DOCSLIB.ORG

Universal Dependency Parsing from Scratch

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Universal Dependency Parsing from Scratch Universal Dependency Parsing from Scratch Peng Qi,* Timothy Dozat,* Yuhao Zhang,* Christopher D. Manning Stanford University Stanford, CA 94305 fpengqi, tdozat, yuhaozhang, [email protected] Abstract 2018).1 Harnessing the power of neural sys- tems, this pipeline achieves competitive perfor- This paper describes Stanford’s system at mance in each of the inter-linked stages: tok- the CoNLL 2018 UD Shared Task. We enization, sentence and word segmentation, part- introduce a complete neural pipeline sys- of-speech (POS)/morphological features (UFeats) tem that takes raw text as input, and per- tagging, lemmatization, and finally, dependency forms all tasks required by the shared task, parsing. Our main contributions include: ranging from tokenization and sentence • New methods for combining symbolic statis- segmentation, to POS tagging and depen- tical knowledge with flexible, powerful neu- dency parsing. Our single system sub- ral systems to improve robustness; mission achieved very competitive perfor- • A biaffine classifier for joint POS/UFeats pre- mance on big treebanks. Moreover, after diction that improves prediction consistency; fixing an unfortunate bug, our corrected • A lemmatizer enhanced with an edit classifier nd st system would have placed the 2 , 1 , and that improves the robustness of a sequence- rd 3 on the official evaluation metrics LAS, to-sequence model on rare sequences; and MLAS, and BLEX, and would have out- • Extensions to our parser from (Dozat et al., performed all submission systems on low- 2017) to model linearization. resource treebank categories on all metrics Our system achieves competitive performance by a large margin. We further show the ef- on big treebanks. After fixing an unfortunate bug, fectiveness of different model components the corrected system would have placed the 2nd, through extensive ablation studies. 1st, and 3rd on the official evaluation metrics LAS, MLAS, and BLEX, and would have outperformed 1 Introduction all submission systems on low-resource treebank Dependency parsing is an important component categories on all metrics by a large margin. We in various natural langauge processing (NLP) sys- perform extensive ablation studies to demonstrate tems for semantic role labeling (Marcheggiani and the effectiveness of our novel methods, and high- 2 Titov, 2017), relation extraction (Zhang et al., light future directions to improve the system. 2018), and machine translation (Chen et al., 2017). However, most research has treated dependency 2 System Description parsing in isolation, and largely ignored upstream NLP components that prepare relevant data for the In this section, we present detailed descriptions parser, e.g., tokenizers and lemmatizers (Zeman for each component of our neural pipeline system, et al., 2017). In reality, however, these upstream namely the tokenizer, the POS/UFeats tagger, the systems are still far from perfect. lemmatizer, and finally the dependency parser. To this end, in our submission to the CoNLL 1 2018 UD Shared Task, we built a raw-text- We chose to develop a pipeline system mainly because it allows easier parallel development and faster model tuning in to-CoNLL-U pipeline system that performs all a shared task context. tasks required by the Shared Task (Zeman et al., 2To facilitate future research, we make our implementa- tion public at: https://github.com/stanfordnlp/ ∗These authors contributed roughly equally. stanfordnlp. (&'()) (+,-) ()./) 2.1 Tokenizer Final prediction !$ !$ !$ Second layer To prepare sentences in the form of a list of words !(&'()) !(+,-) !()./) prediction + 0,$ 0,$ 0,$ for downstream processing, the tokenizer compo- BiLSTM nent reads raw text and outputs sentences in the Second layer 1 Step t CoNLL-U format. This is achieved with two sub- First layer systems: one for joint tokenization and sentence prediction & (&'()) (+,-) ()./) !",$ !",$ !",$ σ + segmentation, and the other for splitting multi- gating + word tokens into syntactic words. BiLSTM First layer 1-D CNN 1 Tokenization and sentence segmentation. We Step t treat joint tokenization and sentence segmentation as a unit-level sequence tagging problem. For Input Input Unit t most languages, a unit of text is a single charac- Figure 1: Illustration of the tokenizer/sentence ter; however, in Vietnamese orthography, the most segmenter model. Components in blue represent 3 natural units of text are single syllables. We as- the gating mechanism between the two layers. sign one out of five tags to each of these units: end of token (EOT), end of sentence (EOS), multi-word token (MWT), multi-word end of sentence (MWS), and W1 contains the weights and biases for a lin- 4 and other (OTHER). We use bidirectional LSTMs ear classifier. For each unit, we concatenate its (BiLSTMs) as the base model to make unit-level trainable embedding with a four-dimensional bi- predictions. At each unit, the model predicts hier- nary feature vector