A Rule-Based and MT-Oriented Approach to Prepositional Phrase Attachment

A Rule-Based and MT-Oriented Approach to Prepositional Phrase Attachment Kuang-hua Chen and Hsin-Hsi Chen l)eparl;nmnl; of C, ompul;er Science a.nd Informal;ion Engilleering Na.domd Taiwa. (lnivcrsity I, St'](',. 4, Roosevelt 1{1)., Ta.il)ei TAIWAN, 1{.0.(:. khcheil<~nlg.csie, ntu.ed u.tw, hh_chcn({~csic.nl u.cdu.tw Abstract I'rel)ositional l'hrase is the key issue in structuraJ ambiguity, l{ecently, re- searches in corpora provide the lexical cue of pre[)ositions with other words and the information could be used to l)artly resolve ambiguity resulted from prepositionM phrases. Two possible a~- t.achments are considered in the lit, era- ture: either I]OUll attaehluelll[, or verb attachment. In this paper, we con- sider the problem fl'om viewpoint of m~> chine translation. Four different attach- ments arc told out according to their ['unctiona.lity: llOU[l attatchmellt, verb at.ta.clllllellt, sentence-level att&chlneut, ~md predicate-level attachtllelll,. [~oth lcxi(:al k.owledge and semantic knowl- edge are involved resolving atta.chment in tl~e proposed me(:hanism, t';xperimen- tal results show that considering more types of prepositional phrases is useful in machine translation. 1 Introduction Prepositional phrases are usually ambiguous. The wen-known sentence shown in the following is a good example. Kevin watched the girl with a telescope. Whether the prepositionM phre~se with a tele- scope modifies the hea.d noun girl or the verb watch are not resolved by using only one knowl- edge sour(:e. Many researchers observe text cor- pora and learll some knowledge based on language model I,o determine the plausible attachment. For example, we could expect that the aforementioned prepositional phrase is usually attached to verb according to text corpora. However, the col rect attachment is dependent on world knowledge SOlfleti/TleS. Some al)l)roaches to determination of Pl)s arc rel)orl,cd in literature. (Kimball, 1973; Frazier, 1978; Ford et al., 1982; Shieber, 1983; Wilks ct al., 1985; Liu et aJ., [990; (',hen and C:hcn, [992: Ilindle and Rooth, 1993; |bill and l{esnik, 1.994). 'l.'he possible attachment they conskler are NO UN attachment and VI.]RJ/ aU;aehment. These res- olutions fall into three categories: sy.tax-based, semantics-based and corpus-based a.pproaches. The brief discussion are described in the follow- ing: I. Syntax-tx~sed • Right Association (Kimball, 1973) The PPs ;dwa.ys modifies the nearest component pre(:eding i t. • Minimal At.t.achmcnl. (Frazier, 1978: Shieber, 1!.)83) The correct attaching point of' a 1'1 ) in a sentence is determined by l,he nmnber of nodes in a p~ursing tree. 2. Semantics-based • l,exicM Pret>rence (Ford et a.l., ]982) The real attaching point llulst S;'l..tist~ r some constraints, e.g., w~rb ligatures. I)it[~rent verbs accolnl)anying with the same PPs may have the different att.~u:h- lug points. • l're, fi~rence Semantics (Wilks et al., 1!)85) Wilks amd his colleagues argue the real attaching l)oint must be determined lay the preference of verbs a.nd prepositions. • Propagated Semantics ((:hen a.nd (7:hen, 1992) The attachment of prepositional l)hrasc is co-determined by the semantic usage of noun, w~rt), attd preposition. 3. Corpus-based • Sta.dstical Score (I,iu at a.l., 1990) They use semantic score and synta.ctic 216 s(x)re t,c) (]el,l'rnfitie l;h,+ ai,l,a/:hing i)(>iut,+ 'l'[l(<s(, ,,4(.c)1'c,,,4 Ftl'(~ (,l'+~,ill(+(I [1'()111 Lexl, /:c)r +. ])()l'a. • I,c×i,a.I A++.'+<wia+l.icm (tlindle a.ud I{oc+t+h, …

Content truncated to 500 words