Processing Unknown Words in HPSG

Processing Unknown Words in HPSG Petra Barg and Markus Walther* Seminar ftir Allgemeine Sprachwissenschaft Heinrich~Heine-Universit~it Dfisseldorf Universit:dtsstr. 1, D-40225 Dtisseldorf, Germany {barg, walther}@ling.uni-.duesseldorf, de Abstract The lexical acquisition system presented in this pa- per increlnenlally updates linguistic properties of un- known words inferred lrom their surrounding con- lexI by parsing senlences wilh an liPSG gramnmr for Gerlmm. We employ a gradual, infornmtion- based concept of "tlnknowntless" providing a uni- form treatment for the range oi completely known to maximally unknown lexical entries. "Unknown" in- formation is viewed as revisable inR)rmation, which is either generalizable or specializable. Updating takes place after parsing, which only requires a mod- ified lexical lookup. Revisable pieces of informa- lion are idenlilied by grammar-specilied declaralions which provide access pmhs into lhe parse feature slructure. The updating nlechanism revises the col responding places in the lexical lcalure stmclures iff the conlext actually provides new information. For revising generalizable inlbrmalioi< (ype union is re- quired. A worked-out example demonstrates the in- Iercntial capacily of our implemented system. 1 Introduction It is a ,emarkable fact thai humans can often un- derstand sentences containing unknown words, in- let their grammatical properties and incrementally refine hypotheses aboul these words when encoun- tering laler instances. In conlrasl, many current NLP systems still prestlppose a complele lexicon. Notable exceptions include Zernik (1989), Erbach (199()), Hastings & Lylinen (1994). See Zernik lot an intro- duction to the general issues involved. 11fis paper describes an HPSG-based system which can incrementally learn and reline proper- lies of unknown words after parsing individual sen- *This work was tan'ted out within the ,S'onde@)rschungs- bereich 282 '77worie des Lexikol~s' (project B3), funded by the German Federal Research Agency DFG. We thank James Kil- bury and lnembers of the B3 group for fruitful discussion. tences. It focusses oll extracting linguislic proper- ties, as compared to e.g. general concept learning (Hahn, Klenner & Schnatiinger 1996). Unlike Er- bach (1990), however, it is not conlined to siul- pie morpho-synlactic information but can also han- (lie selectional reslriciions, senianlic types and argu- ment slructure. Finally, while statistical approaches like Brenl (1991) can gather e.g. valence inloruia- lion lioni large corpora, we rote more interested in full gralmnatical processing of individual sentences to maximally exploil each context. The following three goals serve to struclure our model. It should i) incorporate a gradual, informalion-based conceptualization of "unknown- ness". Words are not unknown as a whole, but may contain ux~k~mw~t, i.e. revisable pieces of b(for- ,la*iom Consequently, even known words can un- der(,o> revision 1o e.>.<) acquire new senses. This view replaces the binm-y distinclion belween open and closed class wordx. 1l should it) maxinmlly exph)il lhe rich represeniaiions and nlodelling convenlions of IiPSG and associaled formalisnls, wiih essen- iially the same grammar anti lexicon as compared (o closed-lexicon approaches. This is important bolh lo facilitate reuse of existing grammars and li) en- able meaningful feedback for linguistic theorizing. Finally, it should iii) possess donlain-independenl in- ference and lexicon-updating capabililies. The gram- mar writer must be …

Content truncated to 500 words