as input, each dimension cor- archically: it first decides whether a given unit is responding to one of the following feature func- at the end of a token with a score s(tok), then clas- tions: (1) does the unit start with whitespace; (2) sifies token endings into finer-grained categories does it start with a capitalized letter; (3) is the unit with two independent binary classifiers: one for fully capitalized; and (4) is it purely numerical. sentence ending s(sent), and one for MWT s(MWT). To incorporate token-level information at the Since sentence boundaries and MWTs usually second layer, we use a gating mechanism to sup- require a larger context to determine (e.g., periods press representations at non-token boundaries be- following abbreviations or the ambiguous word fore propagating hidden states upward: “des” in French), we incorporate token-level infor- (tok) g1 = h1 σ(s1 ) (5) mation into a two-layer BiLSTM as follows (see −! − h = [h ; h ] = BiLSTM (g ); (6) also Figure1). The first layer BiLSTM operates 2 2 2 2 1 (tok) (sent) (MWT) directly on raw units, and makes an initial predic- [s2 ; s2 ; s2 ] = W2h2; (7) tion over the categories. To help capture local unit where is an element-wise product broadcast patterns more easily, we also combine the first- over all dimensions of h1 for each unit. This layer BiLSTM with 1-D convolutional networks, can be viewed as a simpler alternative to multi- by using a one hidden layer convolutional network resolution RNNs (Serban et al., 2017), where the (CNN) with ReLU nolinearity at its first layer, giv- first-layer BiLSTM operates at the unit level, and ing an effect a little like a residual connection (He the second layer operates at the token level. Unlike et al., 2016). The output of the CNN is simply multi-resolution RNNs, this formulation is end-to- added to the concatenated hidden states of the Bi- end differentiable, and can more easily leverage LSTM for downstream computation: efficient off-the-shelf RNN implementations. RNN −! − h1 = [h1; h1] = BiLSTM1(x); (1) To combine predictions from both layers of the CNN BiLSTM, we simply sum the scores to obtain h1 = CNN(x); (2) (X) (X) s(X) = s +s , where X 2 ftok, sent, MWTg. h = hRNN + hCNN; (3) 1 2 1 1 1 The final probability over the tags is then (tok) (sent) (MWT) [s ; s ; s ] = W1h1; (4) 1 1 1 pEOT = p+−− pEOS = p++−; (8) x where is the input character representations, pMWT = p+−+ pMWS = p+++; (9) 3 (tok) (sent) (MWT) In this case, we define a syllable as a consecutive run where p±±± = σ(±s )σ(±s )σ(±s ), of alphabetic characters, numbers, or individual symbols, to- gether with any leading white spaces before them. 4We will omit bias terms in affine transforms for clarity. and σ(·) is the logistic sigmoid function. pOTHER is To bring the symbolic and neural systems to- simply σ(−s(tok)). The model is trained to mini- gether, we train them separately and use the fol- mize the standard cross entropy loss. lowing protocol during evaluation: for each MWT, we first look it up in the dictionary, and return the Multi-word Token Expansion. The tokenizer/ expansion recorded there if one can be found. If sentence segmenter produces a collection of sen- this fails, we retry by lowercasing the incoming tences, each being a list of tokens, some of which token. If that fails again, we resort to the neural are labeled as multi-word tokens (MWTs). We system to predict the final expansion. This allows must expand these MWTs into the underlying syn- us to not only account for languages with flexi- tactic words they correspond to (e.g., “im” to “in ble MWTs patterns (Arabic and Hebrew), but also dem” in German), in order for downstream sys- leverage the training set statistics to cover both tems to process them properly. To achieve this, we languages with simpler MWT rules, and MWTs take a hybrid approach to combine symbolic statis- in the flexible languages seen in the training set tical knowledge with the power of neural systems. without fail. This results in a high-performance, The symbolic statistical side is a frequency lex- robust system for multi-word token expansion. icon. Many languages, like German, have only a handful of rules for expanding a few MWTs. We 2.2 POS/UFeats Tagger leverage this information by simply counting the number of times a MWT is expanded into differ- Our tagger follows closely that of (Dozat et al., ent sequences of words in the training set, and re- 2017), with a few extensions. As in that work, taining the most frequent expansion in a dictionary the core of the tagger is a highway BiLSTM (Sri- to use at test time. When building this dictionary, vastava et al., 2015) with inputs coming from the we lowercase all words in the expansions to im- concatenation of three sources: (1) a pretrained prove robustness. However, this approach would word embedding, from the word2vec embeddings fail for languages with rich clitics, a large set of provided with the task when available (Mikolov unique MWTs, and/or complex rules for MWT ex- et al., 2013), and from fastText embeddings oth- pansion, such as Arabic and Hebrew.
