3 Model Description

The Universal Dependencies 1.2 models contain 36 models, each consisting of a tokenizer, tagger, lemmatizer

and dependency parser, all trained using the UD data. The model for Japanese is missing, because we do not

have the license for the required corpus of Mainichi Shinbun 1995.

The tokenizer is trained using the SpaceAfter=No features. If the features are not present in the data, they can

be ﬁlled in using raw text in the language in question (surprisingly, quite little data suﬃces, we use 500kB).

The tagger, lemmatizer and parser are trained using gold UD data.

Details about model architecture and training process can be found in the (Straka et al. 2016) paper.

5.6.4 Model Performance

We present the tagger, lemmatizer and parser performance, measured on the testing portion of the data.

Only the segmentation and the tokenization of the testing data is retained before evaluation. Therefore, the

dependency parser is evaluated without gold POS tags.

Treebank UPOS XPOS Feats All Tags Lemma UAS LAS

Ancient Greek 91.1% 77.8% 88.7% 77.7% 86.9% 68.1% 61.6%

Ancient Greek-PROIEL 96.7% 96.4% 89.3% 88.4% 93.4% 75.8% 69.6%

Arabic 98.8% 97.7% 97.8% 97.6% - 80.4% 75.6%

Basque 93.3% - 87.2% 85.4% 93.5% 74.8% 69.5%

Bulgarian 97.8% 94.8% 94.4% 93.1% 94.6% 89.0% 84.2%

Croatian 94.9% - 85.5% 85.0% 93.1% 78.6% 71.0%

Czech 98.4% 93.2% 92.6% 92.2% 97.8% 86.9% 83.0%

Danish 95.8% - 94.8% 93.6% 95.2% 78.6% 74.8%

Dutch 89.7% 88.7% 91.2% 86.4% 88.9% 78.1% 70.7%

English 94.5% 93.8% 95.4% 92.5% 97.0% 84.2% 80.6%

Estonian 88.0% 73.7% 80.0% 73.6% 77.0% 79.9% 71.5%

Finnish 94.9% 96.0% 93.2% 92.1% 86.8% 81.0% 76.5%

Finnish-FTB 94.0% 91.6% 93.3% 91.2% 89.1% 81.5% 76.9%

French 95.8% - - 95.8% - 82.8% 78.4%

German 90.5% - - 90.5% - 78.2% 72.2%

Gothic 95.5% 95.7% 88.0% 86.3% 93.4% 76.4% 68.2%

Greek 97.3% 97.3% 92.8% 91.7% 94.8% 80.3% 76.5%

Hebrew 94.9% 94.9% 91.3% 90.5% - 82.6% 76.8%

Hindi 95.8% 94.8% 90.2% 87.7% 98.0% 91.7% 87.5%

Hungarian 92.6% - 89.9% 88.9% 86.9% 77.0% 70.6%

Indonesian 93.5% - - 93.5% - 79.9% 73.3%

Irish 91.8% 90.3% 79.4% 76.6% 87.3% 74.4% 66.1%

Italian 97.2% 97.0% 97.1% 96.2% 97.7% 88.6% 85.8%

Latin 91.2% 75.8% 79.3% 75.6% 79.9% 57.1% 46.7%

Latin-ITT 98.8% 94.0% 94.6% 93.8% 98.3% 79.9% 76.4%

Latin-PROIEL 96.4% 96.0% 88.9% 88.2% 95.3% 75.3% 68.3%

Norwegian 97.2% - 95.5% 94.7% 96.9% 86.7% 84.1%

Old Church Slavonic 95.3% 95.1% 89.1% 88.2% 92.9% 80.6% 73.4%

Persian 97.0% 96.3% 96.5% 96.2% - 83.8% 79.4%

Polish 95.8% 84.0% 84.1% 83.8% 92.8% 86.3% 79.6%

Portuguese 97.6% 92.3% 95.3% 92.0% 97.8% 85.8% 81.9%

Romanian 89.0% 81.0% 82.3% 81.0% 75.3% 68.6% 56.9%

Slovenian 95.7% 88.2% 88.6% 87.5% 95.0% 84.1% 80.3%

Spanish 95.3% - 95.9% 93.4% 96.3% 84.2% 80.3%

Swedish 95.8% 93.9% 94.8% 93.2% 95.5% 81.4% 77.1%

Tamil 85.9% 80.8% 84.3% 80.2% 88.0% 67.2% 58.8%

6 UDPipe API Reference

The UDPipe API is deﬁned in header udpipe.h and resides in ufal::udpipe namespace. The API allows only

using existing models, for custom model creation you have to use the train parser binary.

The strings used in the UDPipe API are always UTF-8 encoded (except from ﬁle paths, whose encoding is

system dependent).

6.1 UDPipe Versioning

UDPipe is versioned using Semantic Versioning. Therefore, a version consists of three numbers ma-

jor.minor.patch, optionally followed by a hyphen and pre-release version info, with the following semantics:

•Stable versions have no pre-release version info, development have non-empty pre-release version info.

•Two versions with the same major.minor have the same API with the same behaviour, apart from bugs.

Therefore, if only patch is increased, the new version is only a bug-ﬁx release.

