THE USER'S GUIDE FOR THE AUTHORS
Contents
1. 323 330 Springer Verlag 2009 TRANDABAT D M CRISTEA D and TUFIS D eds Proceedings of the Workshop Linguistic Resources and Tools for Processing Romanian Language in Romanian Lucr rile atelierului 16 Violeta SERETAN and Eric WEHRLI Resurse lingvistice i instrumente pentru prelucrarea limbii rom ne University Alexandru loan Cuza Publishing House 2008 22 WEHRLI E L analyse syntaxique des langues naturelles Probl mes et m thodes Masson 1997 23 WEHRLI E Fips a deep linguistic multilingual parser In ACL 2007 Workshop on Deep Linguistic Processing Prague pp 120 127 2007 24 WEHRLI E and NERIMA L L analyseur syntaxique Fips In Proceedings of the IWPT 2009 ATALA Workshop What French parsing systems Paris France 20092. Only 2 of the tokens in a corpus are unknown 17 The extraction precision is the percentage of collocations in the results investigated In contrast the grammatical precision i e the percentage of grammatical pairs in the results investigated is much higher We did not quantify it as the focus of our work was to collect collocations 14 Violeta SERETAN and Eric WEHRLI Concordance Collocations 100 uniune european stat unit na iune unit stat unit alt parte atrage aten ie punct de vedere scurt timp imigrant ilegal mia de metru fi vorb miliard de dolar milion de dolar campionat mondial datorie istoric alegere preziden ial da n judecat cas alb cup mondial persoan cu handicap s pt m n trecut campion mondial test nuclear Display Collocations Source D corpus po journaux 162 odc in planul de creare a unei zone de securitate eful Statului Major israelian generalul Dan Halutz a declarat c bateriile de artilerie si fortele aeriene israeliene vor primi ordin s deschid focul asupra oric rui combatant al Hezbollah care intr in aceast zon Postul de radio militar a ad ugat c buldozerele israeliene au intrat in sudul Libanului pentru a demola pozitile abandonate de Hezbollah in urma ofensivei israeliene Armata israelian a dezmintit ins c fortele terestre ar fi p truns in sudul Libanului Nu exist forte terestre israeliene in Liban
3. dat fetelor flori They cl Dat have given girls Dat flowers I gave the girls flowers In Romanian unlike in French and Italian no material can intervene between the auxiliary and the participial verb except for the monosyllabic adverbs mentioned above Our implementation is consistent with Monachesi s 2000 theoretical account which postulates a compound structure for the Romanian auxiliary verb system rather than a flat structure which is suitable for other Romance languages Another difference between Romanian and other Romance languages pertains to the wh elements Soare 2009 Similarly to Slavic languages the Romanian language permits multiple wh fronting which is not possible for instance in French or Italian cf Example 4a c The wh phrases are rigidly ordered obey the Superiority Condition and no material can intervene between them cf Example 4d Broadly speaking the Nominative precedes the Dative and the Dative precedes the Accusative 8 Violeta SERETAN and Eric WEHRLI 4 a Cine cui ce a spus Whowho Dat what has said Who said what to whom b Qui qui quoi a dit c Chi a chi cosa ha detto d Ce cine cui a spus To sum up the Romanian grammar is fully described in our project in a formalism adapted to Fips its implementation currently in progress accounts for whole gamut of possible constituent attachments and while building upon the existing Romance parsing module for d
4. inflected form and the right hand side lexeme related information pertaining to the corresponding lexeme There are multiple lexemes for the verb a veni with different subcategorization as shown in the upper list The features displayed below apply to the currently selected lexeme Among other things the interface displays the collocations in which a lexeme participates Information about the coverage of the Romanian lexicon when tested on a collection of texts will be provided in the Section 5 that is dedicated to the experimental results Monolingual 3 venit Romanian veni V 1911000813 NP_ venit N veni de la din dinspre V 911000814 NP _ PP Add New Paradigm veni la spre V 911000815 NP _ PP Paradigm Category y Base form veni inflection 4s T participe pass en t v Word Lexeme Number v Singular Gender v Masculine Type ordinary v Plural LI Feminine Features tergative noPassive v Neuter Features 2 mouvement Modify Person v 4 M2 Cs v Ms Ls Penalty o Frequency 4 Used in propunere vine v Show collocations Sub mp v Number of 1 i Category Arguments J Tense jparticipePasse Modify Arguments Argument v Mood In Number Flags p Modify Grammatical subject L J Function P
5. language due the high number and variety of lexical resources and morphological tools that are available Dictionaries thesauri e g WordNet FrameNet word aligners annotated parallel corpora POS taggers and automated processing architectures are some examples of resources that have been produced over the past years and that are being successfully used for different purposes computational lexicography question answering word sense disambiguation anaphora resolution and textual entailment among many others A brief review of the resources and tools that exist for Romanian can be found in Cristea 2009 A more detailed report is given in Cristea amp For scu 2006 In addition the proceedings the workshop series ConsILR Consortium for the Romanian Language Resources amp Tools provides a complete picture on the advances in Romanian language processing In contrast with this situation little or no resources exist for Romanian insofar as the syntactic level is concerned Despite the recent efforts made by C l cean amp Nivre 2009 to adapt a stochastic dependency parser to Romanian there is currently no large scale syntactic parser for Romanian able to parse unrestricted text To a certain extent shallow parsing is already available For example Tradab et all 2008 Extending a Multilingual Symbolic Parser to Romanian 3 not only NP and VP chunking i e identification of word sequences constituting noun p
