<?xml version="1.0"?><!DOCTYPE article SYSTEM "/project/take/software/searchbench_offline_processing/paperxml_generator/aclextractor/src/python/../resource/dtd/paperxml.dtd"><article><header><firstpageheader><page local="1" global="477"/><title>Word Sense Disambiguation Incorporating Lexical and Structural Semantic Information</title><pubinfo>Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning,pp. 477-485, Prague, June 2007. ©2007 Association for Computational Linguistics</pubinfo><author surname="Tanaka" givenname="Takaaki"><org  name="NTT Communication Science Laboratories" country="Japan" city="Kyoto"/></author><author surname="Bond" givenname="Francis"><org  name="NTT Communication Science Laboratories" country="Japan" city="Kyoto"/></author><author surname="Baldwin" givenname="Timothy"><org  name="University of Melbourne" country="Australia" city="Melbourne"/></author><author surname="Fujita" givenname="Sanae"><org  name="NTT Communication Science Laboratories" country="Japan" city="Kyoto"/></author><author surname="Hashimoto" givenname="Chikara"><org  name="National Institute of Information and Communications Technology" country="Japan" city="Kyoto"/></author></firstpageheader><frontmatter><p><b>Word Sense Disambiguation Incorporating Lexical and Structural Semantic Information</b></p><p><b>Takaaki Tanaka^ Francis Bond^ Timothy Baldwin* Sanae Fujita^ Chikara Hashimoto*</b></p><p>Î (takaaki, sanaej@cslab.kecl.ntt.co.jp ^ bond@nict.go.jp * tim@csse.unimelb.edu.au * ch@yz.yamagata-u.ac.jp</p><p>î NTT Communication Science Laboratories, Nippon Telegraph and Telephone Corporation ^ National Institute of Information and Communications Technology * The University of Melbourne * Yamagata University</p></frontmatter><abstract>We present results that show that incorporat­ing lexical and structural semantic informa­tion is effective for word sense disambigua­tion. We evaluated the method by using pre­cise information from a large treebank and an ontology automatically created from dic­tionary sentences. Exploiting rich semantic and structural information improves preci­sion 2-3%. The most gains are seen with verbs, with an improvement of 5.7% over a model using only bag of words and n-gram features. </abstract></header><body><section number="1" title="Introduction"><p>Recently, significant improvements have been made in combining symbolic and statistical approaches to various natural language processing tasks. In parsing, for example, symbolic grammars are be­ing combined with stochastic models (Riezler et al., 2002; Oepen et al., 2002; Malouf and van Noord, 2004). Statistical techniques have also been shown to be useful for word sense disambiguation (Steven­son, 2003). However, to date, there have been few combinations of sense information together with symbolic grammars and statistical models. Klein and Manning (2003) show that much of the gain in statistical parsing using lexicalized models comes from the use of a small set of function words. Features based on general relations provide little improvement, presumably because the data is too sparse: in the Penn treebank normally used to train and test statistical parsers <i>stocks </i>and <i>skyrocket </i>never appear together. They note that this should motivate the use of similarity and/or class based approaches:</p><p>the superordinate concepts <i>capital </i><i>(D</i><i> stocks) </i>and <i>move upward </i><i>(D</i><i> sky rocket) </i>frequently appear to­gether. However, there has been little success in this area to date. For example, Xiong et al. (2005) use se­mantic knowledge to parse Chinese, but gain only a marginal improvement. Focusing on WSD, Steven­son (2003) and others have shown that the use of syntactic information (predicate-argument relations) improve the quality of word sense disambiguation the effectiveness of the selectional preference infor­mation for WSD. However, there is still little work on combining WSD and parse selection.</p><doubt alpha="66.7" length="45" tooSmall="False" monospace="0.0">(WSD). McCarthy and Carroll (2003) have shown</doubt><p>We hypothesize that one of the reasons for the lack of success is that there has been no resource annotated with both syntactic (or structural seman­tic information) and lexical semantic information. For English, there is the SemCor corpus (Fellbaum, 1998) which is annotated with parse trees and Word­Net senses, but it is fairly small, and does not ex­plicitly include any structural semantic information. Therefore, we decided to construct and use a tree-bank with both syntactic information (e.g. HPSG parses) and lexical semantic information (e.g. sense tags): the Hinoki treebank (Bond et al., 2004). This can be used to train word sense disambiguation and parse ranking models using both syntactic and lexi­cal semantic features. In this paper, we discuss only word sense disambiguation. Parse ranking is dis­cussed in Fujita et al. (2007).</p></section><section number="2" title="The Hinoki Corpus"><p>The Hinoki corpus consists of the Lexeed Seman­tic Database of Japanese (Kasahara et al., 2004) and corpora annotated with syntactic and semantic information.<page local="2" global="478"/></p><subsection number="2.1" title="Lexeed"><p>Lexeed is a database built from on a dictionary, which defines word senses used in the Hinoki cor­pus and has around 49,000 dictionary definition sen­tences and 46,000 example sentences which are syn­tactically and semantically annotated. Lexeed con­sists of all words with a familiarity greater than or equal to five on a scale of one to seven. This gives a fundamental vocabulary of 28,000 words, divided into 46,347 different senses. Each sense has a defi­nition sentence and example sentence written using only these 28,000 familiar words (and some function words). Many senses have more than one sentence in the definition: there are 75,000 defining sentences in all.