•If two versions vand uhave the same major, but minor(v) is greater than minor(u), version vcontains

only additions to the API. In other words, the API of uis all present in vwith the same behaviour (once

again apart from bugs). It is therefore safe to upgrade to a newer UDPipe version with the same major.

•If two versions diﬀer in major, their API may diﬀer in any way.

Models created by UDPipe have the same behaviour in all UDPipe versions with same major, apart from obvious

bugﬁxes. On the other hand, models created from the same data by diﬀerent major.minor UDPipe versions

may have diﬀerent behaviour.

6.2 Struct string piece

struct string_piece {

const char* str;

size_t len;

string_piece();

string_piece(const char* str);

string_piece(const char* str, size_t len);

string_piece(const std::string& str);

}

The string piece is used for eﬃcient string passing. The string referenced in string piece is not owned by

it, so users have to make sure the referenced string exists as long as the string piece.

6.3 Class token

class token {

public:

string form;

string misc;

token(string_piece form = string_piece(), string_piece misc = string_piece());

// CoNLL-U defined SpaceAfter=No feature

bool get_space_after() const;

void set_space_after(bool space_after);

// UDPipe-specific all-spaces-preserving SpacesBefore and SpacesAfter features

void get_spaces_before(string& spaces_before) const;

void set_spaces_before(string_piece spaces_before);

void get_spaces_after(string& spaces_after) const;

void set_spaces_after(string_piece spaces_after);

void get_spaces_in_token(string& spaces_in_token) const;

void set_spaces_in_token(string_piece spaces_in_token);

// UDPipe-specific TokenRange feature

bool get_token_range(size_t& start, size_t& end) const;

void set_token_range(size_t start, size_t end);

};

The token class represents a sentence token, with form and misc ﬁelds corresponding to CoNLL-U ﬁelds. The

token class acts mostly as a parent to word and multiword token classes.

The class also oﬀers several methods for manipulating features in the misc ﬁeld. Notably, UDPipe uses custom

misc ﬁelds to store all spaces in the original document. This markup is backward compatible with CoNLL-U v2

SpaceAfter=No feature. This markup can be utilized by plaintext output format, which allows reconstructing

the original document.

The markup uses the following misc ﬁelds:

•SpacesBefore=content (by default empty): spaces/other content preceding the token

•SpacesAfter=content (by default a space if SpaceAfter=No feature is not present, empty otherwise):

spaces/other content following the token

•SpacesInToken=content (by default equal to the FORM of the token): FORM of the token including

original spaces (this is needed only if tokens are allowed to contain spaces and a token contains a tab or

newline characters)

The content of all above three ﬁelds must be escaped to allow storing tabs and newlines. The following C-like

schema is used:

• \s: space

• \t: tab

• \r: CR character

• \n: LF character

• \p:|(pipe character)

• \\:\(backslash character)

6.3.1 token::get space after()

bool get_space_after() const;

Returns true if the token should be followed by a spaces, false if not, according to the absence or presence of

the SpaceAfter=No feature in the misc ﬁeld.

6.3.2 token::set space after()

void set_space_after(bool space_after);

Adds or removes the SpaceAfter=No feature in the misc ﬁeld.

6.3.3 token::get spaces before()

void get_spaces_before(string& spaces_before) const;

Return spaces preceding current token, stored in the SpacesBefore feature in the misc ﬁeld. If SpacesBefore

is not present, empty string is returned.

6.3.4 token::set spaces before()

void set_spaces_before(string_piece spaces_before);

Set the SpacesBefore feature in the misc ﬁeld.

6.3.5 token::get spaces after()

void get_spaces_after(string& spaces_after) const;

Return spaces after current token, stored in the SpacesAfter feature in the misc ﬁeld.

If SpacesAfter is not present and SpaceAfter=No is present, return an empty string; if neither feature is

present, one space is returned.

6.3.6 token::set spaces after()

void set_spaces_after(string_piece spaces_after);

Set the SpacesAfter and SpaceAfter=No features in the misc ﬁeld.

6.3.7 token::get spaces in token()

void get_spaces_in_token(string& spaces_in_token) const;

Return the value of the SpacesInToken feature, if present. Otherwise, empty string is returned.

6.3.8 token::set spaces in token()

void set_spaces_in_token(string_piece spaces_in_token);

Set the SpacesInToken feature in the misc ﬁeld.

6.3.9 token::get token range()

bool get_token_range(size_t& start, size_t& end) const;

If present, return the value of the TokenRange feature in the misc ﬁeld. The format of the feature (inspired by

Python) is TokenRange=start:end, where start is zero-based document-level index of the start of the token

(counted in Unicode characters) and end is zero-based document-level index of the ﬁrst character following the

token (i.e., the length of the token is end-start).

6.3.10 token::set token range()

void set_token_range(size_t start, size_t end);

Set the TokenRange feature in the misc ﬁeld. If string::npos is passed in the start argument, TokenRange

feature is removed from the misc ﬁeld.