6. of the partial parses is 5 3 tokens The percentage of fully parsed sentences is much lower that that obtained by the parser Fips for other languages for comparison the English and French versions the most developed achieve around 80 complete analyses on journalistic data 12 Note that contrary to many shallow parsing systems the POS tags is a product not a prerequisite of parsing P http www bbc co uk romanian http www gandul info http www adevarulonline ro Accessed 2006 14 http www mesagerul ro Accessed 2006 2008 5 In the Romanian alphabet there are a few symbols that use diacritics a 4 i s t The online Romanian mass media is not consistent in using these diacritics Only a minority of online newspapers currently uses the correct alphabet Extending a Multilingual Symbolic Parser to Romanian 13 As for the lexical coverage this can be judged as rather satisfactory as only 6 596 of the input tokens were not found in the Romanian lexicon The token type ratio is 4 02 Moreover about half of the unknown words are a priori proper nouns as they begin with a capital letter 39 296 of them Fips adopts a guessing strategy for the unknown words trying to assign them a category compatible with the analysis pursued and to attach them in the output tree Thus the impact of the unknown words on the success rate of the parser is rather reduced The key factor for improving FipsRomanian is continuing the implem
7. the collocations continuum McKeown amp Radev 2000 On the other extreme one finds collocations that are more similar to free combinations like mare importan high importance lit big importance Since collocations allow the insertion of additional material between the component items they cannot be stored in the lexicon in the same way as compounds They are stored as binary associations of lexemes where each item can be either a single word or an existing multi word 10 There are about 35 inflected forms for a Romanian verb l For instance http ro wikipedia org and http www archeus ro Accessed October 2009 10 Violeta SERETAN and Eric WEHRLI entry in particular a compound or another collocation Thanks to the availability of FipsRomanian we recently could run our first experiments of collocation extraction from corpora based on syntactically informed methods Seretan 2008 The results obtained allowed us to the put the basis of the Romanian collocation lexicon which currently contains a few hundred collocations This work will be detailed in a later section Figure 6 displays the interface to the lexical database of the parser It shows the two entries that exist in the Romanian lexicon for the word venit corresponding to the two possible readings as past participle of the verb a veni to come and as a common noun venit income The left hand side present word related information applying to the
8. A avut loc o mic incursiune in timpul noptii pentru a distruge c teva avanposturi ale Hezbollah situate imediat de partea cealalt a granitei Asta a fost tot a subliniat un purt tor de cuv nt al armatei Rachetele libanezilor lovesc pentru prima dat Afula i Nazaret Replica Hezbollah nu s a l sat asteptat gruparea iit intensific nd atacurile cu rachete asupra nordului Israelului Posturile de televiziune libaneze au anuntat c un avion israelian care se afla in spa iul aerian al Libanului a fost dobor t iar cei doi piloti ai aparatului F16 ar fi fost uci i Televiziunile libaneze au prezentat imagini cu resturi ale aparatului fumeg nd Informatia a fost dezmintit de o surs din serviciile de securitate israeliene Atacurile cu rachete sunt ins o certitudine leri Rezistenta Islamic aripa armat a Hezbollah a revendicat un tir de rachete Katiu a lansate asupra ora elor Acre i Nahariya Potrivit armatei israeliene trei persoane au fost r nite in satul Talel de explozia a circa 15 rachete lansate din Liban Pentru prima dat rachetele lansate de Hezbollah au lovit localit ile Afula i Nazaret Proiectilele vizau M Open source fie Rank 1 Language ot aof s2 med ON C Group by syntactic type score O frequency O alphabetical Details Type Merb Object Score 339 90 Status New collocation Validate Cancel Fig 8 Interface of the collocation c
9. VP denoted by b if the case of a is nominative and the conditions of agreement in number and person between a and b are satisfied D N un student a IsType indefinite a student b IsType commonNoun a AgreeWith b gender number Fig 4 Sample right attachment rule Figure 4 shows the example of a right attachment a type of attachment which is relatively more frequent in Romanian This rule enables the attachment of a NP as a right sub constituent of a DP when the determiner is indefinite the noun is common and there is agreement in gender and number In the current Romanian specifications about 100 grammar rules are described a quarter of which concern left attachments and the others right attachments This number is comparable to that for other languages supported by the parser and the coverage of these rules is judged at least in 7 http www technolangue net article198 html Accessed May 2010 http atoll inria fr passage Accessed May 2010 Extending a Multilingual Symbolic Parser to Romanian 7 principle thorough However the implementation of these rules is far from complete Currently about half of the attachment rules are fully implemented and tested and the status of the Romanian parser is that of work in progress Apart from the attachment rules the implementation of the grammar also requires accounting for the syntactic transformations possible in Romanian and correspondingly for the creation of co