</p><p>A (simplified) example of the entry for <i>Mfi</i><i> </i><i>"f-</i><i> un-tenshu </i>"chauffeur" is given in Figure 1. Each word contains the word itself, its part of speech (POS) and lexical type(s) in the grammar, and the familiarity score. Each sense then contains definition and ex­ample sentences, links to other senses in the lexicon (such as hypernym), and links to other resources, such as the Goi-Taikei (Ikehara et al., 1997) and WordNet (Fellbaum, 1998). Each content word in the definition and example sentences is annotated with sense tags from the same lexicon.</p></subsection><subsection number="2.2" title="Lexical Semantics Annotation"><p>The lexical semantic annotation uses the sense in­ventory from Lexeed. All words in the fundamental vocabulary are tagged with their sense. For example, the word <i>ookii </i>"big" (in <i>ookiku naru </i>"grow up") is tagged as sense 5 in the example sentence (Figure 1), with the meaning "elder, older".</p><p>Each word was annotated by five annotators. We use the majority choice in case of disagreements (Tanaka et al., 2006). Inter-annotator agreements among the five annotators range from 78.7% to 83.3%: the lowest agreement is for the Lexeed def­inition sentences and the highest is for Kyoto cor­pus (newspaper text). These agreements reflect the difficulties in disambiguating word sense over each corpus and can be considered as the upper bound of precision for WSD.</p><table caption="Table 1 shows the distribution of word senses ac­cording to the word familiarity in Lexeed."></table></subsection><subsection number="2.3" title="Ontology"><p>The Hinoki corpus comes with an ontology semi-automatically constructed from the parse results of definitions in Lexeed (Nichols and Bond, 2005). The ontology includes more than 80 thousand relation­ships between word senses, e.g. synonym, hyper-nym, abbreviation, etc. The hypernym relation for <i>untenshu </i>"chauffeur" is shown in Figure 1. Hypernym or synonym relations exist for almost all content words.</p></subsection><subsection number="2.4" title="Thesaurus"><p>As part of the ontology verification, all nominal and most verbal word senses in Lexeed were linked to semantic classes in the Japanese thesaurus, Nihongo Goi-Taikei (Ikehara et al., 1997). These were then hand verified. Goi-Taikei has about 400,000 words including proper nouns, most nouns are classified into about 2,700 semantic classes. These seman­tic classes are arranged in a hierarchical structure (11 levels). The Goi-Taikei Semantic Class for <i>M untenshu </i>"chauffeur" is shown in Figure 1: <b>(C292:driver) </b>at level 9 which is subordinate to <b>(C4:person).</b></p></subsection><subsection number="2.5" title="Syntactic and Structural Semantics Annotation"><p>Syntactic annotation is done by selecting the best parse (or parses) from the full analyses derived by a broad-coverage precision grammar. The gram­mar is an HPSG implementation (JACY: Siegel and Bender, 2002), which provides a high level of de­tail, marking not only dependency and constituent structure but also detailed semantic relations. As the grammar is based on a monostratal theory of gram­mar (HPSG: Pollard and Sag, 1994) it is possible to simultaneously annotate syntactic and semantic structure without overburdening the annotator. Us­ing a grammar enforces treebank consistency — all sentences annotated are guaranteed to have well-<page local="3" global="479"/></p><table caption="Table i: Word Senses in Lexeed" class="main" frame="box" rules="all" border="0" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p></p></td><td class="cell"><p>PoIy-</p></td><td class="cell"><p></p></td><td class="cell"><p>#Mono-</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Fa m</p></td><td class="cell"><p>#Words</p></td><td class="cell"><p>semous</p></td><td class="cell"><p>#WS</p></td><td class="cell"><p>semous(%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>6.5 -</p></td><td class="cell"><p>368</p></td><td class="cell"><p>182</p></td><td class="cell"><p>4.0</p></td><td class="cell"><p>186 (50.5)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>6.G -</p></td><td class="cell"><p>4,445</p></td><td class="cell"><p>1,902</p></td><td class="cell"><p>3.4</p></td><td class="cell"><p>2,543 (57.2)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>B.B -</p></td><td class="cell"><p>9,814</p></td><td class="cell"><p>3,502</p></td><td class="cell"><p>2.7</p></td><td class="cell"><p>6,312 (64.3)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>5.0 -</p></td><td class="cell"><p>11,430</p></td><td class="cell"><p>3,457</p></td><td class="cell"><p>2.5</p></td><td class="cell"><p>7,973 (69.8)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><p>Index POS</p><p>Lex-Type Familiarity</p><doubt alpha="83.3" length="6" tooSmall="False" monospace="0.0">Sense1</doubt><p><b>jSIs^ </b><i>untenshu </i>noun</p><doubt alpha="29.2" length="24" tooSmall="False" monospace="0.0">noun-lex 6.2 [1-7] (&gt; 5)</doubt><p>Definition</p><doubt alpha="100.0" length="7" tooSmall="False" monospace="0.0">Example</doubt><p>Hypernym Sem. Class WordNet</p><p>Ï1 <b>^</b><b> </b><b>dWJ</b>&amp;i 2r Sfc <i>~ti&gt; </i>à4   a person who drives trains and cars</p><p>I dream of growing up and becoming a train driver</p><doubt alpha="65.3" length="49" tooSmall="False" monospace="0.0">A4hito"person" (292:driver)(C(4:person))motorman1</doubt><figure caption='Figure 1: Dictionary Entry forjHk^untenshu"chauffeur"'></figure><p>formed parses. The flip side to this is that any sen­tences which the parser cannot parse remain unan-notated, at least unless we were to fall back on full manual mark-up of their analyses. The actual anno­tation process uses the same tools as the Redwoods treebank of English (Oepen et al., 2002).