Load more

Recommended publications
  • Arxiv:1908.07448V1
    Evaluating Contextualized Embeddings on 54 Languages in POS Tagging, Lemmatization and Dependency Parsing Milan Straka and Jana Strakova´ and Jan Hajicˇ Charles University Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics {strakova,straka,hajic}@ufal.mff.cuni.cz Abstract Shared Task (Zeman et al., 2018). • We report our best results on UD 2.3. The We present an extensive evaluation of three addition of contextualized embeddings im- recently proposed methods for contextualized 25% embeddings on 89 corpora in 54 languages provements range from relative error re- of the Universal Dependencies 2.3 in three duction for English treebanks, through 20% tasks: POS tagging, lemmatization, and de- relative error reduction for high resource lan- pendency parsing. Employing the BERT, guages, to 10% relative error reduction for all Flair and ELMo as pretrained embedding in- UD 2.3 languages which have a training set. puts in a strong baseline of UDPipe 2.0, one of the best-performing systems of the 2 Related Work CoNLL 2018 Shared Task and an overall win- ner of the EPE 2018, we present a one-to- A new type of deep contextualized word repre- one comparison of the three contextualized sentation was introduced by Peters et al. (2018). word embedding methods, as well as a com- The proposed embeddings, called ELMo, were ob- parison with word2vec-like pretrained em- tained from internal states of deep bidirectional beddings and with end-to-end character-level word embeddings. We report state-of-the-art language model, pretrained on a large text corpus. results in all three tasks as compared to results Akbik et al.
    [Show full text]
  • Extended and Enhanced Polish Dependency Bank in Universal Dependencies Format
    Extended and Enhanced Polish Dependency Bank in Universal Dependencies Format Alina Wróblewska Institute of Computer Science Polish Academy of Sciences ul. Jana Kazimierza 5 01-248 Warsaw, Poland [email protected] Abstract even for languages with rich morphology and rel- atively free word order, such as Polish. The paper presents the largest Polish Depen- The supervised learning methods require gold- dency Bank in Universal Dependencies for- mat – PDBUD – with 22K trees and 352K standard training data, whose creation is a time- tokens. PDBUD builds on its previous ver- consuming and expensive process. Nevertheless, sion, i.e. the Polish UD treebank (PL-SZ), dependency treebanks have been created for many and contains all 8K PL-SZ trees. The PL- languages, in particular within the Universal De- SZ trees are checked and possibly corrected pendencies initiative (UD, Nivre et al., 2016). in the current edition of PDBUD. Further The UD leaders aim at developing a cross- 14K trees are automatically converted from linguistically consistent tree annotation schema a new version of Polish Dependency Bank. and at building a large multilingual collection of The PDBUD trees are expanded with the en- hanced edges encoding the shared dependents dependency treebanks annotated according to this and the shared governors of the coordinated schema. conjuncts and with the semantic roles of some Polish is also represented in the Universal dependents. The conducted evaluation exper- Dependencies collection. There are two Polish iments show that PDBUD is large enough treebanks in UD: the Polish UD treebank (PL- for training a high-quality graph-based depen- SZ) converted from Składnica zalezno˙ sciowa´ 1 and dency parser for Polish.