10. Violeta Seretan and Eric Wehrli Extending a multilingual symbolic parser to Romanian In Dan Tufis and Corina For scu editors Multilinguality and Interoperability in Language Processing with Emphasis on Romanian Romanian Academy Publishing House Bucharest Romania 2010 EXTENDING A MULTILINGUAL SYMBOLIC PARSER TO ROMANIAN Violeta SERETAN Eric WEHRLI LATL Language Technology Laboratory University of Geneva Switzerland E mail violeta seretan eric wehrli unige ch Abstract A syntactic parser a system that analyses the structure of natural language sentences is a fundamental tool for any language providing information that is essential in virtually any other language application With a single exception C l cean amp Nivre 2009 such a tool was missing from the otherwise vast repertory of language tools available for Romanian In this paper we report on ongoing work aimed to develop a symbolic syntactic parser able to fully analyse unrestricted Romanian text in contrast the existing parser provides an analysis in terms of dependency relations is data driven and was only trained on simple sentences Our parser is based on the Fips multilingual parsing architecture Wehrli 2007 We present the preliminary tasks that enabled the implementation of the Romanian version i e lexicon compilation and grammar specification We describe the current status of the parser and present experimental results both on parsing a collec
11. and the most common verbs Extending a Multilingual Symbolic Parser to Romanian 9 Verbs in particular require detailed specific information about subcategorization selectional features grammatical function for arguments thematic functions and other information for instance on aspect which can only be specified manually Nonetheless the morphological generation process is still very useful since verb paradigms contain very numerous forms and it would be too time consuming to add them manually For the time being the FipsRomanian lexicon contains slightly more than 3600 verbal lexemes a low coverage relative to that of nouns above 38000 lexemes and adjectives above 14000 lexemes As far as non content words are concerned part of them have been added automatically e g interjections the others semi automatically or manually The human intervention was necessary in order to specify for each entry detailed information like for instance subtype for adverbs negation location time etc which is not available in DEX In addition to the entries compiled from DEX the FipsRomanian lexicon also contains proper nouns mostly for places cities countries rivers mountains and persons the most usual first names and surnames separately The majority of them has been collected from different repositories available on the Internet and has been automatically entered in the lexicon Table 1 displays the current composition
12. e appropriate lexical entries This procedure greatly simplifies the construction of the lexicon INFL it cat V inflClass 18 base 1 tense pastPart gender masc neut pers 1 2 3 num sing Fig 5 Sample inflection rule used for morphological generation Figure 5 above gives the example of a simple inflection rule used for generating the past participle form for verbs of a given inflection class in this case 18 by appending the suffix it to the root the root is obtained from the present infinitive This rule specifies the morphological features for this form e g masculine or neuter genders and singular number Other rules will be used for other gender number combinations leading to the production of different forms The present rule will generate for instance the past participle venit of the verb a veni to come once it is stated that a veni is a verb belonging to the inflection class 18 The process of compilation of the FipsRomanian lexicon went through several stages A list of base word forms was first obtained from the DEX dictionary DEX 1998 An inflection class number was automatically assigned to most nouns and adjectives based on their suffix This in conjunction with the inflection rules defined for each class allowed for the automatic generation of inflected forms for these lexemes A part of the remaining nouns and adjectives were manually entered into the lexicon as were pronouns determiners
13. ealing with grammatical transformations it pays particular attention to the proper treatment of language specific phenomena 42 Lexicon compilation The Romanian language has a rich morphology since it inherited in part the Latin declension system There nominal case system comprises the nominative accusative and the dative genitive syncretic cases as well as the vocative Inflected forms are obtained by suffixation A particularity of Romanian among all other languages of the same family is that the definite declension is also obtained through suffixation There are three genders masculine feminine and neuter the neuter gender behaves like masculine for singular and feminine for plural Adjectives also exhibit declension in case number gender and definiteness Verbs are highly inflected for person number tense and mood The Fips parser conveniently includes a morphological generation module which produces all the inflected forms of a lexeme according to the appropriate declension paradigm Generation rules are specified in a Fips specific format automatically processed by the morphological generator A rule accounts for the production of a specific form which will be inserted in the lexicon together with the applicable morphological features Thus given a base word form its lexical category and a numerical code representing the inflection class the morphological generator reads all the rules applying to that class and prepares th