</p><p>There were 4 parses for the definition sentence shown in Figure 1. The correct parse, shown as a phrase structure tree, is shown in Figure 2. The two sources of ambiguity are the conjunction and the rel­ative clause. The parser also allows the conjunction to join to <i>densha </i>and <i>hito. </i>In Japanese, rel­ative clauses can have gapped and non-gapped read­ings. In the gapped reading (selected here), <i>hito </i>is the subject of <i>unten </i>"drive". In the non-gapped reading there is some underspecified relation between the thing and the verb phrase. This is sim­ilar to the difference in the two readings of <i>the day he knew </i>in English: "the day that he knew about" (gapped) vs "the day on which he knew (some­thing)" (non-gapped). Such semantic ambiguity is resolved by selecting the correct derivation tree that includes the applied rules in building the tree.</p><p>The parse results can be automatically given by the HPSG parser PET (Callmeier, 2000) with the</p><p>Japanese grammar JACY. The current parse ranking model has an accuracy of 70%: the correct tree is ranked first 70% of the time (for Lexeed definition sentences) (Fujita et al., 2007).</p><p>The full parse is an HPSG sign, containing both syntactic and semantic information. A view of the semantic information is given in Figure 3<footnote anchor="1"/>.</p><p>The specific meaning representation language used in</p><p>UTTERANCE</p><doubt alpha="100.0" length="1" tooSmall="False" monospace="0.0">i</doubt><doubt alpha="100.0" length="2" tooSmall="False" monospace="0.0">NP</doubt><doubt alpha="33.3" length="9" tooSmall="False" monospace="0.0">VP_____ N</doubt><doubt alpha="61.5" length="13" tooSmall="False" monospace="0.0">mm4&gt;imm.-th x</doubt><doubt alpha="55.4" length="101" tooSmall="False" monospace="0.0">densha    ya    jidousha       o       unten   suru hitotrain      or        caraccdrive    do person</doubt><p><b>jSls</b>^i "chauffeur": "a person who drives a train or car"</p><figure caption='Figure 2: Syntactic View of the Definition ofWM.1untenshu"chauffeur"'></figure><p>The semantic view shows some ambiguity has been resolved that is not visible in the purely syn­tactic view.</p><p>The semantic view can be further simplified into a dependency representation, further abstracting away from quantification, as shown in Figure 4. One of the advantages of the HPSG sign is that it contains all this information, making it possible to extract the particular view needed. In order to make linking to other resources (such as the sense annotation) easier, predicates are labeled with pointers back to their po­sition in the original surface string. For example, the predicate densha_n_l links to the surface characters between positions 0 and 3: .</p><p>JACY is Minimal Recursion Semantics (Copestake et al., 2005).</p><page local="4" global="480"/><p>TEXT TOP</p><doubt alpha="100.0" length="2" tooSmall="False" monospace="0.0">hl</doubt><p>"propos it ion_m_r el"</p><p>LBL <i>hl </i>ARGO <i>e2</i></p><doubt alpha="83.3" length="6" tooSmall="False" monospace="0.0">MARGh3</doubt><p>"unknown_rel</p><p>LBL <i>h4 </i>ARGO <i>e2</i></p><doubt alpha="80.0" length="5" tooSmall="False" monospace="0.0">ARGx5</doubt><doubt alpha="80.0" length="5" tooSmall="False" monospace="0.0">LBLh6</doubt><doubt alpha="83.3" length="6" tooSmall="False" monospace="0.0">ARGOx7</doubt><doubt alpha="66.7" length="18" tooSmall="False" monospace="0.0">"udef _rel "LBLhl5</doubt><p>ARGO <i>xl2 </i>RSTR <i>hl6 </i>BODY <i>hl7_</i></p><doubt alpha="66.7" length="9" tooSmall="False" monospace="0.0">"_unten_s</doubt><p>_j idousha_n</p><doubt alpha="83.3" length="6" tooSmall="False" monospace="0.0">LBLhl8</doubt><doubt alpha="85.7" length="7" tooSmall="False" monospace="0.0">ARGOxl2</doubt><doubt alpha="71.4" length="7" tooSmall="False" monospace="0.0">_hito_n</doubt><doubt alpha="66.7" length="6" tooSmall="False" monospace="0.0">LBLh24</doubt><doubt alpha="83.3" length="6" tooSmall="False" monospace="0.0">ARGOx5</doubt><doubt alpha="66.7" length="6" tooSmall="False" monospace="0.0">LBLh22</doubt><p>ARGO <i>e23 </i>tense=present ARGl <i>x5</i></p><p>Larg2 <i>xl3</i></p><doubt alpha="63.5" length="63" tooSmall="False" monospace="0.0">HCONS{h3qeqh4,h9qeqh6,hl6qeqhll,h2Oqeqhl8,h26qeqh24,h28qeq h22}</doubt><p>ING    <i>{h24 </i>ing <i>hlOOOl}</i></p><p>udef _rel</p><p>LBL <i>hl9 </i>ARGO <i>xl2 </i>RSTR <i>h2O </i>BODY <i>h2l</i></p><p>RSTR <i>h26 </i>BODY <i>h27</i></p><p>"udef _rel LBL <i>h8</i></p><p>RSTR <i>h9 </i>BODY <i>hlO</i></p><doubt alpha="60.0" length="5" tooSmall="False" monospace="0.0">-ya-P</doubt><doubt alpha="100.0" length="3" tooSmall="False" monospace="0.0">LBL</doubt><doubt alpha="100.0" length="4" tooSmall="False" monospace="0.0">ARGO</doubt><doubt alpha="100.0" length="3" tooSmall="False" monospace="0.0">hll</doubt><doubt alpha="66.7" length="3" tooSmall="False" monospace="0.0">xl3</doubt><p>L-INDEX <i>x7 </i>R-INDEX <i>xl2</i></p><p>propos it ion_m_r el</p><p>LBL <i>hlOOOl </i>ARGO <i>e23 </i>tense=present MARG <i>h28</i></p><figure caption='Figure 3: Semantic View of the Definition ofjSIk^iuntenshu"chauffeur"'></figure><doubt alpha="61.5" length="39" tooSmall="False" monospace="0.0">_1:proposition_m&lt;0:13&gt;[MARG e2:unknown]</doubt><doubt alpha="53.1" length="32" tooSmall="False" monospace="0.0">e2:unknown&lt;0:13&gt;[ARG x5:_hito_n]</doubt><doubt alpha="35.7" length="14" tooSmall="False" monospace="0.0">x7:udef&lt;0:3&gt;[]</doubt><doubt alpha="44.4" length="18" tooSmall="False" monospace="0.0">x7:densha_n_1&lt;0:3&gt;</doubt><doubt alpha="33.3" length="15" tooSmall="False" monospace="0.0">x12:udef&lt;4:7&gt;[]</doubt><doubt alpha="47.6" length="21" tooSmall="False" monospace="0.0">x12:_j idousha_n&lt;4:7&gt;</doubt><doubt