    [Show full text]
  • Universal Dependencies According to BERT: Both More Specific and More General
    Universal Dependencies According to BERT: Both More Specific and More General Tomasz Limisiewicz and David Marecekˇ and Rudolf Rosa Institute of Formal and Applied Linguistics, Faculty of Mathematics and Physics Charles University, Prague, Czech Republic {limisiewicz,rosa,marecek}@ufal.mff.cuni.cz Abstract former based systems particular heads tend to cap- ture specific dependency relation types (e.g. in one This work focuses on analyzing the form and extent of syntactic abstraction captured by head the attention at the predicate is usually focused BERT by extracting labeled dependency trees on the nominal subject). from self-attentions. We extend understanding of syntax in BERT by Previous work showed that individual BERT examining the ways in which it systematically di- heads tend to encode particular dependency verges from standard annotation (UD). We attempt relation types. We extend these findings by to bridge the gap between them in three ways: explicitly comparing BERT relations to Uni- • We modify the UD annotation of three lin- versal Dependencies (UD) annotations, show- ing that they often do not match one-to-one. guistic phenomena to better match the BERT We suggest a method for relation identification syntax (x3) and syntactic tree construction. Our approach • We introduce a head ensemble method, com- produces significantly more consistent depen- dency trees than previous work, showing that bining multiple heads which capture the same it better explains the syntactic abstractions in dependency relation label (x4) BERT. • We observe and analyze multipurpose heads, At the same time, it can be successfully ap- containing multiple syntactic functions (x7) plied with only a minimal amount of supervi- sion and generalizes well across languages.
    [Show full text]
  • Universal Dependencies for Japanese
    Universal Dependencies for Japanese Takaaki Tanaka∗, Yusuke Miyaoy, Masayuki Asahara}, Sumire Uematsuy, Hiroshi Kanayama♠, Shinsuke Mori|, Yuji Matsumotoz ∗NTT Communication Science Labolatories, yNational Institute of Informatics, }National Institute for Japanese Language and Linguistics, ♠IBM Research - Tokyo, |Kyoto University, zNara Institute of Science and Technology [email protected], [email protected], [email protected], [email protected], [email protected], [email protected], [email protected] Abstract We present an attempt to port the international syntactic annotation scheme, Universal Dependencies, to the Japanese language in this paper. Since the Japanese syntactic structure is usually annotated on the basis of unique chunk-based dependencies, we first introduce word-based dependencies by using a word unit called the Short Unit Word, which usually corresponds to an entry in the lexicon UniDic. Porting is done by mapping the part-of-speech tagset in UniDic to the universal part-of-speech tagset, and converting a constituent-based treebank to a typed dependency tree. The conversion is not straightforward, and we discuss the problems that arose in the conversion and the current solutions. A treebank consisting of 10,000 sentences was built by converting the existent resources and currently released to the public. Keywords: typed dependencies, Short Unit Word, multiword expression, UniDic 1. Introduction 2. Word unit The definition of a word unit is indispensable in UD an- notation, which is not a trivial question for Japanese, since The Universal Dependencies (UD) project has been de- a sentence is not segmented into words or morphemes by veloping cross-linguistically consistent treebank annota- white space in its orthography.
    [Show full text]
  • Universal Dependencies for Persian
    Universal Dependencies for Persian *Mojgan Seraji, **Filip Ginter, *Joakim Nivre *Uppsala University, Department of Linguistics and Philology, Sweden **University of Turku, Department of Information Technology, Finland *firstname.lastname@lingfil.uu.se **figint@utu.fi Abstract The Persian Universal Dependency Treebank (Persian UD) is a recent effort of treebanking Persian with Universal Dependencies (UD), an ongoing project that designs unified and cross-linguistically valid grammatical representations including part-of-speech tags, morphological features, and dependency relations. The Persian UD is the converted version of the Uppsala Persian Dependency Treebank (UPDT) to the universal dependencies framework and consists of nearly 6,000 sentences and 152,871 word tokens with an average sentence length of 25 words. In addition to the universal dependencies syntactic annotation guidelines, the two treebanks differ in tokenization. All words containing unsegmented clitics (pronominal and copula clitics) annotated with complex labels in the UPDT have been separated from the clitics and appear with distinct labels in the Persian UD. The treebank has its original syntactic annota- tion scheme based on Stanford Typed Dependencies. In this paper, we present the approaches taken in the development of the Persian UD. Keywords: Universal Dependencies, Persian, Treebank 1. Introduction this paper, we present how we adapt the Universal Depen- In the past decade, the development of numerous depen- dencies to Persian by converting the Uppsala Persian De- dency parsers for different languages has frequently been pendency Treebank (UPDT) (Seraji, 2015) to the Persian benefited by the use of syntactically annotated resources, Universal Dependencies (Persian UD). First, we briefly de- or treebanks (Bohmov¨ a´ et al., 2003; Haverinen et al., 2010; scribe the Universal Dependencies and then we present the Kromann, 2003; Foth et al., 2014; Seraji et al., 2015; morphosyntactic annotations used in the extended version Vincze et al., 2010).