6.4 Class word

class word : public token {

public:

// form and misc are inherited from token

int id; // 0 is root, >0 is sentence word, <0 is undefined

string lemma; // lemma

string upostag; // universal part-of-speech tag

string xpostag; // language-specific part-of-speech tag

string feats; // list of morphological features

int head; // head, 0 is root, <0 is undefined

string deprel; // dependency relation to the head

string deps; // secondary dependencies

vector<int> children;

word(int id = -1, string_piece form = string_piece());

};

The word class represents a sentence word. The word ﬁelds correspond to CoNLL-U ﬁelds, with the children

ﬁeld representing the opposite direction of head links (the elements of the children array are in ascending

order).

6.5 Class multiword token

class multiword_token : public token {

public:

// form and misc are inherited from token

int id_first, id_last;

multiword_token(int id_first = -1, int id_last = -1, string_piece form = string_piece(),

string_piece misc = string_piece());

};

The multiword token represents a multi-word token described in CoNLL-U format. The multi-word token has

aform and a misc ﬁeld, other CoNLL-U word ﬁelds are guaranteed to be empty.

6.6 Class empty node

class empty_node {

public:

int id; // 0 is root, >0 is sentence word, <0 is undefined

int index; // index for the current id, should be numbered from 1, 0=undefined

string form; // form

string lemma; // lemma

string upostag; // universal part-of-speech tag

string xpostag; // language-specific part-of-speech tag

string feats; // list of morphological features

string deps; // secondary dependencies

string misc; // miscellaneous information

empty_node(int id = -1, int index = 0) : id(id), index(index) {}

};

The empty node class represents an empty node from CoNLL-U 2.0, with the ﬁelds corresponding to CoNLL-U

ﬁelds. For a speciﬁed id, the index are numbered sequentially from 1.

6.7 Class sentence

class sentence {

public:

sentence();

vector<word> words;

vector<multiword_token> multiword_tokens;

vector<empty_node> empty_nodes;

vector<string> comments;

static const string root_form;

// Basic sentence modifications

bool empty();

void clear();

word&add_word(string_piece form = string_piece());

void set_head(int id, int head, const string& deprel);

void unlink_all_words();

// CoNLL-U defined comments

bool get_new_doc(string* id = nullptr) const;

void set_new_doc(bool new_doc, string_piece id = string_piece());

bool get_new_par(string* id = nullptr) const;

void set_new_par(bool new_par, string_piece id = string_piece());

bool get_sent_id(string& id) const;

void set_sent_id(string_piece id);

bool get_text(string& text) const;

void set_text(string_piece text);

};

The sentence class represents a sentence CoNLL-U sentence, which consists of:

•sequence of words stored in ascending order, with the ﬁrst word (with index 0) always being a technical

root with form root form

•sequence of multiword tokens also stored in ascending order

•sequence of empty nodes also stored in ascending order

•comments

Although you can manipulate the words directly, the sentence class oﬀers several simple node manipulation

methods. There are also several methods manipulating CoNLL-U v2 comments.

6.7.1 sentence::empty()

bool empty();

Returns true if the sentence is empty. i.e., if it contains only a technical root node.

6.7.2 sentence::clear()

void clear();

Removes all words, multi-word tokens and comments (only the technical root word is kept).

6.7.3 sentence::add word()

word& add_word(string_piece form = string_piece());

Adds a new word to the sentence. The new word has ﬁrst unused id, speciﬁed form and is not linked to any

other node. Reference to the new word is returned so that other ﬁelds can be also ﬁlled.

6.7.4 sentence:set head()

void set_head(int id, int head, const std::string& deprel);

Link the word id to the word head, with the speciﬁed dependency relation. If the head is negative, the word

id is unlinked from its current head, if any.

6.7.5 sentence::unlink all words()

void unlink_all_words();

Unlink all words.

6.7.6 sentence::get new doc()

bool get_new_doc(string* id = nullptr) const;

Return true if # newdoc comment is present. Optionally, document id is also returned (in # newdoc id = ...

format).

6.7.7 sentence::set new doc()

void set_new_doc(bool new_doc, string_piece id = string_piece());

Adds/removes # newdoc comment, optionally with a given document id.

6.7.8 sentence::get new par()

bool get_new_par(string* id = nullptr) const;

Return true if # newpar comment is present. Optionally, paragraph id is also returned (in # newpar id =

... format).

6.7.9 sentence::set new par()

void set_new_par(bool new_par, string_piece id = string_piece());

Adds/removes # newpar comment, optionally with a given paragraph id.

6.7.10 sentence::get sent id()

bool get_sent_id(string& id) const;

Return true if # sent id = ... comment is present, and ﬁll given id with sentence id. Otherwise, return

false and clear id.

6.7.11 sentence::set sent id()

void set_sent_id(string_piece id);

Set the # sent id = ... comment using given sentence id; if the sentence id is empty, remove all present #

sent id comment.

6.7.12 sentence::get text()

bool get_text(string& text) const;

Return true if # text = ... comment is present, and ﬁll given text with sentence text. Otherwise, return

false and clear text.

6.7.13 sentence::set text()

void set_text(string_piece text);

Set the # text = ... comment using given text; if the given text is empty, remove all present # text comment.