alpha="48.6" length="144" tooSmall="False" monospace="0.0">x13:_ya_p_conj&lt;0:4&gt;[L-INDEX x7:_densha_n_1, R-INDEX x12:_jidousha_n] e23:_unten_s_2&lt;8:10&gt;[ARG1 x5:_hito_n, ARG2 x13:_ya_p_conj] x5:udef&lt;12:13&gt;[]</doubt><doubt alpha="53.5" length="43" tooSmall="False" monospace="0.0">_2:proposition_m&lt;0:13&gt;[MARG e23:_unten_s_2]</doubt><figure caption='Figure 4: Dependency View of the Definition ofjSIk^iuntenshu"chauffeur"'></figure><doubt alpha="66.7" length="6" tooSmall="False" monospace="0.0">3 Task</doubt><p>We define the task in this paper as "allocating the word sense tags for all content words included in Lexeed as headwords, in each input sentence". This task is a kind of all-words task, however, a unique point is that we focus on fundamental vocabulary (basic words) in Lexeed and ignore other words. We use Lexeed as the sense inventory. There are two problems in resolving the task: how to build the model and how to assign the word sense by using the model for disambiguating the senses. We de­scribe the word sense selection model we use in sec­tion 4 and the method of word sense assignment in section 5.</p></subsection></section><section number="4" title="Word Sense Selection Model"><p>All content words (i.e. basic words) in Lexeed are classified into six groups by part-of-speech: noun, verb, verbal noun, adjective, adverb, others. We treat the first five groups as targets of disambiguat­ing senses. We build five words sense models corre­sponding to these groups. A model contains senses for various words, however, features for a word are discriminated from those for other words so that the senses irrelevant to a target word are not selected. For example, an n-gram feature following a target word "has-a-tail" for <i>dog </i>is distinct from that for <i>cat.</i></p><p>In the remainder of this section, we describe the features used in the word sense disambiguation. First we used simple n-gram collocations, then a bag of words of all words occurring in the sentence. This was then enhanced by using ontological information and predicate argument relations.</p><subsection number="4.1" title="Word Collocations"><p>Word collocations (WORD-Col) are basic and effec­tive cues for WSD. They can be modelled by n-gram and bag of words features, which are easily extracted from a corpus. We used all unigrams, bi-grams and trigrams which precede and follow the target words (N-gram) and all content words in the sentences where the target words occur (BOW).</p><doubt alpha="66.7" length="6" tooSmall="False" monospace="0.0">LBLh25</doubt><page local="5" global="481"/><p># sample features</p><doubt alpha="61.5" length="13" tooSmall="False" monospace="0.0">CI (COLWS:A4)</doubt><p>C2 (COLWSSC:C33:other person)</p><p>C3 (COLWShyp^ISJi)</p><p>C4 (COLWSHYPSC:C5:person)</p><doubt alpha="46.2" length="13" tooSmall="False" monospace="0.0">C1 (COLWS:!^)</doubt><doubt alpha="66.7" length="30" tooSmall="False" monospace="0.0">C2 &lt;COLWSSC:C988:land vehicle)</doubt><p>C3 (COLWSHYP^Mi)</p><p>C4 (COLWS HYPSC:C988:land vehicle)</p><doubt alpha="57.1" length="14" tooSmall="False" monospace="0.0">C1 (COLWS^Rj^)</doubt><doubt alpha="64.7" length="17" tooSmall="False" monospace="0.0">C3 (COLWSHYP::fe)</doubt><p>Table 2: Example semantic collocation features (SEM-Col) extracted from the word sense tagged cor­pus and the dictionary (Lexeed and GoiTaikei) and the ontology which have the word senses and the se­mantic classes linked to the semantic tags. The first column numbers the feature template corresponding to each example.</p></subsection><subsection number="4.2" title="Semantic Features"><p>We use the semantic information (sense tags and on­tologies) in two ways. One is to enhance the collo­cations and the other is to enhance dependency rela­tions.</p><subsubsection number="4.2.1" title="Semantic Collocations"><p>Word surface features like N-gram and BOW in­evitably suffer from data sparseness, therefore, we generalize them to more abstract words or concepts and also consider words having the same mean­ings. We used the ontology described in Sec­tion 2.3 to get hypernyms and synonyms and the Goi-Taikei thesaurus to abstract the words to the se­mantic classes. The superordinate classes at level 3, 4 and 5 are also added in addition to the original semantic class. For example, <i>densha </i>"train" and § Uj <i>M jidousha </i>"automobile" are both gener­alized to the semantic class <b>(C988: land vehicle) </b>(level 7). The superordinate classes are also used: <b>(C706:inanimate) </b>(level 3), <b>(C760:artifact) </b>(level 4) and <b>(C986:vehicle) </b>(level 5).</p></subsubsection><subsubsection number="4.2.2" title="Semantic Dependencies"><p>The semantic dependency features are based on a predicate and its arguments taken from the ele­mentary dependencies. For example, consider the semantic dependency representation for <i>densha ya</i></p><doubt alpha="66.7" length="30" tooSmall="False" monospace="0.0"># sample features forM%,~i~Z&gt;1</doubt><doubt alpha="38.9" length="36" tooSmall="False" monospace="0.0">D322 (PRED:C2003, ARG1:A4,ARG2:H$.1)</doubt><p>Table 3: Example semantic features extracted from the dependency tree in Figure 4. The first column numbers the feature template corresponding to each example.</p><p><i>jidousha-wo unten suru hito </i>"a person who drives a train or car" given in Figure 4. The predicate <i>un­ten </i>"drive", has two arguments: arg1 <i>hito </i>"person" and arg2 <i>ya </i>"or". The coordinate conjunction is expanded out into its children, giving arg2 <i>densha </i>"train" and <i>jidousha </i>"automobile".</p><p>From these, we produce several features, a sam­ple of them are shown in Table 3. One has all argu­ments and their labels (D11). We also produce var­ious back offs, for example the predicate with only one argument at a time (D1-D3). Each combination of predicate and its related argument(s) becomes a feature.