    [Show full text]
  • Language Technology Meets Documentary Linguistics: What We Have to Tell Each Other
    Language technology meets documentary linguistics: What we have to tell each other Language technology meets documentary linguistics: What we have to tell each other Trond Trosterud Giellatekno, Centre for Saami Language Technology http://giellatekno.uit.no/ . February 15, 2018 . Language technology meets documentary linguistics: What we have to tell each other Contents Introduction Language technology for the documentary linguist Language technology for the language society Conclusion . Language technology meets documentary linguistics: What we have to tell each other Introduction Introduction I Giellatekno: started in 2001 (UiT). Research group for language technology on Saami and other northern languages Gramm. modelling, dictionaries, ICALL, corpus analysis, MT, ... I Trond Trosterud, Lene Antonsen, Ciprian Gerstenberger, Chiara Argese I Divvun: Started in 2005 (UiT < Min. of Local Government). Infrastructure, proofing tools, synthetic speech, terminology I Sjur Moshagen, Thomas Omma, Maja Kappfjell, Børre Gaup, Tomi Pieski, Elena Paulsen, Linda Wiechetek . Language technology meets documentary linguistics: What we have to tell each other Introduction The most important languages we work on . Language technology meets documentary linguistics: What we have to tell each other Language technology for the documentary linguist Language technology for documentary linguistics ... Language technology meets documentary linguistics: What we have to tell each other Language technology for the documentary linguist ... what’s in it for the language community I work for? I Let’s pretend there are two types of language communities: 1. Language communities without plans for revitalisation or use in domains other than oral use 2. Language communities with such plans . Language technology meets documentary linguistics: What we have to tell each other Language technology for the documentary linguist Language communities without such plans I Gather empirical material and do your linguistic analysis I (The triplet: Grammar, text collection and dictionary) I ..
    [Show full text]
  • Foreword to the Special Issue on Uralic Languages
    Northern European Journal of Language Technology, 2016, Vol. 4, Article 1, pp 1–9 DOI 10.3384/nejlt.2000-1533.1641 Foreword to the Special Issue on Uralic Languages Tommi A Pirinen Hamburger Zentrum für Sprachkorpora Universität Hamburg [email protected] Trond Trosterud HSL-fakultehta UiT Norgga árktalaš universitehta [email protected] Francis M. Tyers Veronika Vincze HSL-fakultehta MTA-SZTE UiT Norgga árktalaš universitehta Szegedi Tudomány Egyetem [email protected] [email protected] Eszter Simon Jack Rueter Research Institute for Linguistics Helsingin yliopisto Hungarian Academy of Sciences Nykykielten laitos [email protected] [email protected] March 7, 2017 Abstract In this introduction we have tried to present concisely the history of language tech- nology for Uralic languages up until today, and a bit of a desiderata from the point of view of why we organised this special issue. It is of course not possible to cover everything that has happened in a short introduction like this. We have attempted to cover the beginnings of the (Uralic) language-technology scene in 1980’s as far as it’s relevant to much of the current work, including the ones presented in this issue. We also go through the Uralic area by the main languages to survey on existing resources, to also form a systematic overview of what is missing. Finally we talk about some possible future directions on the pan-Uralic level of language technology management. Northern European Journal of Language Technology, 2016, Vol. 4, Article 1, pp 1–9 DOI 10.3384/nejlt.2000-1533.1641 Figure 1: A map of the Uralic language area show approximate distribution of languages spoken by area.
