Smoothing of Automatically Generated Selectional Constraints

SMOOTHING OF AUTOMATICALLY GENERATED SELECTIONAL CONSTRAINTS Ralph Grishman and John Sterling Department of Computer Science New York University New York, NY 10003 ABSTRACT Frequency information on co-occurrence patterns can be automati- cally collected from a syntactically analyzed corpus; this information can then serve as the basis for selectional constraints when analyz- ing new text from the same domain. Better coverage of the domain can be obtained by appropriate generalization of the specific word patterns which are collected. We report here on an approach to au- tomatically make suitable generalizations: using the co-occurrence data to compute a confusion matrix relating individual words, and then using the confusion matrix to smooth the original frequency data. 1. INTRODUCTION Semantic (selectional) constraints are necessary for the accu- rate analysis of natural language text. Accordingly, the acqui- sition of these constraints is an essential yet time-consuming part of porting a natural language system to a new domain. Several research groups have attempted to automate this pro- cess by collecting co-occurrence patterns (e.g., subject-verb- object patterns) from a large training corpus. These patterns are then used as the source of selectional constraints in ana- lyzing new text. However, the patterns collected in this way involve specific word combinations from the training corpus. Unless the train- ing corpus is very large, this will provide only limited cover- age of the range of acceptable semantic combinations, even within a restricted domain. In order to obtain better coverage, it will be necessary to generalize from the patterns collected so that patterns with semantically related words will also be considered acceptable. In most cases this has been done by manually assigning words to semantic classes and then gen- eralizing from specific words to their classes. This approach still implies a substantial manual burden in moving to a new domain, since at least some of the semantic word classes will be domain-specific. In order to fully automate the process of semantic constraint acquisition, we would like to be able to automatically identify semantically related words. This can be done using the co- occurrence data, by identifying words which occur in the same contexts (for example, verbs which occur with the same subjects and objects). From the co-occurrence data one can compute a similarity relation between words, and then cluster words of high similarity. This approach was taken by Sekine et al. at UMIST, who then used these clusters to generalize semantic patterns [6]. A similar approach to word clustering was reported by Hirschman et al. in 1975 [5]. For our current experiments, we have adopted a slightly dif- ferent approach. We compute from the co-occurrence data a confusion matrix, which also measures the interchangeability of words in particular contexts. We then use the confusion matrix directly to generalize the semantic patterns. 2. THE NATURE OF THE CONSTRAINTS The constraints we wish to acquire are local semantic con- straints; more specifically, constraints on which words can occur together in specific syntactic relations. These include head-argument relations (e.g., subject-verb-object) and head- modifier relations. …

Content truncated to 500 words