14. enalty 5 Frequency 9 Thematical scent Rh l Function S Phonetics Add Features p Modify Fig 6 Lexicon interface screen capture Extending a Multilingual Symbolic Parser to Romanian 11 5 EXPERIMENTAL RESULTS As mentioned in the previous section the preparation of the resources needed for extending Fips to Romanian namely the lexicon and the grammar has already reached completion however the implementation of the grammar is still at an incipient stage We started with simple structures for phrases short declarative sentences and then moved to more complex structures subordination and coordination of clauses interrogation etc Example 5 illustrates the kind of structures that FipsRomanian is able to analyze correctly The parse trees are shown in the typical parenthetical notation 5 a un copac cu flori DP un NP copac PP cu NP flori b Copiii scriu scrisori p rin ilor TP DP Copiii scriu VP NP scrisori DP p rin ilor c Ghici cine vine m ine la cin TP DP Ghici VP TP NP cine vine VP AdvP maine PP la NP cin d Toti angaja ii societ ii particip la aceast sedinf d de informare i socializare TP DP Toti DP angaja ii DP societ ii particip VP PP la DP aceast NP sedint PP de NP ConjP NP informare si NP socializare Figure 7 shows in graphical form the output obtained for a se
15. entation of the grammatical component For the time being given the incipient stage of implementation we cannot provide more meaningful evaluation results for the parser This will be a concern of our future work in which we plan to test the precision and recall of FipsRomanian on the dependency treebanks mentioned in Section 2 5 3 Application In order to test to what extent the often fragmented parsing analyses produced by FipsRomanian are useful from a practical point of view we attempted to use them in specific application as suggested in Section 4 the application considered was a lexical acquisition application whose goal is more specifically to detect collocations from parsed corpora Until now the collocation extraction application Seretan 2008 has systematically been run on corpora in the other languages supported by Fips and the results have been used as raw material for manual inclusion in the lexical database Furthermore bilingual correspondences for the collocations indentified are being added into bilingual lexicons and used in an in house rule based machine translation system Collocations are also being used in turn in further parsing processes in order to inform attachment disambiguation decisions But the most directly perceived application of collocation is the language generation since collocations constitute a major means to ensure language fluency Collocations not always translate literally compare faire atte
16. for simple lexemes along with complex lexemes i e compound words collocations and idioms which are enriched with morphosyntactic and semantic information whose role is to guide the parser Extending Fips to a new language therefore means performing two main preliminary tasks grammar specification and lexicon compilation In this section we detail this process for Romanian 4 1 Grammar specification A description of the Romanian syntax for the purposes of Fips consistent with a traditional grammar Popescu 2004 and a generative grammar Dobrovie Sorin 1994 for Romanian and adapted to the theoretical model used by Fips has been made available by Soare 2005 Based on this description grammar rules for Romanian have been specified in a pseudo formalism specific to Fips This formalism has the advantage that it is easy to adopt by linguists and at the same time it is close to the program code of the parser Most rules in the grammar specification refer to the conditions under which two adjacent constituents can be attached by the Merge operation in order to yield a larger constituent An example of such a rule is provided in Figure 3 D T subjectAttachment studentul scrie a HasCase nominative student the writes a AgreeWith b number person the student writes b IsTensed Fig 3 Sample left attachment rule This rule describes the subject attachment in which a DP denoted by a is attached as a left sub constituent of a tensed
17. ft to right bottom up fashion by applying at each step one of the main operations enumerated above The application of these operations is constrained by both language independent grammar rules implemented in the generic module and language specific rules defined for each language supported by the parser Thus the application of the Merge operation in which a left or right sub constituent is attached to the current structure is constrained by language specific licensing rules like the agreement rules Moreover the attachments can only be made to a node that is active i e a node that accepts sub constituents The alternatives are pursued in parallel and several pruning heuristics are employed for limiting the search space Given a sentence the parser provides both the phrase structure representation same as the constituent structure c structure in LFG and the representation of grammatical functions for constituents in the form of a predicate argument table similar to the f structure in LFG The parser also provides an interpretation for clitics wh elements and relative pronouns a process which can be assimilated to intra sentential pronoun resolution Since it adopts the theory of movement the parser considers that words may leave their original deep or canonical position and move to their final surface position due to grammatical transformations for instance interrogation and relativization The parser therefore builds co
