Generative Operations in a Constraint-based Grammar ?

One of the challanges for building natural language processing systems is that of reducing the size of the lexicon while still accounting for the polymorphic behavior of language. One instance of verbal polymorphism is the ability of certain predicates (e.g. enjoy, promise , begin, etc.) to surface with diierent complement structures, which requires, under traditional approaches, multiple listing of lexical entries. To avoid such redundancies in the lexicon we adopt the semantic operation of type coercion as deened in Generative Lexicon theory (GL, 13],,14],,15]) and show that integrating it into HPSG requires only some tuning of existing principles, without the need for additional stipulations 3. In particular, we will discuss how the control properties of coerced nom-inals can be inferred from the general HPSG control theory, and then, argue that the introduction of type coercion mechanisms (henceforth TC) within an HPSG grammar will greatly reduce the size of the lexicon , while still accounting for all the available semantic interpretations. Abstract. Wir unterbreiten einen Vorschlag zur Integration einer The-orie lexikalischer Semantik wie 'Generative Lexicon Theory (GL)' in HPSG. Wir gehen das Problem an indem wir untersuchen, wie Typ-Coercion (in seiner GL Formulierung) in HPSG behandelt werden kann, und zeigen, dass dies nur einige kleine Anpassungen der existierenden HSPG Prinzipien erfordert. Zwei Dinge machen es interessant, die beiden Theorien zu kombinieren: ein GL Ansatz traegt dazu bei, unter Beibehal-tung der notwendigen Menge von Information die Groesse des Lexikons zu reduzieren; und HPSG bietet ohne zusaetzliche Stipulationen Erklaer-ungen fuer einige Beobachtungen in GL. ? We would like to thank P. Sag and two anonymous reviewers for helpful comments and discussion. 3 The approach described here has been wholly implemented using an HPSG grammar developped at CSLI, which augments Penn and Carpenter's version. Our implementation runs in ALE, but the approach will t any constraint-based framework with a type inference mechanism.