</p><p>For the next class of features, we used the sense information from the corpus combined with the se­mantic classes in the dictionary to replace each predicate by its disambiguated sense, its hypernym, its synonym (if any) and its semantic class.<page local="6" global="482"/> The seman­tic classes for and g iS^a are both <b>(988: land vehicle), </b>while jSfjh is <b>(2003:motion) </b>and A4is <b>(4:human). </b>We also expand @f§jj^4 into its syn­onym ^ — <i>9 — ij </i><b>— i </b><i>motaka </i>"motor car".</p><table class="main" frame="box" rules="all" border="0" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Dl Dl Dl</p></td><td class="cell"><p>(PRED (PRED (PRED</p></td><td class="cell"><p>, , ,</p></td><td class="cell"><p>ARG1:A)</p><p>ARG2: ) ARG2: )</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>D2 D2 D2</p></td><td class="cell"><p>(PRED (PRED (PRED</p></td><td class="cell"><p>, , ,</p></td><td class="cell"><p>ARGl: 4) ARG2:    1 ) ARG2:      1 )</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>D3 D3</p></td><td class="cell"><p>(PRED (PRED</p></td><td class="cell"><p>, ,</p></td><td class="cell"><p>ARGlSC:C33)</p><p>ARG2SC :C988)</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>D4</p></td><td class="cell"><p>(PRED</p></td><td class="cell"><p>,</p></td><td class="cell"><p>ARG2syn::E-*-#-1)</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>D5 D5 D5</p></td><td class="cell"><p>(PRED (PRED (PRED</p></td><td class="cell"><p>, , ,</p></td><td class="cell"><p>ARGlHYP:   1 ) ARG2HYP:   1 ) ARG2HYP: 2)</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Da Da</p></td><td class="cell"><p>(PRED (PRED</p></td><td class="cell"><p>, ,</p></td><td class="cell"><p>ARGlHYPSC:C5)ARG2HYPSC:C988)</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Dll</p><p>D22 D23</p></td><td class="cell"><p>(PRED (PRED (PRED</p></td><td class="cell"><p>, , ,</p></td><td class="cell"><p>ARGl: , ARG2: ) ARGl: 4, ARG2:    1 ) ARGl: 4, ARG2:Cl4a0 )</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>D24</p></td><td class="cell"><p>(PRED</p></td><td class="cell"><p>,</p></td><td class="cell"><p>ARG1:à4, ARG2syn:*-:?-</p></td><td class="cell"><p>1)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>D32 D33</p></td><td class="cell"><p>(PRED (PRED</p></td><td class="cell"><p>, ,</p></td><td class="cell"><p>ARGl:C5, ARG2:    1 )</p><p>ARG1:C5, ARG2:C988)</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>D55 D5a Da5</p></td><td class="cell"><p>(PRED (PRED (PRED</p></td><td class="cell"><p>, , ,</p></td><td class="cell"><p>ARG1HYP:ABÜ4, ARG2HYP:jg ARG1hyp:AIH4, ARG2hypsc ARG1HYPSC :C5 , ARG2HYP:|</p></td><td class="cell"><p>1) :C988)</p><p>1)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><p>The semantic class features provide a seman­tic smoothing, as words are binned into the 2,700 classes. The hypernym/synonym features provide even more smoothing. Both have the effect of mak­ing more training data available for the disambigua-tor.</p></subsubsection></subsection><subsection number="4.3" title="Domain"><p>Domain information is a simple and sometimes strong cue for disambiguating the target words (Gliozzo et al., 2005). For instance, the sense of the word <i>"record" </i>is likey to be different in the mu­sical context, which is recalled by domain-specific words like <i>"orchestra", "guitar", </i>than in the sport­ing context. We use 12 domain categories like "cul­ture/art", "sport", etc. which are similar to ones used in directory search web sites. About 6,000 words are automatically classified into one of 12 domain categories by distributions in web sites (Hashimoto and Kurohashi, 2007) and 10% of them are manually checked. Polysemous words which belong to multi­ple domains and neutral words are not classified into any domain.</p></subsection></section><section number="5" title="Search Algorithm"><p>The conditional probability of the word sense for each word is given by the word sense selection model described in Section 4. In the initial state, some of the semantic features, e.g. semantic col­locations (SEM-Col) and word sense extensions for semantic dependencies (SEM-Dep) are not available, since no word senses for polysemous words have been determined. It is not practical to count all com­binations of word senses for target words, therefore, we first try to decide the sense for that word which is most plausible among all the ambiguous words, then, disambiguate the next word by using the sense.</p><p>We use the beam search algorithm, which is sim­ilar to that used for decoder in statistical machine translation (Watanabe, 2004), for finding the plausi­ble combination of word sense tags.</p><p>The algorithm is described as follows. For a pol-ysemous word set in an input sentence {w1,..., <i>wn</i>}, <i>tWkk </i>is the k-th word sense of word wi, <i>W </i>is a set having words to be disambiguated, <i>T </i>is a list of re­solved word senses. A search node <i>N </i>is defined as <i>[W,</i><i> </i><i>T</i>] and a score of a node N, <i>s(N)</i><i> </i>is defined as the probability that the word sense set <i>T </i>occurs in the context. The beam search can be done as fol­lows (beam width is b):</p><p>1. Create an initial node N0 = [T0, W0] (T0 = {}, W0 = {}) and insert the node into an initial queue <i>Q0.</i></p><p>2. For each node <i>N </i>in the queue Q, do the follow­ing steps.