Selection and information: a class-based approach to lexical relationships

Selectional constraints are limitations on the applicability of predicates to arguments. For example, the statement "The number two is blue" may be syntactically well formed, but at some level it is anomalous-- scBLUE is not a predicate that can be applied to numbers. In this dissertation, I propose a new, information-theoretic account of selectional constraints. Unlike previous approaches, this proposal requires neither the identification of primitive semantic features nor the formalization of complex inferences based on world knowledge. The proposed model assumes instead that lexical items are organized in a conceptual taxonomy according to class membership, where classes are defined simply as sets--that is, extensionally, rather than in terms of explicit features or properties. Selection is formalized in terms of a probabilistic relationship between predicates and concepts: the selectional behavior of a predicate is modeled as its distributional effect on the conceptual classes of its arguments, expressed using the information-theoretic measure of relative entropy. The use of relative entropy leads to an illuminating interpretation of what selectional constraints are: the strength of a predicate's selection for an argument is identified with the quantity of information it carries about that argument. In addition to arguing that the model is empirically adequate, I explore its application to two problems. The first concerns a linguistic question: why some transitive verbs permit implicit direct objects ("John ate $\emptyset$") and others do not ("*John brought $\emptyset$"). It has often been observed informally that the omission of objects is connected to the ease with which the object can be inferred. I have made this observation more formal by positing a relationship between inferability and selectional constraints, and have confirmed the connection between selectional constraints and implicit objects in a set of computational experiments. Second, I have explored the practical applications of the model in resolving syntactic ambiguity. A number of authors have recently begun investigating the use of corpus-based lexical statistics in automatic parsing; the results of computational experiments using the present model suggest that often lexical relationships are better viewed in terms of underlying conceptual relationships such as selectional preference and concept similarity. Thus the information-theoretic measures proposed here can serve not only as components in a theory of selectional constraints, but also as tools for practical natural language processing.