UHTIVELTSTTV 0E EHGIIIEETTIG AIID TEchlI0L0GV VIETHAM HA TI0HAL UHTIVETSITV, hAH01I HGUYEH ThAIIh huy BUILDITIG A SEMATITIcrOLELABELITIG SYSTEM FOr VIETITAMESE SENTENMcES Major : cCOmputer Science c0de 60 48 01 MASTET ThESIS Supervised by: PhD. Iguyen Phu0ng Thai han0i - 2011 'Table 0f c0nlenis Ackn0wledgeimeiÌs. ¿+ 6 + 1t 12T HH TT TH TT TH TT TH TT TH Hàng HH 3 _Y vn. 7 List Of Tables.
8 chapter I: Inlr0duCHŨNH. cOrpOra fOr furrchOn lag label1IIg. curreni sỈudies 0n FuncHOH Ìagg1IIg. 0pjecHve Of the lhes1s.
Làng HH HH Hàn HH 12 5. 14 chapter Il: I'elaled WT[KS. ó6 + 1S 1n HT HH HH TH HT TH HH nghiệt 16 1. FuncHOn Tags Lapelimg by PaTSIHE.- ch TH TH HT HH HH ch Hy 17 1.
Sequenhial FuncHOn Tag LabeliIE. --¿- + 1x11 11T TH TT HH TH ngư, 19 3. FuncH0n Tag Lapeling by classIficaHH. 5 «55s S*+*£s£serseeexree 19 SN: nh.
20 chapler III: The pr0p0sed apprOachh.- kh TT TH HT TH HT HH TH Hành 22 2. FuncHOn Tags In VielH1aIT€§€.--- + ST TT HH nghi, 23 3. «hàn HT TH Hàn HH 24 "nh. << kh TH TH TH Hành 26 5.1 Maximum Enlr0py sy M0Hivaling Example.
Maximum Enir0py M0deling. -- -¿-¿- <5 S*+*EsE‡keEskerrskeee 28 “uc ch .4 Fealures ad CŨHSỈTaiTIÌS.5 Maximum EHlr0py PTiICIDÏ€. «+ 1k S11 TT TH HH TH HT HH nh nh 32 chapter IV: Exp€TlI€HIÌ.---- ¿6 + 1t 1111 TH TH Tà TH TH HH nghiệt 34 1. cOrp0Ora arid TÖUÌs.
---- 6 6S S121 11 21 1121 11 0121 H1 Hà TH Hà HH tr 34 2. FuncHOnal Lapeling PT€CISIÔHS. SÁCH HH TH TH TH TH TT HH TH HH HH 38 4. EffecHveness 0f W0rd clusler FeaÌure .--¿-« 5<Sscxsxtersreeree 39 Bi an.
40 chapter V: cOmclusiOm ard Fulure WTÍK. 5 «+5 **kS*EEkEEkerrekrkrrrke 41 I Hi 1 ae. 45 List Of Figures Figure 1. Sample d0main and frame element Of Frame ÏIIel.
A parsing with funcH0un lags In V1Iel TTeebarIE.The perceplr0n m0del fOr funcHOn †ags labeling pr0blem. MOdel Of FumchOn Tag Labeling System fOr Vietnamese semtences. An example fOr selecled fealures In Viel TreebaH[k. Example 0f w0rd clusler hierarChyy.---- «6c St sxtereeeree 26 Figure 7.
Scenar10s 1n c0nsirained 0pHm1zaHOH. Pseud0-c0de fOr exiracHng funcHOH labeÌs. An example Of wÔrd CÏUSÏ€T. - «6 + xxx HH, 36 Figure 10.
- - ¿+ E111 S111 11T TT TH nhân 38 Figure 11. The dependency between †w0 funcHOH labeÌs .---- -«-«- 39 List Of Tables Taple1. FuncHOnal Lapeling Appr0aches.---- - 6 xà kg rê, 16 Tabple 2. I'esuli 0Ÿ laseling sy parsing appr0ach f0Il0wing c0llin m0del.
FunchOn Tags Ôn V'1e†l 'TT€€b. Vieliamese TreebaHIk sÌaÏiSÏICS. EvaluahOn Of Vieltamese funcHOnal labeling syslem. Increases im precisiOn by using wOrd clusler fealure .---- 40 chapter I: Intr0duchOn In this chapter, I intrOduce funcliOn tags and value Of funcHOn lag In IILP - applicaliOns, sOme current apprOaches, Onjective Of thesis and Our cOntrinuliOn.
Finally, I descrise the structure Of the thesis. Euncli0n tags There are tw0 kinds Of tags in linguistics: syntachic tags and funchOn tags. FOr syutache fags there are several theOries and prOQjects research result in English, Spanish, chinese and [4][13][14][18]. These research mainly fOcus On finding the parl-Of-speech and tagging fOr their cOnstituents.
