Computational Lexicons: the Neat Examples and the Odd Exemplars

Computational Lexicons: the Neat Examples and the Odd Exemplars Roberto Basili, Maria Teresa Pazienza Dip. di Ingegneria Elettronica, Universita' "Tor Vergata", Roma, Italy Paola Velardi Ist. di Informatica, Universita' di Ancona, Ancona, Italy Abstract When implementing computational lexicons it is important to keep in mind the texts that a NLP system must deal with. Words relate to each other in many different, often queer, ways: this information is rarely found in dictionaries, and it is quite hard to be invented a priori, despite the imagination that linguists exhibit at inventing esoteric examples. In this paper we present the results of an experiment in learning from corpora the frequent selectional restrictions holding between content words. The method is based on the analysis of word associations augmented with syntactic markers and semantic tags. Word pairs are extracted by a morphosyntactic analyzer and clustered according to their semantic tags. A statistical measure is applied to the data to evaluate the significance of a detected relation. Clustered association data render the study of word associations more interesting with several respects: data are more reliable even for smaller corpora, more easy to interpret, and have many practical applications in NLP. 1. Introduction One of the fundamental property of computational lexicons is an account of the relations between verbs and its arguments. Arguments are identified by their position in a predicate-argument structure, or by conceptual relations names (e.g. agent, purpose, location, etc). Arguments are annotated with selectional restrictions, that impose type constraints on the set of content words that may fill a relation. Selectional restrictions often do not provide all the semantic information that is necessary in NLP systems, however they are at the basis of the majority of computational approaches to syntactic and semantic disambiguation. It has been noticed that representing only the semantics of verbs may be inadequate (Velardi et al. 1988; Boguraev 1991; Macpherson 1991). The notion of spreading the semantic load supports the idea that every content word should be represented in the lexicon as the union of all the situations in which it could potentially participate. Unfortunately, hand writing selectional restrictions is not an easy matter, because it is time consuming and it is hard to keep consistency among the data when the lexicon has several hundred or thousand words. However the major difficulty is that words relate to each other in many different, often domain dependent ways. The nowadays vast literature on computational lexicons is filled with neat examples of the eat(animate,food) flavour, but in practice in many language domains selectional constraints between words are quite odd. It is not just a matter of violating the semantic expectations, such as in "kill the process" or "my car drinks gasoline", neither it is that kind of fancifulness that linguists exhibit at finding queer sentences. Rather, there exist statistically relevant linguistic relations that are hard to imagine a-priori, almost never found in dictionaries, and even harder to assign to the appropriate slot in the whatever conceptual structure adopted for lexical representation. Several …

Content truncated to 500 words