6.8 Class input format

class input_format {

public:

virtual ~input_format() {}

virtual bool read_block(istream& is, string& block) const = 0;

virtual void reset_document(string_piece id = string_piece()) = 0;

virtual void set_text(string_piece text, bool make_copy = false) = 0;

virtual bool next_sentence(sentence& s, string& error) = 0;

// Static factory methods

static input_format*new_input_format(const string& name);

static input_format*new_conllu_input_format(const string& options = std::string());

static input_format*new_generic_tokenizer_input_format(const string& options =

std::string());

static input_format*new_horizontal_input_format(const string& options = std::string());

static input_format*new_vertical_input_format(const string& options = std::string());

static input_format*new_presegmented_tokenizer(input_format* tokenizer);

static const string CONLLU_V1;

static const string CONLLU_V2;

static const string GENERIC_TOKENIZER_NORMALIZED_SPACES;

static const string GENERIC_TOKENIZER_PRESEGMENTED;

static const string GENERIC_TOKENIZER_RANGES;

};

The input format class allows loading sentences in various formats.

Th class instances may store internal state and are not thread-safe.

6.8.1 input format::read block()

virtual bool read_block(istream& is, string& block) const = 0;

Read a portion of input, which is guaranteed to contain only complete sentences. Such portion is usually a

paragraph (text followed by an empty line) or a line, but it may be more complex (i.e., in a XML-like format).

6.8.2 input format::reset document()

virtual void reset_document(string_piece id = string_piece()) = 0;

Resets the input format instance state. Such state is needed not only for remembering unprocessed text of the

last set text call, but also for correct inter-block state tracking (for example to track document-level ranges

or inter-sentence spaces – if you pass only spaces to set text, these spaces has to accumulate and be returned

as preceding spaces of the next sentence).

If applicable, ﬁrst read sentence will have the # newdoc comment, optionally with given document id.

6.8.3 input format::set text()

virtual void set_text(string_piece text, bool make_copy = false) = 0;

Set the text from which the sentences will be read.

If make copy is false, only a reference to the given text is stored and the user has to make sure it exists until

the instance is destroyed or set text is called again. If make copy is true, a copy of the given text is made

and retained until the instance is destroyed or set text is called again.

6.8.4 input format::next sentence()

virtual bool next_sentence(sentence& s, string& error) = 0;

Try reading another sentence from the text speciﬁed by set text. Returns true if the sentence was read and

false if the text ended or there was a read error. The latter two conditions can be distinguished by the error

parameter – if it is empty, the text ended, if it is nonempty, it contains a description of the read error.

6.8.5 input format::new input format()

static input_format* new_input_format(const string& name);

Create new input format instance, given its name. The individual input formats can be parametrized by using

format=data syntax. The following input formats are currently supported:

•conllu: return the new conllu input format

•generic tokenizer: return the new generic tokenizer input format

•horizontal: return the new horizontal input format

•vertical: return the new vertical input format

The new instance must be deleted after use.

6.8.6 input format::new conllu input format()

static input_format* new_conllu_input_format(const string() options = std::string());

Create input format instance which loads sentences in the CoNLL-U format. The new instance must be deleted

after use.

Supported options:

•v2 (default): use CoNLL-U v2

•v1: allow loading only CoNLL-U v1 (i.e., no empty nodes and no spaces in forms and lemmas)

6.8.7 input format::new generic tokenizer input format()

static input_format* new_generic_tokenizer_input_format(const string() options =

std::string());

Create rule-based generic tokenizer for English-like languages (with spaces separating tokens and English-like

punctuation). The new instance must be deleted after use.

Supported options:

•normalized spaces: by default, UDPipe uses custom misc ﬁelds to exactly encode spaces in the original

document. If normalized spaces option is given, only standard CoNLL-U v2 markup (SpaceAfter=No

and # newpar) is used.

•presegmented: input is assumed to be already segmented, with every sentence on a line, and is only

tokenized (respecting sentence breaks)

•ranges: for every token, range in the original document is stored in a format described in token class

6.8.8 input format::new horizontal input format()

static input_format* new_horizontal_input_format(const string() options = std::string());

Create input format instance which loads forms from a simple horizontal format – each sentence on a line,

with word forms separated by spaces. The new instance must be deleted after use.

In order to allow spaces in tokens, Unicode character ’NO-BREAK SPACE’ (U+00A0) is considered part of

token and converted to a space during loading.

6.8.9 input format::new vertical input format()

static input_format* new_vertical_input_format(const string() options = std::string());

Create input format instance which loads forms from a simple vertical format – each word on a line, with

empty line denoting end of sentence. The new instance must be deleted after use.

6.8.10 input format::new presegmented tokenizer()

static input_format* new_presegmented_tokenizer(input_format* tokenizer);

Create input format instance which acts as a tokenizer adapter – given a tokenizer which segments anywhere,

it creates a tokenizer which segments on newline characters (by calling the tokenizer on individual lines, and if

the tokenizer segments in the middle of the line, it calls it repeatedly and merges the results).

The new instance must be deleted after use. Note that the new instance takes ownership of the given tokenizer

and deletes it during its own deletion.

