The Semantic Interpretation of Compound Nominals | CSEE Online Publication Manager

This thesis is an investigation of how a computer can be programmed to understand the class of linguistic phenomena loosely referred to as nominal compounds, i.e. sequences of two or more nouns related through modification. Examples of the kinds of nominal compounds dealt with are: "engine repairs", "aircraft flight arrival", "aluminum water pump", and "noun noun modification". <p> The interpretation of nominal compounds is divided into three intertwined subproblems: lexical interpretation (mapping words into concepts), modifier parsing (discovering the structure of strings with more than two nominals) and concept modification (assigning an interpretation to the modification or one concept by another). This last problem is the focus of this research. The essential feature of this form of modification is that the underlying semantic relationship which exists between the two concepts is not explicit. Moreover, a large number of relationships might, in principal, exist between the two concepts. The selection of the most appropriate one depends on a host of semantic, pragmatic and contextual factors. <p> As a part of this research, a computer program has been written which builds an appropriate semantic interpretation when given a string of nouns. This program has been designed as one component of the natural language question answering system JETS. The interpretation is done by a set of semantic interpretation rules. Some of the rules are very specific, capturing the meaning of idioms and canned-phrases. Other rules are very general, representing fundamental case-like relationships which can hold between concepts. A strong attempt has been made to handle as much as possible with the more general, highly productive rules. <p> The approach has been built around a frame-based representational system which represents concepts and the relationships between them. The concepts are organized into an abstraction hierarchy which supports inheritance of attributes. The same representational system is used to encode the semantic interpretation rules. An important part of the system is the concept matcher which, given two concepts, determines whether the first describes the second and, if it does, how well.