18. her languages are also supported to a certain extent Fips is able to handle a wide range of constructions in these languages such as the ones illustrated below for English e relativization the record which he broke e passivization the record was broken e interrogation which record did he break e cleft constructions it is the world record that was broken e coordination the record set by X and later broken by Y e apposition the record previously held by X was broken One of the key features of Fips is its robustness It can process large text collections at a reasonable speed of approximately 150 tokens s Its precision is currently being measured in the framework of two parsing evaluation campaigns for French namely EASy Evaluation des Analyseurs 6 Violeta SERETAN and Eric WEHRLI SYntaxiques and PASSAGE Produire des Annotations Syntaxiques Grande chelle For the time being no definitive evaluation results have been made available 4 EXTENDING FIPS TO ROMANIAN TWO MAIN TASKS The language specific part of the Fips parser consists on the one hand of the grammar rules of a given language and on the other hand of the lexicon of that language The grammar rules specify under which condition the parser s main operations Project Merge and Move cf Section 3 may apply in order to enable the creation of a parse tree given an input sentence The lexicon of the language contains entries
19. hrases and verb phrases cf Cristea amp For scu 2006 but also identification of syntactic relations by using regular expressions applied to POS tagged text The latter is work in progress aimed at extending a lexicographic corpus tool the Sketch Engine Kilgarriff et al 2004 to Romanian The work of C l cean amp Nivre 2009 can be considered as the most significant to date in the area of syntactic parsing for Romanian It concerns the application to Romanian of a language independent dependency parser the MaltParser Nivre et al 2007 MaltParser is data driven i e it uses syntactically annotated corpora treebanks to induce a parsing model C l cean amp Nivre 2009 made use of a small Romanian dependency treebank developed in the framework of the BALRIC LING project the Balkan Regional Information Centers for HLT This treebank consist of about 36000 tokens corresponding to slightly more than 4000 sentences with an average length of 8 94 tokens Only simple sentences were included and the treebank is reported as rather homogenous The authors reported a high level of precision for the Romanian parser similar to that achieved by MaltParser for English Italian and Catalan 88 6 for labelled attachments and 92 096 for unlabelled attachments The evaluation was performed on a held out portion of 10 of the same treebank The parser has not been tested on more complex sentences thus in spite of the promising result
20. indexation chains cf Section 3 These processes are already dealt with in FipsRomance the grammatical component of Fips that models the Romance family of languages to which Romanian belongs Some refinements are however needed because in certain respects the Romanian language exhibit distinctive properties A case in point is the clitic system which is richer in Romanian than in other Romance languages such as French or Italian Monachesi 2000 The Romanian clitic complex involves pronouns negation auxiliaries as well as a restricted subclass of monosyllabic adverbs mai again cam little prea too i also tot still The order of these elements is rigid negation is the leftmost element preceding the clitics and the auxiliary dative clitics precede accusative clitics and monosyllabic adverbs fill a position between the auxiliary and the participial verb cf Example 1a 1 a nu i l am mai prezentat pe Dan not you cl Dat he cl Acc haveagain introduce PE Dan didn t introduce Dan to you anymore b nu fi am mai prezentat pe Dan Romanian is similar to Spanish but differs from French and Italian in that the Accusative DP marked by the preposition pe must be clitic doubled cf Example 1b and 2 2 a L am v zut pe Ion He cl Acc have seen PE Ton I have seen Ion b Am v zut pe Ion Clitic doubling is however optional for full Dative DPs as shown in Example 3 3 Le am
21. indexation chains which link Extending a Multilingual Symbolic Parser to Romanian 5 extraposed moved elements to their empty original position where empty constituents are created to mark the trace left by the movement TP difficult Fig 2 Sample output of FipsEnglish showing co indexation chains Figure 2 above displays graphically the constituent structure returned by Fips for the English sentence The problem which we try to solve is very difficult The extraposed noun problem leaved its deep canonical position of direct object for the predicate solve and took the surface position of subject due to a relativization transformation The parser links the trace of the noun DP to the relative pronoun which and further to the noun in the subject position NP Therefore in the predicate argument table associated with solve problem will be found on the direct object position Similarly the parser builds a co indexation chains between the empty subject of solve the lower DP and the overt subject of try the higher DPy The pronoun we will thus be found on the subject position in the predicate argument table of solve The co indexation denoted by k therefore allows us to infer that the subject of the verb solve is in fact the subject of the c commanding higher verb try Fips is currently available for English French Spanish Italian German and Greek and has a high grammatical coverage for these languages ot