6.9 Class output format

class output_format {

public:

virtual ~output_format() {}

virtual void write_sentence(const sentence& s, ostream& os) = 0;

virtual void finish_document(ostream& os) {};

// Static factory methods

static output_format*new_output_format(const string& name);

static output_format*new_conllu_output_format(const string() options = std::string());

static output_format*new_epe_output_format(const string() options = std::string());

static output_format*new_matxin_output_format(const string() options = std::string());

static output_format*new_horizontal_output_format(const string() options =

std::string());

static output_format*new_plaintext_output_format(const string() options = std::string());

static output_format*new_vertical_output_format(const string() options = std::string());

static const string CONLLU_V1;

static const string CONLLU_V2;

static const string HORIZONTAL_PARAGRAPHS;

static const string PLAINTEXT_NORMALIZED_SPACES;

static const string VERTICAL_PARAGRAPHS;

};

The output format class allows printing sentences in various formats.

The class instances may store internal state and are not thread-safe.

6.9.1 output format::write sentence()

virtual void write_sentence(const sentence& s, ostream& os) = 0;

Write given sentence to the given output stream.

When the output format requires document-level markup, it is written automatically when the ﬁrst sentence is

written using this output format instance (or after finish document call).

6.9.2 output format::ﬁnish document()

virtual void finish_document(ostream& os) {};

When the output format requires document-level markup, write the end-of-document mark and reset the

output format instance state (i.e., the next write sentence will write start-of-document mark).

6.9.3 output format::new output format()

static output_format* new_output_format(const string& name);

Create new output format instance, given its name. The following output formats are currently supported:

•conllu: return the new conllu output format

•epe: return the new epe output format

•matxin: return the new matxin output format

•horizontal: return the new horizontal output format

•plaintext: return the new plaintext output format

•vertical: return the new vertical output format

The new instance must be deleted after use.

6.9.4 output format::new conllu output format()

static output_format* new_conllu_output_format(const string() options = std::string());

Creates output format instance for writing sentences in the CoNLL-U format. The new instance must be

deleted after use.

Supported options:

•v2 (default): use CoNLL-U v2

•v1: produce output in CoNLL-U v1 format. Note that this is a lossy process, as empty nodes are ignored

and spaces in forms and lemmas are converted to underscores.

6.9.5 output format::new epe output format()

static output_format* new_epe_output_format(const string() options = std::string());

Creates output format instance for writing sentences in the EPE (Extrinsic Parser Evaluation 2017) interchange

format. The new instance must be deleted after use.

6.9.6 output format::new matxin output format()

static output_format* new_matxin_output_format(const string() options = std::string());

Creates output format instance for writing sentences in the Matxin format – UDPipe produces a XML with

the following DTD:

<!ELEMENT corpus (SENTENCE*)>

<!ELEMENT SENTENCE (NODE*)>

<!ATTLIST SENTENCE ord CDATA #REQUIRED

alloc CDATA #REQUIRED>

<!ELEMENT NODE (NODE*)>

<!ATTLIST NODE ord CDATA #REQUIRED

alloc CDATA #REQUIRED

form CDATA #REQUIRED

lem CDATA #REQUIRED

mi CDATA #REQUIRED

si CDATA #REQUIRED

sub CDATA #REQUIRED>

The new instance must be deleted after use.

6.9.7 output format::new plaintext output format()

static output_format* new_plaintext_output_format(const string() options = std::string());

Creates output format instance for writing sentence tokens (in the UD sense) using original spacing. By

default, UDPipe custom misc features (see description of token class) are used to reconstruct the exact original

spaces. However, if the document does not contain these features or if only normalized spacing is wanted, you

can use the following option:

•normalized spaces: write one sentence on a line, and either one or no space between tokens, using the

SpaceAfter=No feature

6.9.8 output format::new horizontal output format()

static output_format* new_horizontal_output_format(const string() options = std::string());

Creates output format instance for writing sentences in a simple horizontal format – each sentence on a line,

with word forms separated by spaces. The new instance must be deleted after use.

Because words can contain spaces in CoNLL-U v2, the spaces in words are converted to Unicode character

’NO-BREAK SPACE’ (U+00A0).

Supported options:

•paragraphs: if given, an empty line is printed after the end of a paragraph or a document (recognized by

# newpar or # newdoc comments)

6.9.9 output format::new vertical output format()

static output_format* new_vertical_output_format(const string() options = std::string());

Creates output format instance for writing sentences in a simple vertical format – each word form on a line,

with empty line denoting end of sentence. The new instance must be deleted after use.

Supported options:

•paragraphs: if given, an empty line is printed after the end of a paragraph or a document (recognized by

# newpar or # newdoc comments)

6.10 Class model

class model {

public:

virtual ~model() {}

static model*load(const char* fname);

static model*load(istream& is);

virtual input_format*new_tokenizer(const string& options) const = 0;

virtual bool tag(sentence& s, const string& options, string& error) const = 0;

virtual bool parse(sentence& s, const string& options, string& error) const = 0;

static const string DEFAULT;

static const string TOKENIZER_NORMALIZED_SPACES;

static const string TOKENIZER_PRESEGMENTED;

static const string TOKENIZER_RANGES;

};

Class representing UDPipe model, allowing to perform tokenization, tagging and parsing.

