A computer method is disclosed for determining predicate-argument structures in input prose sentences of English. The input sentence, in the form of a string of words separated by blanks, is first analyzed (parsed) by a rule component that has access only to morphological and syntactic information about the words. The output of this rule component, in the form of a data structure consisting of attribute-value pairs, is then processed by the argument structure component, which consists of a set of partially ordered procedures that incorporate further linguistic knowledge. The output of these procedures is the same attribute-value structure, now enhanced by the presence of semantic (i.e. meaningful, non-syntactic) attributes. These semantic attributes, taken together, form the argument structure of the input sentence.
The resulting invention constitutes a fully modular, comprehensive and efficient method for passing from syntax to the first stage of semantic processing of natural (human) language. The invention applies to all prose sentences of the language for which it is designed, and not just to a subset of those sentences. It does not use domain-specific semantic information to improve the accuracy or efficiency of the syntactic component. It therefore constitutes an unrestricted broad-coverage method for natural language processing (NLP), as opposed to the restricted methods used in most NLP applications today.
Although the specific rules and procedures will be different for different natural languages, the general concept embodied in this invention is applicable to all natural languages.