22. llaborators An important role was also played by the availability of suitable lexical resources such as comprehensive electronic dictionaries The two main preliminary tasks required for building the Romanian version i e lexicon compilation and grammar specification recently approached completion Therefore the implementation of the grammar itself could be started and the Romanian parser took shape Although its development is far from completed the parser is operational and could already be used in a specific application In the remaining sections we describe the current state of FipsRomanian and present its first results We first review in Section 2 the previous work aimed at developing syntactic parsers for Romanian In Section 3 we introduce the Fips multilingual parser specify the kind of information it provides and briefly present its parsing algorithm In Section 4 we focus on the development of the Romanian version and discuss the preparation of the resources needed as well as the language specific issues encountered In Section 5 we present some preliminary results obtained by parsing a one million word collection of newspaper articles We also show how we used these results in an application concerned with the identification of a particular subtype of multi word expressions The last section provides concluding remarks and points out directions for future work 2 RELATED WORK Romanian can be considered as a rather privileged
23. nal Grammar Bresnan 2001 Each syntactic constituent is represented as a simplified X bar structure of the form xp L X R with no intermediate levels X stands for one of the following lexical or functional categories N noun A adjective D determiner V verb Adv adverb P preposition Conj conjunction Interj interjection C complementizer T tensed VP head of a sentence and F functional phrase representing predicative objects L and R stand for possibly empty lists of left and right sub constituents respectively The lexical level contains detailed morphosyntactic and semantic information available from the manually built lexicons namely selectional properties subcategorization information and syntactico semantic features likely to influence the syntactic analysis Thus the parser relies on a strong lexicalist grammar framework Written in Component Pascal Fips adopts an object oriented implementation design that enables the coupling of language specific processing modules to a generic module The generic module defines the basic data types and is responsible of the parser s main operations Project assignment of constituent structures to lexical entries Merge combination of adjacent constituents into larger structures and Move creation of chains by linking surface positions of moved constituents to their corresponding canonical positions The parsing algorithm proceeds in a le
24. neva over the past years since 2004 to build a lexicon and a syntactic parser for Romanian This work is part of a long term parsing project that led since its inception in the 1990s 2 Violeta SERETAN and Eric WEHRLI to the development of a large scale robust syntactic parser called Fips which 1s now available in six major indo European languages French English German Spanish Italian and Greek Wehrli 1997 Wehrli 2007 Wehrli amp Nerima 2009 A number of other languages have also been considered and the corresponding parsers reached different stages of development The work on Romanian has mainly been done in the framework of a recent project on multilingual extension of the Fips parser to four new languages including another Romance language Romansh another major indo European language Russian and a language much more distant than all the other languages currently supported by Fips namely Japanese The aim of the project was also to challenge the extent to which Fips succeeded to evolve over the years from a parser supporting a single language French Fips stands for French Interactive Parsing System to a multilingual parsing architecture separating the core language independent implementation from the language specific extensions The extension to Romanian was the one that advanced the most rapidly thanks to the sustained efforts of lexicon compilation and grammar description made previously by different co
25. ng of this section presents the experiment performed to this end 5 1 Data The FipsRomanian parser was applied on collection of newspaper articles totalling 1 2 million tokens corresponding to 1 05 million words which have been gathered from various sources on the Internet These included the BBC Romanian website G ndul Adev rul 3 articles from 2006 and as the main source the newspaper Mesagerul articles from the period 2006 2008 The collection of articles has been done manually The main criterion for choosing the source sites was the proper encoding of diacritics A few more statistics on these texts could be derived after parsing was performed the average sentence length is 26 9 tokens including punctuation marks and there are in average 113 2 sentences per file the whole collection contains 393 files The corpus was not pre processed in any way therefore there are spans of texts corresponding to the article header e g title date which might affect the parser since they do not constitute full sentences 5 2 Results FipsRomanian succeeded in processing the whole text collection at a speed of 429 tokens s The processing took slightly less than one hour and was done on a standard PC configuration A total number of 44483 sentences have been indentified and analysed The parser could build a complete parse tree in 16 2 of the cases For the others it returned multiple disconnected parse trees The average length
26. ngs of the Natural Language Pacific Rim Symposium NLPRS 99 Beijing China pp 346 351 1999 MCKEOWN K R and RADEV D R Collocations In Robert Dale Hermann Moisl and Harold Somers editors A Handbook of Natural Language Processing pp 507 523 Marcel Dekker 2000 MONACHESI P Clitic placement in the Romanian verbal complex In B Gerlach and J Grijzenhout eds Clitics in Phonology Morphology and Syntax John Benjamins 2000 NIVRE J HALL J NILSSON J CHANEV A ERYIGIT G K BLER S MARINOV S and MARSI E Maltparser A language independent system for data driven dependency parsing Natural Language Engineering 13 pp 95 135 2007 PAD S and LAPATA M Dependency based Construction of Semantic Space Models Computational Linguistics 33 2 pp 161 199 2007 POPESCU S Gramatica practic a limbii rom ne Editura TEDIT FZH 2004 SERETAN V Collocation Extraction Based on Syntactic Parsing Ph D thesis University of Geneva 2008 SOARE G Romanian syntax Technical report Language Technology Laboratory University of Geneva 2005 SOARE G The Syntax Information Structure Interface A Comparative View from Romanian Ph D thesis University of Geneva 2009 SAUPE A TEODORESCU L R ORDEAN M A BOLDIZSAR R ORDEAN M and SILAGHI GH C Efficient parsing of Romanian language for text to speech purposes In V Matoucek and P Mautner eds Text Speech and Dialogue 2009 pp
