Japanese Dependency Structure Analysis based on Lexicalized Statistics

Japanese Dependency Structure Analysis based on Lexicalized Statistics MASAKAZU Fujio NAIST 8916-5 Takayama, Ikorna Nara, 630-0101 JAPAN masaka-h@is.aist-nara.ac.jp Abstract \Ve JH't'::wnt statistical rnodels of Japanese dependency analysis aud report results of some experiments to in- vestigate the pnfonuancc of the models for the use fo a partical parsing system. The statistical modeb <:u'<' rather simple compared with the recent complex mod- els <wd intesivcl,y usc lexical level information 1 such a.s morplH:mcs 1 and part-of-speech tags .. \Ve conducted several expcrimcllts to show the fol- lowing properties of the modeb: lb performance of the models according to feature selec- tion ;{I performance of the models as a partial pa.rsiHg s:y:::;- tclll. The EDH .. [6] corpus was used for both tra.iuing a.ud evaluatioH of the syst(_'HL 1. Introduction A HUilllwr of statistical parsing methods have lwe11 pro·· pOS('Cl. most of the systous focus on full parsing of scl\- t<'Hces, and do not discuss the performance of partial parses, whkh b nucia.l for some application;;, such as i11forma.t.ion n'trieva.l or pre pron'ssing of corpus <HlllO- ta.liou. Early approa.clH'S of statistical pa.r,':>iHg [15, 10, 1~3] conditioned probabilitic::; on syHta(tic rules. To take lllOn: contextual iufonna.tioll into a.ccomtt, word collo- cation i.':> applied to syntactk fonnalization, Buch as k'x- ica.lized PCFG.lexica.li:zc'd tree adjoining gra.mmar 1 and lcxicalir.:cd lillk gra.nuna.r. The leugth of phrase::; or the distcuKe between head- words were also cmtsidere<l in the severa.l mocleb [16, 8] Th('re a.re parsing methods tlta.t do uot require a grammar. Colliils [a] propo.':\c's a statistical pa.nwr bas('d on probabilities of d<'].H?llclcHcics bet.W('eH head-words in parse tn'<'S. Yasnha.ra. [18L co11stnH'ts a. syst<'m based 88 YUJI Matsumoto NAIST 8916-5 Takayama, Ikoma Nara, 630-0101 JAPAN matsu@is.aist-nara.ac.jp on collocation counts as the Oilly source of grammati- cal iilfonna.tiOit. He uses co-occurr<'Hce pa.th'nts of tlw POS tags of head-words. The method, how<'H'f, is not statistical, iu that it only accumulates correct pattems for direct. use. Ma.genna.n [4] proposes a statistical parser based Oil a decision tree model, in which the probabilitiC':; arc conditioned on the derivation history of the parse tr<.'cs [4, 10]. He compares the decision tiT(' model with the n-gra.m model, a.ud cla.ims that the a.mouut of parauw- tcrs in t.h<' resulting model remaiHs relatively constant, depending mostly on the numher of traiuing exa.mph's. Cha.ntiak [5] proposes a lH'W model alld compared it with Colli11s' 1 aud IVJa.germaiL's models and shows what aspects of these systems affect. their relative per- fonnance. In general, statistical models suffc'r from the probkut of data. spa.rs<'IH.'ss. Instead of usiilg a complex sta.t.istiralinodd combi11nl with vnrions smoothillg techniq11es [1 1 2, I, 0], \Ve stick to a. statistical model of simple scttillg aiming at au ('asy implt'mentatioll, and pursue a wa.y to select usdul information for achieving higlwr parse a.c·naa.cy. T'he basic model is close to Colli11s' model[3] Ja.pa.HcS<~ dep<'ndeilcy structure are usually bct-':iNI Oil phrasa.l units (called '' b-un8ct.su 1 ' ). A bnnsetsn b<:t.sica.lly COilSifits of OilC' (or a. sequ<'ucc of) content wonl(fi) and its sucn~cding function words (that forms the …

Content truncated to 500 words