Structural Methods For Lexical/Semantic Patterns

This paper represents initial work on corpus methods for acquiring lexical/semantic pattern lexicons for text understanding. Recently, implementors of information extraction (IE) systems have moved away from using conventional syntactic parsing methods, instead adopting a variety of pattern based approaches for complex IE tasks. While there has been much work towards automated acquisition of lexicons for conventional syntactic processing, little progress has been made towards those for pattern systems, due primarily, in the author 's opinion, to a lack of a linguistic framework in which to view their use. In combining a functional view of both denotational semantics and syntactic structure, this paper provides a basis for examining the structural constraints between the two. This functional viewpoint is the starting point for methods to investigate the characteristics of the interaction between text and denotation, from the perspective of patternbased systems. An approach for determining and exploiting these structural constraints is outlined in terms of building hierarchical lexical structures for text understanding. Experiment results for such a method are given, demonstrating the functionality of the approach. 1. I n t r o d u c t i o n . Recently, implementors of information extraction (IE) systems have moved away from using conventional syntactic parsing methods, instead adopting a variety of pattern based approaches for complex IE tasks (such as the MUC contests and the ARPA-sponsored TIPSTER research). These pattern-based systems use short and fairly specific lexical patterns to specify the relation between strings in the source text and particular entries in a problem-dependent knowledge representation. This one-step process substitutes for a conventional two-level process of a full syntactic parse followed by semantic interpretation. With considerably less time and development effort (notably demonstrated by [11, 8]), these systems achieve performance comparable to more standard systems that rely heavily on full syntactic analysis ([9, 5]). However, because these pattern-based systems are still viewed as linguistically ungrounded and somewhat ad hoc, formal work in the application and acquisition of lexical patterns has lagged system development. In most current systems, patterns are produced through tedious hand analysis of text ([11, 4, 8]), while system coverage is gained either through extensive linguistic knowledge on the part of the researcher (in judging appropriate pattern generalizations), or by generating and testing massive numbers of patterns. One exception to hand analysis is Lehnert's work in [12], in which machine learning techniques are used to infer possible patterns for extraction. While this AutoSlog technique has dramatically reduced system development time, the inference techniques use only sparse linguistic information, provide no means of generalizing patterns across domains, and still require that the rules be checked by the researcher for applicability. 1 .1 . A t h e o r e t i c a l f r a m e w o r k By placing pattern-based approaches in a lexical semantic framework, such as Pustejovsky's Generative Lexicon theory ([16]), my aim is to provide a basis for patternbased understanding systems which can be used to relate pattern-based approaches to more conventional linguistic approaches. Within such a framework, methods for pattern acquisition can be studied and developed and the effectiveness of patterns can be assessed. My main contention is that this framework can be developed by viewing the lexieal patterns as structural mappings from text to denotation in a compositional lexical semantics, merging the distinction between syntactic and semantic analysis, and obviating the need for separate syntactic and semantic processing systems. This interpretation follows directly from an appeal to functional semantic principles. In the framework I present, patterns indexed to individual words relate semantic interpretations to lexical constraints, in a manner dictated by context. Patterns for multiple words in context can be combined to provide a consistent interpretation for large constructions a mechanism that could be viewed as a lexically distributed semantic grammar. A combined approach to pattern acquisition is outlined here, with two orthogonal methods whose combination leads to the construction of organized sets of lex-