    [Show full text]
  • The Universal Dependencies Treebank of Spoken Slovenian
    The Universal Dependencies Treebank of Spoken Slovenian Kaja Dobrovoljc1, Joakim Nivre2 1Institute for Applied Slovene Studies Trojina, Ljubljana, Slovenia 1Department of Slovenian Studies, Faculty of Arts, University of Ljubljana 2Department of Linguistics and Philology, Uppsala University [email protected], joakim.nivre@lingfil.uu.se Abstract This paper presents the construction of an open-source dependency treebank of spoken Slovenian, the first syntactically annotated collection of spontaneous speech in Slovenian. The treebank has been manually annotated using the Universal Dependencies annotation scheme, a one-layer syntactic annotation scheme with a high degree of cross-modality, cross-framework and cross-language interoper- ability. In this original application of the scheme to spoken language transcripts, we address a wide spectrum of syntactic particularities in speech, either by extending the scope of application of existing universal labels or by proposing new speech-specific extensions. The initial analysis of the resulting treebank and its comparison with the written Slovenian UD treebank confirms significant syntactic differences between the two language modalities, with spoken data consisting of shorter and more elliptic sentences, less and simpler nominal phrases, and more relations marking disfluencies, interaction, deixis and modality. Keywords: dependency treebank, spontaneous speech, Universal Dependencies 1. Introduction actually out-performs state-of-the-art pipeline approaches (Rasooli and Tetreault, 2013; Honnibal and Johnson, 2014). It is nowadays a well-established fact that data-driven pars- Such heterogeneity of spoken language annotation schemes ing systems used in different speech-processing applica- inevitably leads to a restricted usage of existing spoken tions benefit from learning on annotated spoken data, rather language treebanks in linguistic research and parsing sys- than using models built on written language observation.
    [Show full text]
  • Proceedings of the First Workshop on Natural Language Processing for Indigenous Languages of the Americas
    Zurich Open Repository and Archive University of Zurich Main Library Strickhofstrasse 39 CH-8057 Zurich www.zora.uzh.ch Year: 2021 Proceedings of the First Workshop on Natural Language Processing for Indigenous Languages of the Americas Edited by: Mager, Manuel ; Oncevay, Arturo ; Rios, Annette ; Meza Ruiz, Ivan Vladimir ; Palmer, Alexis ; Neubig, Graham ; Kann, Katharina Posted at the Zurich Open Repository and Archive, University of Zurich ZORA URL: https://doi.org/10.5167/uzh-203436 Edited Scientific Work Published Version The following work is licensed under a Creative Commons: Attribution 4.0 International (CC BY 4.0) License. Originally published at: Proceedings of the First Workshop on Natural Language Processing for Indigenous Languages of the Americas. Edited by: Mager, Manuel; Oncevay, Arturo; Rios, Annette; Meza Ruiz, Ivan Vladimir; Palmer, Alexis; Neubig, Graham; Kann, Katharina (2021). Online: Association for Computational Linguistics. NAACL-HLT 2021 Natural Language Processing for Indigenous Languages of the Americas (AmericasNLP) Proceedings of the First Workshop June 11, 2021 ©2021 The Association for Computational Linguistics These workshop proceedings are licensed under a Creative Commons Attribution 4.0 International License. Order copies of this and other ACL proceedings from: Association for Computational Linguistics (ACL) 209 N. Eighth Street Stroudsburg, PA 18360 USA Tel: +1-570-476-8006 Fax: +1-570-476-0860 [email protected] ISBN 978-1-954085-44-2 ii Preface This area is in all probability unmatched, anywhere in the world, in its linguistic multiplicity and diversity. A couple of thousand languages and dialects, at present divided into 17 large families and 38 small ones, with several hundred unclassified single languages, are on record.
    [Show full text]
  • Conll-2017 Shared Task
    TurkuNLP: Delexicalized Pre-training of Word Embeddings for Dependency Parsing Jenna Kanerva1,2, Juhani Luotolahti1,2, and Filip Ginter1 1Turku NLP Group 2University of Turku Graduate School (UTUGS) University of Turku, Finland [email protected], [email protected], [email protected] Abstract word and sentence segmentation as well as mor- phological tags for the test sets, which they could We present the TurkuNLP entry in the choose to use as an alternative to developing own CoNLL 2017 Shared Task on Multilingual segmentation and tagging. These baseline seg- Parsing from Raw Text to Universal De- mentations and morphological analyses were pro- pendencies. The system is based on the vided by UDPipe v1.1 (Straka et al., 2016). UDPipe parser with our focus being in ex- In addition to the manually annotated treebanks, ploring various techniques to pre-train the the shared task organizers also distributed a large word embeddings used by the parser in or- collection of web-crawled text for all but one of der to improve its performance especially the languages in the shared task, totaling over 90 on languages with small training sets. The billion tokens of fully dependency parsed data. system ranked 11th among the 33 partici- Once again, these analyses were produced by the pants overall, being 8th on the small tree- UDPipe system. This automatically processed banks, 10th on the large treebanks, 12th on large dataset was intended by the organizers to the parallel test sets, and 26th on the sur- complement the manually annotated data and, for prise languages. instance, support the induction of word embed- dings.