FunchOn tags are undersi00d as abstract labels because they are n0t similar 10 syntactic lapels. If a syntactic label has One n0tahOn fOr a batch Of wOrds im a paragraph, fumclOn lags present the relaHOnship selween a phrase and 11s uHerance im each difference cOntext. SO fOr each phrase, funchOn lags might se transfOrming. It depends On cOntext Of its nieighbOrs.
FOr example we cOnsider a phrase: “basevall vat” syntactic Of this phrase is “nOun phzase” (im almOst research they are anmOtated as //P). but its funchOn tag might be a subject in this semtence: This vasevall vat is veey expensive In Other case its furrchOn tag might be a direct Object: 1 o0ughi this vasevall vat last mOnth Or instrument, agent in a passive vOice: That man was attacked by this basevall vat FunctiOn tags are directly mentiOned by blaheta (2003) [2]. There are a 10t Of research fOcuses im hOw 10 tag fumctiOn tags fOr a semtence. This kind Of research pr0blem is called funcHOn tags labeling pr0slem, a class Of prOblems 10 finding semantic infOrmalOn Of phrase.
TO sum up, funcHOn lag labeling is defimed as a pr0plem hOw 10 find the semantic infOrmatiOn Of a batch Of wOrds, and then tag them with a given ann0laH0u in 11s cOntext. c0rp0ra fOr functiOn tag labeling II0wadays machine learning is the pOpular methOd fOr mOst Of mOdern pr0blems especially in [lature Language PrOcessing subject. TO Build up a machine learning 10 syslem we need a lraining dala set. There are s0me funcH0n tag labeling cOrpOra that are applied fOr languages such as English and chinese.
In English, there are tw0 main cOrpOra which are used fOr semantic rOle labeling and funchOn tag labeling pr0blems. They are Frame Ilel(baker, 1998; FillMOre and baker 2000) and Pr0p bank (Palmer et al,2005).The main idea Of Frame [let is that grOup all similar wOrds in the same grOup, and then represents relatiOnship Of this grOup with Other gr0ups in a grOup-netwOrk. That why it is called Frame Ilet. Figure 1shOws a small example Domain: | branch 0n Frame IIel Frame: Questioning Frame:Conversation Elements: Elements: e Speaker e Protagonist — 1 e Message © = Prot ist -2 ic cOnfer-v re a Sons + Topic ¢ Topic © Medium © Medium Dispute-n Talk-v GOssip-v Discussi0n-n Dexale-v Figure 1.
Sample d0main and frame element Of Frame [et The secOnd cOrpus is PrOpbank (Palmer et al., 2005) which is a mOdificaliOn Of Penn Treebank by annOtated additiOn infOrmatiOn: fumchOn tags. Perm Treebank and Pr0p bank are Organized as a sel Of trees; each tree illustrates a sentence which is tagged symtactic labels (fOr Perm Treepank) and with 50th syntactic and furctiOnal labels (fOr Pr0p bank). Pr0pbank and chinese Treebank are linguistics resOurces which have been available fOr research purpOse fOr a 10ng time. Whereas, Viet Treesank! have been develped recently by Ilguyem [17] by using experiment and appr0aches from Penn Treepank.
hence, Viet Treebank has the same structure as Penm Treesank and chinese Treepank in which each wOrd is presented as leaf n0de Of a tree, n0ne terminal n0des are lagged syntactic label Or furchOnal label. Figure 2 will shOw an 11 example fr0m Viel Treebark with funcHOn †ags.Org:8080/dem0/?page=resOurces 12 NP- SUB AP-PRD | dy A-H SBAR He | /~ nỗi tiếng NPB-SUB NP-DOB is known | | la mot nha Có tải as a talented poet Figure 2. A parsing with functi0n lags in Viet Treepank 3. current studies 0n Funch0n lagging FunchOn tags labeling is am impOrtamt prOcessing step fOr mamy natural language pr0cessing applicatiOns such as queshOn answering, infOrmalOn exirachOn, and summarizaliOn.
Thus, there were sOme research that fOcused On functiOn tagging pr0blem 10 cOver additiOmal semantic infOrmatiOn which is mOre useful than syntactic labels. In 1997, cOllins [7] intrOduced the idea 10 add sOme useful syntactic infOrmatiOn, and then he pr0p0sed a parser 10 have enOugh abilily im guessing the cOmplement fag. This parser is called cOllins‘s parser, and it is cOnsidered as first system 1m lagging label. The furchOn tags labeling is defined precisely by blaheta (2003) [2].
his research used data frOm Pen Tree II which is cOvered extra funchOn lags. With blaheta“s pr0pOsal there are variOus inveshgaliOn fOcusing On functiOn tag labeling such as Merl0 and Mussill0 (2005), blaheta and charniak (2004), chrupala and Genabith (2006), Sum, Sui (2009). These studies extend furcliOn tags labeling t0pic by fOcusing On new language such as chinese, pr0p0sing mew apprOaches, Or 1nvesHgaling new features. [JOwadays, there are three maim appr0ach strategies fOr funcHOn tags labeling pr0slem: 13 The first apprOach is called pazsing, which is tagging funcHOn labels during the parsing prOcess, this appr0ach is a mOdificaliOn Of cOllin‘s parser.