6.10.1 model::load(const char*)

static model* load(const char* fname);

Load a new model from a given ﬁle, returning NULL on failure. The new instance must be deleted after use.

6.10.2 model::load(istream&)

static model* load(istream& is);

Load a new model from a given input stream, returning NULL on failure. The new instance must be deleted

after use.

6.10.3 model::new tokenizer()

virtual input_format* new_tokenizer(const string& options) const = 0;

Construct a new tokenizer (or NULL if no tokenizer is speciﬁed by the model). The new instance must be deleted

after use.

6.10.4 model::tag()

virtual bool tag(sentence& s, const string& options, string& error) const = 0;

Tag the given sentence.

6.10.5 model::parse()

virtual bool parse(sentence& s, const string& options, string& error) const = 0;

Parse the given sentence.

6.11 Class pipeline

class pipeline {

public:

pipeline(const model* m, const string& input, const string& tagger, const string& parser,

const string& output);

void set_model(const model* m);

void set_input(const string& input);

void set_tagger(const string& tagger);

void set_parser(const string& parser);

void set_output(const string& output);

void set_immediate(bool immediate);

void [set_document_id #pipeline_set_document_id[(const string& document_id);

bool process(istream& is, ostream& os, string& error) const;

static const string DEFAULT;

static const string NONE;

};

The pipeline class allows simple ﬁle-to-ﬁle processing. A model and input/tagger/parser/output options can

be speciﬁed in the pipeline.

The input ﬁle can be processed either after fully loaded (default), or in immediate mode, in which case is

the input processed and printed as soon as a block of input guaranteed to contain whole sentences is loaded.

Speciﬁcally, for most input formats the input is processed after loading an empty line (with the exception of

horizontal input format and presegmented tokenizer, where the input is processed after loading every line).

6.11.1 pipeline::set model()

void set_model(const model* m);

Use the given model.

6.11.2 pipeline::set input()

void set_input(const string& input);

Use the given input format. In addition to formats described in new input format, a special tokenizer or

tokenizer=options format allows using the model tokenizer.

6.11.3 pipeline::set tagger()

void set_tagger(const string& tagger);

Use the given tagger options.

6.11.4 pipeline::set parser()

void set_parser(const string& parser);

Use the given parser options.

6.11.5 pipeline::set output()

void set_output(const string& output);

Use the given output format (see new output format for a list).

6.11.6 pipeline::set immediate()

void set_immediate(bool immediate);

Set or reset the immediate mode (default is immediate=false).

6.11.7 pipeline::set document id()

void set_document_id(const string& document_id);

Set document id, which is passed to input format::reset document).

6.11.8 pipeline::process()

bool process(istream& is, ostream& os, string& error) const;

Process the given input stream, writing results to the given output stream. If the processing succeeded, true

is returned; otherwise, false is returned with an error stored in the error argument.

6.12 Class trainer

class trainer {

public:

static bool train(const string& method, const vector<sentence>& train, const

vector<sentence>& heldout,

const string& tokenizer, const string& tagger, const string& parser,

ostream& os, string& error);

static const string DEFAULT;

static const string NONE;

};

Class allowing training a UDPipe model.

6.12.1 trainer::train()

static bool train(const string& method, const vector<sentence>& train, const

vector<sentence>& heldout,

const string& tokenizer, const string& tagger, const string& parser,

ostream& os, string& error);

Train a UDPipe model. The only supported method is currently morphodita parsito. Use the supplied train

and heldout data, and given tokenizer, tagger and parser options (see the Training UDPipe Models section in

the User’s Manual).

If the training succeeded, true is returned and the model is saved to the given os stream; otherwise, false is

returned with an error stored in the error argument.

6.13 Class evaluator

class evaluator {

public:

evaluator(const model* m, const string& tokenizer, const string& tagger, const string&

parser);

void set_model(const model* m);

void set_tokenizer(const string& tokenizer);

void set_tagger(const string& tagger);

void set_parser(const string& parser);

bool evaluate(istream& is, ostream& os, string& error) const;

static const string DEFAULT;

static const string NONE;

};

Class evaluating performance of given model on CoNLL-U ﬁle.

Three diﬀerent settings (depending on whether tokenizer, tagger and parser is used) can be evaluated. For

details, see Measuring Model Accuracy in User’s Manual.

6.13.1 evaluator::set model()

void set_model(const model* m);

Use the given model.

6.13.2 evaluator::set tokenizer()

void set_tokenizer(const string& tokenizer);

Use the given tokenizer options; pass DEFAULT to use default options or NONE not to use a tokenizer.

6.13.3 evaluator::set tagger()

void set_tagger(const string& tagger);

Use the given tagger options; pass DEFAULT to use default options or NONE not to use a tagger.

6.13.4 evaluator::set parser()

void set_parser(const string& parser);

Use the given parser options; pass DEFAULT to use default options or NONE not to use a parser.

6.13.5 evaluator::evaluate()

bool evaluate(istream& is, ostream& os, string& error) const;

Evaluate the speciﬁed model on the given CoNLL-U input read from is stream.