</p><p>• For each <i>wi </i>(e <i>W</i>), create <i>W\</i><i> </i>by picking out <i>w </i>from <i>W</i></p><p>• Create new lists <i>T[</i><i>T{</i><i> </i>by adding one of word sense candidates tw1,... <i>,twii </i>for <i>wi</i></p><doubt alpha="100.0" length="3" tooSmall="False" monospace="0.0">toT</doubt><p>• Create new nodes <i>[Wj,</i><i> </i>T0'],... <i>,[WJ,</i><i> </i><i>T/]</i><i> </i>and insert them into the queue <i>Q</i></p></section><section number="3." title="Sort the nodes in Q  by the score s(N)"><p>4. If the top node <i>W </i>in the queue <i>Q </i>is empty, adopt <i>T </i>as the combination of word senses and terminate. Otherwise, pick out the top <i>b </i>nodes from <i>Q</i><i> </i>and insert them into new queue <i>Q,</i><i> </i>then go back to 2</p></section><section number="6" title="Evaluation"><p>We trained and tested on the Lexeed Dictionary Def­inition (LXD-DEF) and Example sections (LXD-EX) of the Hinoki corpus (Bond et al., 2007). These have about 75,000 definition and 46,000 example sen­tences respectively. Some 54,000 and 36,000 sen­tences of them are treebanked, i.e., they have the syntactic trees and structural semantic information. We used these sentences with the complete informa­tion and selected 1,000 sentences out of each sen­tence class as test sets (LXD-DEFtest, LXD-EXtest), and the remainder is combined and used as a training set (LXD-ALL). We also tested 1,000 sentences from the Kyoto Corpus of newspaper text (KYOTOtest). These sentences have between 3.4 (LXD-EXtest) - 5.2 (KYOTOtest) polysemous words per sentence on av­erage.</p><page local="7" global="483"/><p>We use a <i>maximum entropy / minimum diver­gence </i>(MEMD) modeler to train the word sense se­lection model. We use the open-source Maximun Entropy Modeling Toolkit<footnote anchor="2"/> for training, determining best-performing convergence thresholds and prior sizes experimentally. The models for five differ­ent POSs were trained with each training sets: the base model is word collocation model (WORD-Col), and the semantic models built by semantic colloca­tion (SEM-Col), semantic dependency (SEM-Dep) or domain with WORD-Col (+SEM-Col, +SEM-Dep and +DOMAIN).</p><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">0.9</doubt><doubt alpha="100.0" length="1" tooSmall="False" monospace="0.0">u</doubt><doubt alpha="33.3" length="6" tooSmall="False" monospace="0.0">&lt;D0.4U</doubt><doubt alpha="0.0" length="13" tooSmall="False" monospace="0.0">^ 0.3 0.2 0.1</doubt><doubt alpha="0.0" length="9" tooSmall="False" monospace="0.0">1 2 3 4 5</doubt><p><b>size of training corpus (partition)</b></p><figure caption="Figure 5: Learning Curve"></figure></section><section number="7" title="Results and Discussion"><p>Table 4 shows the precision as the results ofthe word sense disambiguation on the combination of LXD-DEF and LXD-EX (LXD-ALL ). The baseline method selects the senses occurring most frequently in the training corpus. Each row indicates the results us­ing the baseline, word collocation (WORD-Col), the combinations of WORD-Col and one of the seman­tic features (+SEM-Col, +SEM-Dep and +DOMAIN), e.g, +SEM-Col gives the results using WORD-Col and SEM-Col, and all features (FULL).</p><p>There are significant improvements over the base­line and the other results on all corpora. Basic word collocation features (WORD-Col) give a vast improve­ment. Extending this by using the ontological in­formation (+SEM-Col) gives a further improvement over the WORD-Col. Adding the predicate-argument relationships (+SEM-Dep) improves the results even more.</p><footnote label="2">http://homepages.inf.ed.ac.uk/s0450736/ maxent_toolkit.html</footnote><p>Table 6 shows the statistics of the target corpora. The best result of LXD-DEFtest (80.7%) surpasses the inter-annotator agreement (78.7%) in building the Hinoki Sensebank. However, there is a wide gap between the best results of KYOTOtest (60.4%) and the inter-annotator agreement (83.3%), this suggests other information such as the semantic classes for named entities (including proper nouns and multi­word expressions (MWE)) and broader contexts are required. However, a model built on dictionary sen­tences lacks these features. Even, so there is some improvement.</p><p>The domain features (+DOMAIN) give small con­tribution to the precision, since only intra-sentence context is counted in this experiment. Unfortunately dictiory definition and example sentences do not re­ally have a useful context. We expect broader con­text should make the domain features more effective for the newspaper text (e.g. as in Stevenson (2003)),</p><p>Table 5 shows comparison of results of different POSs. The semantic features (+SEM-Col and +SEM-Dep ) are particularly effective for verb and also give moderate improvements on the results of the other</p><p>POSs.</p><p>Figure 5 shows the precisions of LXD-DEFtest in changing the size of a training corpus, which is di­vided into five partitions. The precision is saturated in using four partitions (264,000 tokens).</p><p>These results of the dictionary sentences are close to the best published results for the SENSEVAL-2 task (79.3% by Murata et al. (2003) using a com­bination of simple Bayes learners). However, we are using a different sense inventory (Lexeed not Iwanami (Nishio et al., 1994)) and testing over a dif­ferent corpus, so the results are not directly compa­rable. In future work, we will test over SENSEVAL-2 data so that we can compare directly.</p><p>None of the SENSEVAL-2 systems used onto-logical information, despite the fact that the dic­tionary definition sentences were made available, and there are several algorithms describing how to extract such information from MRDs (Tsurumaru et al., 1991; Wilkes et al., 1996; Nichols et al., 2005).<page local="8" global="484"/> We hypothesize that this is partly due to the way the task is presented: there was not enough time to ex­tract and debug an ontology as well as build a dis­ambiguation system, and there was no ontology dis­tributed. The CRL system (Murata et al., 2003) used a syntactic dependency parser as one source of fea­tures (KNP: Kurohashi and Nagao (2003)), remov­ing it decreased performance by around 0.6%.