    [Show full text]
  • Conferenceabstracts
    TENTH INTERNATIONAL CONFERENCE ON LANGUAGE RESOURCES AND EVALUATION Held under the Honorary Patronage of His Excellency Mr. Borut Pahor, President of the Republic of Slovenia MAY 23 – 28, 2016 GRAND HOTEL BERNARDIN CONFERENCE CENTRE Portorož , SLOVENIA CONFERENCE ABSTRACTS Editors: Nicoletta Calzolari (Conference Chair), Khalid Choukri, Thierry Declerck, Marko Grobelnik , Bente Maegaard, Joseph Mariani, Jan Odijk, Stelios Piperidis. Assistant Editors: Sara Goggi, Hélène Mazo The LREC 2016 Proceedings are licensed under a Creative Commons Attribution- NonCommercial 4.0 International License LREC 2016, TENTH INTERNATIONAL CONFERENCE ON LANGUAGE RESOURCES AND EVALUATION Title: LREC 2016 Conference Abstracts Distributed by: ELRA – European Language Resources Association 9, rue des Cordelières 75013 Paris France Tel.: +33 1 43 13 33 33 Fax: +33 1 43 13 33 30 www.elra.info and www.elda.org Email: [email protected] and [email protected] ISBN 978-2-9517408-9-1 EAN 9782951740891 ii Introduction of the Conference Chair and ELRA President Nicoletta Calzolari Welcome to the 10 th edition of LREC in Portorož, back on the Mediterranean Sea! I wish to express to his Excellency Mr. Borut Pahor, the President of the Republic of Slovenia, the gratitude of the Program Committee, of all LREC participants and my personal for his Distinguished Patronage of LREC 2016. Some figures: previous records broken again! It is only the 10 th LREC (18 years after the first), but it has already become one of the most successful and popular conferences of the field. We continue the tradition of breaking previous records. We received 1250 submissions, 23 more than in 2014. We received 43 workshop and 6 tutorial proposals.
    [Show full text]
  • A Multilingual Collection of Conll-U-Compatible Morphological Lexicons
    A multilingual collection of CoNLL-U-compatible morphological lexicons Benoˆıt Sagot Inria 2 rue Simone Iff, CS 42112, 75589 Paris Cedex 12, France [email protected] Abstract We introduce UDLexicons, a multilingual collection of morphological lexicons that follow the guidelines and format of the Universal Dependencies initiative. We describe the three approaches we use to create 53 morphological lexicons covering 38 languages, based on existing resources. These lexicons, which are freely available, have already proven useful for improving part-of-speech tagging accuracy in state-of-the-art architectures. Keywords: Morphological Lexicons, Universal Dependencies, Freely Available Language Resources 1. Introduction guages following a universal set of guidelines. The obvious Morphological information belongs to the most fundamen- choice would be to make use of the UD guidelines them- tal types of linguistic knowledge. It is often either encoded selves. into morphological analysers or gathered in the form of We have therefore developed a multilingual collection of morphological lexicons. Such lexicons, which constitute morphological lexicons that follow the UD guidelines re- the focus of this paper, are collections of lexical entries that garding part-of-speech and morphological features. We typically associate a wordform with a part-of-speech (or used three main sources of lexical information: morphosyntactic category), morphological features (such • In the context of the CoNLL 2017 UD morphologi- as gender, tense, etc.) and a lemma. Beyond direct lexicon cal and syntactic analysis shared task (Zeman et al., lookup, used in virtually all types of natural language pro- 2017) based on UD treebank data, we used lexical in- cessing applications and computational linguistic studies, formation available in the Apertium2 and Giellatekno3 morphological lexicons have been shown to significantly projects.
    [Show full text]