If the evaluation succeeded, true is returned and the evaluation results are written to the os stream in a plain

text format; otherwise, false is returned with an error stored in the error argument.

6.14 Class version

class version {

public:

unsigned major;

unsigned minor;

unsigned patch;

string prerelease;

static version current();

};

The version class represents UDPipe version. See UDPipe Versioning for more information.

6.14.1 version::current

static version current();

Returns current UDPipe version.

6.15 C++ Bindings API

Bindings for other languages than C++ are created using SWIG from the C++ bindings API, which is a slightly

modiﬁed version of the native C++ API. Main changes are replacement of string piece type by native strings

and removal of methods using istream. Here is the C++ bindings API declaration:

6.15.1 Helper Structures

typedef vector<int> Children;

typedef vector<string> Comments;

class ProcessingError {

public:

bool occurred();

string message;

};

class Token {

public:

string form;

string misc;

Token(const string& form = string(), const string& misc = string());

// CoNLL-U defined SpaceAfter=No feature

bool getSpaceAfter() const;

void setSpaceAfter(bool space_after);

// UDPipe-specific all-spaces-preserving SpacesBefore and SpacesAfter features

string getSpacesBefore() const;

void setSpacesBefore(const string& spaces_before);

string getSpacesAfter() const;

void setSpacesAfter(const string& spaces_after);

string getSpacesInToken() const;

void setSpacesInToken(const string& spaces_in_token);

// UDPipe-specific TokenRange feature

bool getTokenRange() const;

size_t getTokenRangeStart() const;

size_t getTokenRangeEnd() const;

void setTokenRange(size_t start, size_t end);

};

class Word : public Token {

public:

// form and misc are inherited from token

int id; // 0 is root, >0 is sentence word, <0 is undefined

string lemma; // lemma

string upostag; // universal part-of-speech tag

string xpostag; // language-specific part-of-speech tag

string feats; // list of morphological features

int head; // head, 0 is root, <0 is undefined

string deprel; // dependency relation to the head

string deps; // secondary dependencies

Children children;

Word(int id = -1, const string& form = string());

};

typedef vector<Word> Words;

class MultiwordToken : public Token {

public:

// form and misc are inherited from token

int idFirst, idLast;

MultiwordToken(int id_first = -1, int id_last = -1, const string& form = string(), const

string& misc = string());

};

typedef vector<MultiwordToken> MultiwordTokens;

class EmptyNode {

public:

int id; // 0 is root, >0 is sentence word, <0 is undefined

int index; // index for the current id, should be numbered from 1, 0=undefined

string form; // form

string lemma; // lemma

string upostag; // universal part-of-speech tag

string xpostag; // language-specific part-of-speech tag

string feats; // list of morphological features

string deps; // secondary dependencies

string misc; // miscellaneous information

EmptyNode(int id = -1, int index = 0) : id(id), index(index) {}

};

typedef vector<empty_node> EmptyNodes;

class Sentence {

public:

Sentence();

Words words;

MultiwordTokens multiwordTokens;

EmptyNodes emptyNodes;

Comments comments;

static const string rootForm;

// Basic sentence modifications

bool empty();

void clear();

virtual Word& addWord(const char* form);

void setHead(int id, int head, const string& deprel);

void unlinkAllWords();

// CoNLL-U defined comments

bool getNewDoc() const;

string getNewDocId() const;

void setNewDoc(bool new_doc, const string& id = string());

bool getNewPar() const;

string getNewParId() const;

void setNewPar(bool new_par, const string& id = string());

string getSentId() const;

void setSentId(const string& id);

string getText() const;

void setText(const string& id);

};

typedef vector<Sentence> Sentences;

6.15.2 Main Classes

class InputFormat {

public:

virtual void resetDocument(const string& id = string());

virtual void setText(const char* text);

virtual bool nextSentence(Sentence& s, ProcessingError* error = nullptr);

static InputFormat* newInputFormat(const string& name);

static InputFormat* newConlluInputFormat(const string& id = string());

static InputFormat* newGenericTokenizerInputFormat(const string& id = string());

static InputFormat* newHorizontalInputFormat(const string& id = string());

static InputFormat* newVerticalInputFormat(const string& id = string());

static InputFormat* newPresegmentedTokenizer(InputFormat tokenizer);

static const string CONLLU_V1;

static const string CONLLU_V2;

static const string GENERIC_TOKENIZER_NORMALIZED_SPACES;

static const string GENERIC_TOKENIZER_PRESEGMENTED;

static const string GENERIC_TOKENIZER_RANGES;

};

class OutputFormat {

public:

virtual string writeSentence(const Sentence& s);

virtual string finishDocument();

static OutputFormat* newOutputFormat(const string& name);

static OutputFormat* newConlluOutputFormat(const string& options = string());

static OutputFormat* newEpeOutputFormat(const string& options = string());

static OutputFormat* newMatxinOutputFormat(const string& options = string());

static OutputFormat* newHorizontalOutputFormat(const string& options = string());

static OutputFormat* newPlaintextOutputFormat(const string& options = string());

static OutputFormat* newVerticalOutputFormat(const string& options = string());

static const string CONLLU_V1;

static const string CONLLU_V2;

static const string HORIZONTAL_PARAGRAPHS;

static const string PLAINTEXT_NORMALIZED_SPACES;

static const string VERTICAL_PARAGRAPHS;