</p></section><section number="8" title="Conclusions"><p>We used the Hinoki corpus to test the importance of lexical and structural information in word sense dis­ambiguation. We found that basic n-gram features and collocations provided a great deal of useful in­formation, but that better results could be gained by using ontological information and semantic depen­dencies.</p></section><section title="Acknowledgements"><p>We would like to thank the other members of the NTT Natural Language Research Group NTT Com­munication Science laboratories for their support. We would also like to express gratitude to the re­viewers for their valuable comments and Professor Zeng Guangping, Wang Daliang and Shen Bin of the University of Science and Technology Beijing (USTB) for building the demo system.</p><table caption="Table 4: The Precision ofWSD" class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Model</p></td><td class="cell"><p>Test</p></td><td class="cell"><p>Baseline</p></td><td class="cell"><p>WORD-Col</p></td><td class="cell"><p>+SEM-Col</p></td><td class="cell"><p>+SEM-Dep</p></td><td class="cell"><p>+DOMAIN</p></td><td class="cell"><p>FULL</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>LXD-ALL</p></td><td class="cell"><p>LXD-DEFtest</p></td><td class="cell"><p>72.8</p></td><td class="cell"><p>78.4</p></td><td class="cell"><p>79.8</p></td><td class="cell"><p>80.2</p></td><td class="cell"><p>78.1</p></td><td class="cell"><p>80.7</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>LXD-EXtest</p></td><td class="cell"><p>70.4</p></td><td class="cell"><p>75.6</p></td><td class="cell"><p>78.7</p></td><td class="cell"><p>77.9</p></td><td class="cell"><p>76.0</p></td><td class="cell"><p>78.8</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>KYOTOtest</p></td><td class="cell"><p>55.6</p></td><td class="cell"><p>58.5</p></td><td class="cell"><p>60.0</p></td><td class="cell"><p>58.8</p></td><td class="cell"><p>59.8</p></td><td class="cell"><p>60.4</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><table caption="Table 5: The Precision of WSD (per Part-of-Speech)" class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>POS</p></td><td class="cell"><p>Baseline</p></td><td class="cell"><p>WORD-Col</p></td><td class="cell"><p>+SEM-C0I</p></td><td class="cell"><p>+SEM-Dep</p></td><td class="cell"><p>+DOMAIN</p></td><td class="cell"><p>FULL</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Noun</p></td><td class="cell"><p>65.5</p></td><td class="cell"><p>68.7</p></td><td class="cell"><p>69.6</p></td><td class="cell"><p>69.4</p></td><td class="cell"><p>68.9</p></td><td class="cell"><p>69.8</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Verb</p></td><td class="cell"><p>60.3</p></td><td class="cell"><p>66.9</p></td><td class="cell"><p>71.0</p></td><td class="cell"><p>70.a</p></td><td class="cell"><p>67.7</p></td><td class="cell"><p>72.6</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>VN</p></td><td class="cell"><p>72.6</p></td><td class="cell"><p>76.2</p></td><td class="cell"><p>77.7</p></td><td class="cell"><p>74.a</p></td><td class="cell"><p>77.6</p></td><td class="cell"><p>77.5</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Adj</p></td><td class="cell"><p>59.9</p></td><td class="cell"><p>67.2</p></td><td class="cell"><p>69.5</p></td><td class="cell"><p>a8.9</p></td><td class="cell"><p>68.9</p></td><td class="cell"><p>69.5</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Adv</p></td><td class="cell"><p>74.4</p></td><td class="cell"><p>78.6</p></td><td class="cell"><p>79.8</p></td><td class="cell"><p>79.2</p></td><td class="cell"><p>78.6</p></td><td class="cell"><p>79.8</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><table caption="Table 6: Corpus Statistics" class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Annotated</p></td><td class="cell"><p></p></td><td class="cell"><p>Agreement</p></td><td class="cell"><p></p></td><td class="cell"><p></p></td><td class="cell"><p></p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Corpus</p></td><td class="cell"><p>Tokens</p></td><td class="cell"><p>#WS</p></td><td class="cell"><p>token (type)</p></td><td class="cell"><p>%Other Sense</p></td><td class="cell"><p>%Homonym</p></td><td class="cell"><p>%MWE</p></td><td class="cell"><p>%Proper Noun</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>LXD-DEF</p></td><td class="cell"><p>199,268</p></td><td class="cell"><p>5.18</p></td><td class="cell"><p>.787 (.850)</p></td><td class="cell"><p>4.2</p></td><td class="cell"><p>0.084</p></td><td class="cell"><p>1.5</p></td><td class="cell"><p>0.046</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>LXD-EX</p></td><td class="cell"><p>126,966</p></td><td class="cell"><p>5.00</p></td><td class="cell"><p>.820 (.871)</p></td><td class="cell"><p>2.3</p></td><td class="cell"><p>0.035</p></td><td class="cell"><p>0.4</p></td><td class="cell"><p>0.0018</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>KYOTO</p></td><td class="cell"><p>268,597</p></td><td class="cell"><p>3.93</p></td><td class="cell"><p>.833 (.828)</p></td><td class="cell"><p>9.8</p></td><td class="cell"><p>3.3</p></td><td class="cell"><p>7.9</p></td><td class="cell"><p>5.5</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table></section><references><p>Francis Bond, Sanae Fujita, Chikara Hashimoto, Kaname Kasahara, Shigeko Nariyama, Eric Nichols, Akira Ohtani, Takaaki Tanaka, and Shigeaki Amano. 2004. The Hinoki treebank: A treebank for text understanding. In <i>Proceed­ings of the First International Joint Conference on Natural</i> <i>Language Processing (IJCNLP-04)</i>, pages 554-559. Hainan Island.</p><p>Francis Bond, Sanae Fujita, and Takaaki Tanaka. 2007. The Hi­noki syntactic and semantic treebank of Japanese. <i>Language Resources and Evaluation. </i>(Special issue on Asian language technology).