FOUOwing this appr0ach, we cam cOnsider studies Of Gabbard [9], and Marcus [17]. 14 The secOmd appr0ach is called /aveling methOd which includes twO phases: extracting features and classifying funclOn labels. This appr0ach has m0re techniques because Of the diversity Of classificaliOm techniques. The mOst typical research Of this appr0ach is blaheta“s research [3].
Im his research, he has been applied sOme techniques 10 shOw the impact Of each technique fOr funchOn tag labeling pr0blem. The third appr0ach is defined as sequential labeling apprOach. FOr this appr0ach, funcliOn tags are predicted frOm Observed wOrds chain (Yuan [23]). This appr0ach is similar 10 selecting features Of classificaliOn apprOach bul the difference is that it uses a predicHOn m0del instead Of a classificatiOn mOdel.
These apprOaches will be discussed in detail in next chapter. TOday, there is a class prOslem which cOvers funchOn tagging. This class was mentiOned by carreras (2004) [5] and called Semantic [Ole Labeling. Semantic TOle Labeling is similar 10 funchOn tag labeling but it wOrks at a mOre abstract level.
When building a Semantic [Ole Laseling system, the training dala will have m0re infOrmahOn. They n0t Only include time, lOcatiOn, mannez, etc, but alsO Ovjec, insteument, agent, etc. This prOblem is a new prOmised research fOr IILP-applicahiOns which need 10 understand the meaning Of sentences. Onjective Of the thesis As we mentiOned ab0ve, assigning furrchOm tags has wide research, especially fOr English.
Tecemtly, sOme studies were applied fOr Spanish, and chinese. All funchOn lagging system have cOmtributed in their cOrpOra a semantic class which is very useful fOr Other IILP-applicatiOns such as QueshOn Answering, SummarizatiOn, InfOrmahOn Tetrieval, etc. In recent years, [lature Language PrOcessing t0pics in Vieltam have devel0ped rapidly. Especially fOr Vielmamese, many studies have fOcused On hOw 10 recOgmize the syntactic Of Vieltamese senlences by a P0S tagging system.
but unfOrtunalely, these IILP-applicaHOns d0 n0t prOvide semantic infOrmalOn fOr a sentence. 15 Whereas, s0me IILP-applicaHOns need 10 kn0w semanHic infOrmalHOn 10 answer quesh0ns: wh0, wheze, whdi, and wh0m. T0 deal with this pr0blem, Our research fOcuses On building am aul0malic funcHOn tags labeling. In this thesis, I call as stage One, tempOrarily; Our research will build a funcHOn tagging system, a prOblem that is shallOwer than Semantic Ole Labeling 16 (SIL) pr0slem, which 1s applied f0r Vielramese.
Our research Only tags sOme basic funchOn labels such as: fime, lOcatiOn, dizecdiOn etc. Others semantic rOles such as: agent, insteument, etc, which sel0ng 10 Semantic TOle Labeling pr0slem will be append intO Our system in the future. In Our system, we have tw0 phases: first we extract funchOn tags from Viel Treepank, a bank cOvered by hand-craft semantic labels. After that, we select features 10 train the classificaliOmn mOdel.
SOme features extend fr0m studies Of blaheta [2] and Yuan [22], we als0 pr0pOsed a new feature which has significant impact fOr fumclOn tag labeling system. In later chapters, we will intrOduce Our system in detail such as: feature extrachOn, selected mOdel, building mew feature, etc. Our c0nlribuH0ns As we discussed in the previ0us chapter, we aim 10 building a system which is shall0wer than Semantic ['0le Labeling f0r Vieliamese sentences. AccOrding 10 Our kn0wledge, thr0ugh there are have seen Other invesHgaHOn 1m IILPs fÚr 'Vielramese, 0ur syslem research 0n funcHOn lag laseling may se the first One.
Furtherm0re, Our system gives new 1001 10 lag funcHOmal labels fOr Viet Treebank. This will enrich Viet Treepank 10 have autOmatic instead Of hand-crafted lagging as it has dOne befOre. Viet Treebank is One Of impOrtant resOurces fOr research that study in [atural Language Pr0cessing fOr Vietnamese. If it is enriched by funcHOn tags, Others research will have mOre infOrmatiOn 0n Viehiamese especially fOr semantic infOrmatiOn.
MOreOver, im this thesis I will alsO imtrOduce sOme appr0aches 1m funcHOn lag labeling pr0blem 10 shOw sOme applied techniques. These techniques may be inherited frOm sOme researcher later. because Our system is the first One in functiOn fag labeling prOblem, we expect that it gives a base line system fOr funchOn tag labeling research later.