};

class Model {

public:

static Model* load(const char* fname);

virtual InputFormat* newTokenizer(const string& options) const;

virtual bool tag(Sentence& s, const string& options, ProcessingError* error = nullptr)

const;

virtual bool parse(Sentence& s, const string& options, ProcessingError* error) const;

static const string DEFAULT;

static const string TOKENIZER_PRESEGMENTED;

};

class Pipeline {

public:

Pipeline(const Model* m, const string& input, const string& tagger, const string& parser,

const string& output);

void setModel(const Model* m);

void setInput(const string& input);

void setTagger(const string& tagger);

void setParser(const string& parser);

void setOutput(const string& output);

void setImmediate(bool immediate);

void setDocumentId(const string& document_id);

string process(const string& data, ProcessingError* error = nullptr) const;

static const string DEFAULT;

static const string NONE;

};

class Trainer {

public:

static string train(const string& method, const Sentences& train, const Sentences&

heldout,

const string& tokenizer, const string& tagger, const string& parser,

ProcessingError* error = nullptr);

static const string DEFAULT;

static const string NONE;

};

class Evaluator {

public:

Evaluator(const Model* m, const string& tokenizer, const string& tagger, const string&

parser);

void setModel(const Model* m);

void setTokenizer(const string& tokenizer);

void setTagger(const string& tagger);

void setParser(const string& parser);

string evaluate(const string& data, ProcessingError* error = nullptr) const;

static const string DEFAULT;

static const string NONE;

};

class Version {

public:

unsigned major;

unsigned minor;

unsigned patch;

string prerelease;

// Returns current version.

static version current();

};

6.16 C# Bindings

UDPipe library bindings is available in the Ufal.UDPipe namespace.

The bindings is a straightforward conversion of the C++ bindings API. The bindings requires native C++ library

libudpipe csharp (called udpipe csharp on Windows).

6.17 Java Bindings

UDPipe library bindings is available in the cz.cuni.mff.ufal.udpipe package.

The bindings is a straightforward conversion of the C++ bindings API. Vectors do not have native Java interface,

see cz.cuni.mff.ufal.udpipe.Words class for reference. Also, class members are accessible and modiﬁable

using using getField and setField wrappers.

The bindings require native C++ library libudpipe java (called udpipe java on Windows). If the library is

found in the current directory, it is used, otherwise standard library search process is used. The path to the

C++ library can also be speciﬁed using static udpipe java.setLibraryPath(String path) call (before the

ﬁrst call inside the C++ library, of course).

6.18 Perl Bindings

UDPipe library bindings is available in the Ufal::UDPipe package. The classes can be imported into the current

namespace using the :all export tag.

The bindings is a straightforward conversion of the C++ bindings API. Vectors do not have native Perl interface,

see Ufal::UDPipe::Words for reference. Static methods and enumerations are available only through the

module, not through object instance.

6.19 Python Bindings

UDPipe library bindings is available in the ufal.udpipe module.

The bindings is a straightforward conversion of the C++ bindings API. In Python 2, strings can be both unicode

and UTF-8 encoded str, and the library always produces unicode. In Python 3, strings must be only str.

7 Contact

Authors:

•Milan Straka,straka@ufal.mﬀ.cuni.cz

UDPipe website.

UDPipe LINDAT/CLARIN entry.

8 Acknowledgements

This work has been using language resources developed and/or stored and/or distributed by the LIN-

DAT/CLARIN project of the Ministry of Education of the Czech Republic (project LM2010013 ).

Acknowledgements for individual language models are listed in UDPipe User’s Manual.

8.1 Publications

•(Straka et al. 2017) Milan Straka and Jana Strakov´a. Tokenizing, POS Tagging, Lemmatizing and Parsing

UD 2.0 with UDPipe. In Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw

Text to Universal Dependencies, Vancouver, Canada, August 2017.

•(Straka et al. 2016) Straka Milan, Hajiˇc Jan, Strakov´a Jana. UDPipe: Trainable Pipeline for Processing

CoNLL-U Files Performing Tokenization, Morphological Analysis, POS Tagging and Parsing. In Pro-

ceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016),

Portoroˇz, Slovenia, May 2016.

8.2 Bibtex for Referencing

@InProceedings{udpipe:2017,

author = {Straka, Milan and Strakov\’{a}, Jana},

title = {Tokenizing, POS Tagging, Lemmatizing and Parsing UD 2.0 with UDPipe},

booktitle = {Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw

Text to Universal Dependencies},

month = {August},

year = {2017},

address = {Vancouver, Canada},

publisher = {Association for Computational Linguistics},

pages = {88--99},

url = {http://www.aclweb.org/anthology/K/K17/K17-3009.pdf}

}

8.3 Persistent Identiﬁer

If you prefer to reference UDPipe by a persistent identiﬁer (PID), you can use

http://hdl.handle.net/11234/1-1702.

MANUAL

Navigation menu

Versions of this User Manual:

Views

Navigation