</p><p>Ulrich Callmeier. 2000. PET - a platform for experimentation with efficient HPSG processing techniques. <i>Natural Lan­guage Engineering, </i>6(1):99-108.</p><p>Ann Copestake, Dan Flickinger, Carl Pollard, and Ivan A. Sag. 2005. Minimal Recursion Semantics. An introduction. <i>Re­search on Language and Computation, </i>3(4):281-332.</p><p>Christine Fellbaum, editor. 1998. <i>WordNet: An Electronic Lex­ical Database. </i>MIT Press.</p><p>Sanae Fujita, Francis Bond, Stephan Oepen, and Takaaki Tanaka. 2007. Exploiting semantic information for HPSG parse selection. In <i>ACL 2007 Workshop on Deep Linguistic Processing, </i>pages 25-32. Prague, Czech Republic.</p><p>Alfio Massimiliano Gliozzo, Claudio Giuliano, and Carlo Strapparava. 2005. Domain kernels for word sense disam­biguation. In <i>Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL 2005)</i>. Ann</p><p>Arbor, U.S.</p><p>Chikara Hashimoto and Sadao Kurohashi. 2007. Construction of domain dictionary for fundamental vocaburalry. In <i>Pro­ceedings of the ACL 2007 Main Conference Poster Sessions. </i>Association for Computational Linguistics, Prague, Czech Republic.</p><p>Satoru Ikehara, Masahiro Miyazaki, Satoshi Shirai, Akio Yokoo, Hiromi Nakaiwa, Kentaro Ogura, Yoshifumi Ooyama, and Yoshihiko Hayashi. 1997. <i>Goi-Taikei — A Japanese Lexicon. </i>Iwanami Shoten, Tokyo. 5 vol-umes/CDROM.</p><p>Kaname Kasahara, Hiroshi Sato, Francis Bond, Takaaki Tanaka, Sanae Fujita, Tomoko Kanasugi, and Shigeaki Amano. 2004. Construction of a Japanese semantic lexicon: Lexeed. In <i>IPSG SIG: 2004-NLC-159, </i>pages 75-82. Tokyo. (in Japanese).</p><p>Dan Klein and Christopher D. Manning. 2003. Accurate un-lexicalized parsing. In Erhard Hinrichs and Dan Roth, edi­tors, <i>Proceedings ofthe 41st Annual Meeting ofthe Associ­ation for Computational Linguistics, </i>pages 423-430. URL http://www.aclweb.org/anthology/P03-1054.pdf.</p><page local="9" global="485"/><p>Sadao Kurohashi and Makoto Nagao. 2003. Building a Japanese parsed corpus — while improving the parsing sys­tem. In Anne Abeille, editor, <i>Treebanks: Building and Using Parsed Corpora, </i>chapter 14, pages 249-260. Kluwer Aca­demic Publishers.</p><p>Robert Malouf and Gertjan van Noord. 2004. Wide cover­age parsing with stochastic attribute value grammars. In <i>IJCNLP-04 Workshop: Beyond shallow analyses - For­malisms and statistical modeling for deep analyses. </i>JST CREST. URL http: //www-tsujii. is . s .u-tokyo. ac. jp/bsa/papers/malouf.pdf.</p><p>Diana McCarthy and John Carroll. 2003. Disambiguat-ing nouns, verbs and adjectives using automatically ac­quired selectional preferences. <i>Computational Linguistics, </i>29(4):639-654.</p><p>Masaaki Murata, Masao Utiyama, Kiyotaka Uchimoto, Qing Ma, and HItoshi Isahara. 2003. CRL at Japanese dictionary-based task of SENSEVAL-2. <i>Journal of Natural Language Processing, </i>10(3):115-143. (in Japanese).</p><p>Eric Nichols and Francis Bond. 2005. Acquiring ontologies using deep and shallow processing. In <i>11th Annual Meeting ofthe Association for Natural Language Processing, </i>pages 494-498. Takamatsu.</p><p>Eric Nichols, Francis Bond, and Daniel Flickinger. 2005. Ro­bust ontology acquisition from machine-readable dictionar­ies. In <i>Proceedings ofthe International Joint Conference on Artificial Intelligence IJCAI-2005, </i>pages 1111-1116. Edin­burgh.</p><p>Minoru Nishio, Etsutaro Iwabuchi, and Shizuo Mizutani. 1994. <i>Iwanami Kokugo Jiten Dai Go Han [Iwanami Japanese Dic­tionary Edition 5]. </i>Iwanami Shoten, Tokyo. (in Japanese).</p><p>Stephan Oepen, Kristina Toutanova, Stuart Shieber, Christoper D. Manning, Dan Flickinger, and Thorsten Brant. 2002. The LinGO redwoods treebank: Motivation and preliminary applications. In <i>19th International Confer­ence on Computational Linguistics: COLING-2002, </i>pages 1253-7. Taipei, Taiwan.</p><p>Carl Pollard and Ivan A. Sag. 1994. <i>Head Driven Phrase Struc­ture Grammar. </i>University ofChicago Press, Chicago.</p><p>Stefan Riezler, Tracy H. King, Ronald M. Kaplan, Richard Crouch, John T. Maxwell, and Mark Johnson. 2002. Parsing the Wall Street Journal using a Lexical-Functional Grammar and discriminative estimation techniques. In <i>41st Annual Meeting ofthe Association for Computational Linguistics: ACL-2003, </i>pages 271-278.</p><p>Melanie Siegel and Emily M. Bender. 2002. Efficient deep pro­cessing of Japanese. In <i>Proceedings of the 3rd Workshop on Asian Language Resources and International Standardiza­tion at the 19th International Conference on Computational Linguistics, </i>pages 1-8. Taipei.</p><p>Mark Stevenson. 2003. <i>Word Sense Disambiguation. </i>CSLIPub-lications.</p><p>Takaaki Tanaka, Francis Bond, and Sanae Fujita. 2006. The Hi­noki sensebank — a large-scale word sense tagged corpus of Japanese —. In <i>Proceedings of the Workshop on Frontiers in Linguistically Annotated Corpora 2006, </i>pages 62-69. Syd­ney. URL http://www.aclweb.org/anthology/W/W06/ W06-0608, (ACL Workshop).</p><p>Hiroaki Tsurumaru, Katsunori Takesita, Itami Katsuki, Toshi-hide Yanagawa, and Sho Yoshida. 1991. An approach to thesaurus construction from Japanese language dictionary. In <i>IPSJ SIGNotes Natural Language, </i>volume 83-16, pages 121-128. (in Japanese).</p><p>Taro Watanabe. 2004. <i>Example-based Statistical Machine Translation. </i>Ph.D. thesis, Kyoto University.</p><p>Yorick A. Wilkes, Brian M. Slator, and Louise M. Guthrie.</p><p>1996. <i>Electric Words. </i>MIT Press.</p><p>Deyi Xiong, Qun Liu Shuanglong Li and, Shouxun Lin, and Yueliang Qian. 2005. Parsing the Penn Chinese treebank with semantic knowledge. In Robert Dale, Jian Su Kam-Fai Wong and, and Oi Yee Kwong, editors, <i>Natural Language Processing—IJCNLP 005: Second International Joint Con­ference Proceedings, </i>pages 70-81. Springer-Verlag.</p></references></body></article>