27. nication April 2010 MaltParser is freely available at http maltparser org Accessed May 2010 http www larflast bas bg balric Accessed March 2010 Diacritics are unfortunately absent from the treebank used by C l cean amp Nivre 2009 Dan Cristea and Augusto Perez personal communication April 2010 4 Violeta SERETAN and Eric WEHRLI To summarize there is currently no large scale syntactic parser available for Romanian but work is in progress for developing both syntactically annotated resources and shallow analysis tools In this paper we present the first steps towards building a symbolic parser for Romanian able to fully analyse unrestricted text Unlike previous work concerned with dependency relations phrase chunks or specific syntactic relations in a sentence as in the Romanian Sketch Engine our parser aims to create a complete syntactic structure for the input sentence In the cases when this is not possible the parser returns disconnected parse trees for the parts of the input sentence it succeeded to analyse 3 THE FIPS MULTILINGUAL PARSER Fips Wehrli 1997 Wehrli 2007 Wehrli amp Nerima 2009 is a deep symbolic parser developed at the Language Technology Laboratory of the University of Geneva It is based on an adaptation of generative grammar concepts as inspired by the Minimalist Program Chomsky 1995 the Simpler Syntax model Culicover amp Jackendoff 2005 and the Lexical Functio
28. ntence involving a relative construction As can be seen a co indexation chain is created that links the canonical position of subject in the subordinated clause to the relative care and then to the surface position in the matrix clause see also the discussion in Section 3 DP vorbeste VP Pm talks a E A 4 mc AdvP Fata CP i Guil 7 P d AdvP mult p i much N foarte TP very i vine VP comes P dt PP AdvP la DP m ine to tomorrow noi we Fig 7 Output of FipsRomanian for a relative construction showing co indexation 12 Violeta SERETAN and Eric WEHRLI As a by product of parsing a POS tagger version of the output is available which lists detailed information for each input token A sample POS output is provided in Example 6 below As can be seen the POS tagging output also contains information on the predicate argument structure 6 Mama NOM COM SIN FEM NOM ACC 911000300 mam SUBJ spune VER INF PRE 3 SIN 911000443 spune SUB mama DO poveste IO copilului o DET IND SIN FEM NOM ACC 911067636 o OBJ poveste NOM COM SIN FEM NOM ACC 0911054772 poveste copilului NOM COM SIN MAS DAT GEN 911023143 copil IND OBJ PONC point 0 Even though FipsRomanian did not yet reached maturity we were interested in evaluating to which extent the current version can be applied on real data if it is robust enough and if the results obtained would prove already useful The remaini
29. ntion in French lit make attention to the English equivalent pay attention or Entscheidung treffen in German lit encounter a decision to the English counterpart make a decision The extraction application applies a hybrid procedure combining syntactical information and statistical methods Broadly speaking it first builds a list of syntactically valid candidates and then applies association measures to retain the candidates likely to constitute typical expression of a language Applied to Romanian this procedure collected pairs of words in predefined syntactical relations like adjective noun verb object etc from both complete and incomplete parse trees These pairs have been scored using log likelihood ratio a measure typically used for collocation extraction We manually investigated the top 2000 extraction results with the help of our concordance tool shown in Figure 8 Among the top 2000 pair types a number of 606 have been retained as lexicographically interesting and have been entered in our collocation database Thus the extraction precision is 30 3 This is a large difference from the precision of 65 9 obtained for the top 500 pair types in an experiment for French However despite the fact that Romanian version of the parser is comparatively much less developed our finding suggests that the results are already useful for practical applications The lexical coverage of Fips for French data is about 98
30. of the FipsRomanian lexicon by lexical category Table 1 Distribution of lexical categories in the FipsRomanian lexicon May 2010 Lexical category Inflected forms Lexemes Noun common 254410 38635 Noun proper 9269 9241 Pronoun 155 74 Clitic 22 5 Adjective 76203 14280 Verb 49340 3611 Adverb 879 882 Interjection 413 413 Preposition 114 116 Conjunction 72 76 Determiner 161 57 Total 391038 67390 In addition to single word entries the lexical database of the Fips parser also contains multi word entries A first category of such entries is represented by compound words for instance complex prepositions conjunctions and adverbs de jur mprejurul around dat fiind c given that p n c nd until as well as a few complex proper nouns C mpulung Moldovenesc These have however the status of single lexemes since they behave like single words A second category is represented by collocations which cover as a special case the idioms Collocations are semi compositional language specific restricted combinations of words like a atrage aten ia to draw the attention lit to attract attention Like idioms collocations pose production problems to non native language speakers but unlike idioms they do not really pose comprehension problems Idioms e g a pune la punct to fix lit to put to point are the semantically non compositional extreme of
31. oncordancer showing the top extraction results for Romanian 6 CONCLUSION A syntactic parser is a central tool for any language Largely absent from the otherwise vast repertory of Romanian language resources such a tool is under development at the University of Geneva as part of a project of multilingual extension of a symbolic parser We reported on the preliminary efforts made to build the necessary resources for the Romanian version namely the compilation of the lexicon and the specification of the grammar We also described the current state of the grammar implementation Although this is still in an incipient stage we could report on the first results obtained by parsing a large amount of unrestricted text and showed that these are useful for a practical application concerned with lexical acquisition In future work we will continue consolidating the lexicon we will concentrate on the implementation of the grammar and will perform appropriate evaluation experiments in order to provide more insightful information on the performance of the parser Extending a Multilingual Symbolic Parser to Romanian 15 ACKNOWLEDGEMENTS This work has been supported by the Swiss National Science Foundation grant no 100012 117944 The authors acknowledge the contribution of Gianina Aonofriesei Maria Husarciuc Luka Nerima Gabriela Soare and Diana Tradabat to this undertaking Particular thanks are due to the two anonymous reviewers and the a
32. s obtained it is still unclear to what extent the parser can be applied on unrestricted text In addition the reported accuracy holds for perfect POS tagged data as found in the treebank whereas in a more realistic evaluation scenario one should consider data with automatically assigned POS categories Figure 1 shows a sample dependency structure produced by MaltParser for the Romanian sentence La acest efort diplomatic participa premierul britanic Tony Blair The British Prime Minister Tony Blair is part of this diplomatic effort RELPPOZ ATSBS ATADJ ATADJ SUB ATADJ La acest Sort diplomatic participa 3 4 Iu jremierul britanic TonyBlair g 9 e 555 ADJ v5 S85 A sss m S65 Fig 1 Sample output of the Romanian dependency parser of Calacean amp Nivre 2009 In a project with similar goals a new dependency treebank is under construction at the University of Ia i which is expected to overcome the severe limitations imposed by the above mentioned treebank Finally another related report is given in Saupe et all 2009 on work done in the framework of a project on sentence analysis for Romanian Despite the title this work is not on syntactic parsing proper since it is not concerned with sentence structure It is limited to the lexical level and provides a preliminary shallow lexical analysis aimed at identifying paragraph sentence and word boundaries Adam Kilgarriff personal commu
33. tion of journalistic text and on using the parsed data in a collocation extraction application Key words Syntactic parsing symbolic methods grammar lexicon multilingualism 1 INTRODUCTION Syntactic parsing is arguably one of the most important natural language processing applications as its role is to provide the essential structural information that is required by virtually all other language applications in order to produce reliable results Syntactic parsing either symbolic rule based or stochastic statistical has been shown to considerably improve the results of many NLP applications e g term extraction Maynard amp Ananiadou 1999 semantic role labelling Gildea amp Palmer 2002 semantic similarity computation Pad amp Lapata 2007 A syntactic parser is central tool for any language However developing such a tool is an extremely complex task as natural languages as opposed to formal languages pose notoriously difficult challenges Besides this task demands depending on the approach a substantial amount of various language resources which are more or less difficult to obtain For example lexicons containing detailed morphosyntactic information associated to the items of a language are the typical resource used by natural language parsers but their construction may be prohibitively time consuming In this paper we report on the efforts undertaken at the Department of Linguistics of the University of Ge
34. udience of the PROMISE 2010 workshop in la i for useful suggestions that helped enrich the content of this paper n4 95 10 11 12 13 14 15 16 17 18 19 20 2 REFERENCES BRESNAN J Lexical Functional Syntax Blackwell 2001 CHOMSKY N The Minimalist Program MIT Press 1995 CRISTEA D Romanian language resources and tools http www clarin eu files cnl04_web pdf 2009 accessed October 2009 CRISTEA D and FOR SCU C Linguistic Resources and Technologies for Romanian Language Computer Science Journal of Moldova 14 1 40 2006 CALACEAN M and NIVRE J A data driven dependency parser for Romanian In Proceedings of the 7th International Workshop on Treebanks and Linguistic Theories TLT 7 Groningen Holland pp 65 76 2009 CULICOVER P and JACKENDOFF R Simpler Syntax Oxford University Press 2005 DEX Dic ionarul explicativ al limbii rom ne Academia Rom n 1998 DOBROVIE SORIN C The Syntax of Romanian Comparative Studies in Romance Mouton de Gruyter 1994 GILDEA D and PALMER M The necessity of parsing for predicate argument recognition In Proceedings of ACL 2002 Philadelphia 2002 KILGARRIFF A RYCHLY P SMRZ P and TUGWELL D The Sketch Engine In Proceedings of the Eleventh EURALEX International Congress Lorient France pp 105 116 2004 MAYNARD D and ANANIADOU S A linguistic approach to terminological context clustering In Proceedi
Download Pdf Manuals
Related Search