Browse Prior Art Database

A"-PROCESS TO IMPLEMENT SOMEWORD-SENSE~DISAMBIGUATIONS

IP.com Disclosure Number: IPCOM000128625D
Original Publication Date: 1979-Dec-31
Included in the Prior Art Database: 2005-Sep-16
Document File: 15 page(s) / 59K

Publishing Venue

Software Patent Institute

Related People

Philip J. Hayes: AUTHOR [+3]

Abstract

One of the major problems to be faced in the construction of computer programs to understand natural language is that of word sense ambiguity. Although a word (e.g. ball) may have two or more quite different meanings, when a human reads (or hears) it in context he will almost always be able to correctly choose which of those meanings the author intended to convey. He is able to disambiguate the word. Any computer program which claims to understand natural language must be able to do the same. What follows is an outline of a programmable process for making disambiguations. The process deals with a somewhat neglected area in natural language processing, the interface between parsing and inference. As such the process does not address all disambiguation problems. Indeed, if it did it would have addressed almost the entire natural language understanding problem. The disambiguations it does tackle can be thought of in four categories, characterizable in terms of four different ytpes of influece which can determine the meaning of ambiguous words. I shall refer to these four types as: selectional restrictions, preference restrictions, associations, and logical restrictions. They are described below by example. Selectional restrictions (henceforth srs) are perhaps the most straightforward. They restrict the types of entities that can participate in certain actions or relations. Thus in: (1) Fred kicked the ball the sr that only physical objects may be kicked dismbiguates "ball" to mean spherical object rather than formal dance. Srs are absolute and can never be overriden. They are bas3ed on intrinsic and normally unchangeable properties. Srs were first described systematically by Katz and Fodor (1963).Preference restrictions (henceforth prs) are the sam as srs except that they can be violated if need. They, in general, express ides of normality. Thus in the isolated sentence: (2) The stock tasted good "stock" is disambiguated to mean a type of soup by the pr that taste (used intrasitively) takes a food as its subject. However, in: (3) Fred licked his rifle all over. The stock tasted good this restriction is ovoerriden and "stock" is disambiguated to refer to the wooden part of a rifle. The machine translation system of Yorick Wilks (1975) makes considerable use of the notion of preference. Association is the influence which overrides the pr and effects the disambiguation in (3). The appropriate sense of "stock" is associated with rifle because it is part of a rifle. Much vaguer associations can also effect dismbiguation as in:

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 5% of the total text.

Page 1 of 15

THIS DOCUMENT IS AN APPROXIMATE REPRESENTATION OF THE ORIGINAL.

A"-PROCESS TO IMPLEMENT SOMEWORD-SENSE~DISAMBIGUATIONS

Philip J. Hayes Computer Science Department The University of Rochester Word-sense ambiguity is one of the major problems to be faced by any natural language understanding system. A process is proposed to implement disambiguations explainable by four classes of influence --selectional restrictions, preference restrictions, logical restrictions, and association. The process operates as an interface between parsing and deeper processing. The treatment of association is based on a frame-like organizational system for semantic nets. This system provides a means of representing context and facilitates the search for an associa-tive connexion between a word-sense and an already existing context. In addition, it is argued that selectional restrictions really are absolute, even though sentences exist in which they are apparently violated.

This paper reports research undertaken while the author was a member of the Istituto per gli Studs Semantics a Cognitivi, 17 Rue de Candolle, 1205 Geneva, Switzerland: The paper was largely written during that period and is also published as a Working Paper of the above named institution.

Introduction

One of the major problems to be faced in the construction of computer programs to understand natural language is that of word sense ambiguity. Although a word (e.g. ball) may have two or more quite different meanings, when a human reads (or hears) it in context he will almost always be able to correctly choose which of those meanings the author intended to convey. He is able to disambiguate the word. Any computer program which claims to understand natural language must be able to do the same.

What follows is an outline of a programmable process for making disambiguations. The process deals with a somewhat neglected area in natural language processing, the interface between parsing and inference. As such the process does not address all disambiguation problems. Indeed, if it did it would have addressed almost the entire natural language understanding problem. The disambiguations it does tackle can be thought of in four categories, characterizable in terms of four different ytpes of influece which can determine the meaning of ambiguous words. I shall refer to these four types as: selectional restrictions, preference restrictions, associations, and logical restrictions. They are described below by example.

Selectional restrictions (henceforth srs) are perhaps the most straightforward. They restrict the types of entities that can participate in certain actions or relations. Thus in: (1) Fred kicked the ball the sr that only physical objects may be kicked dismbiguates "ball" to mean spherical object rather than formal dance. Srs are absolute and can never be overriden. They are bas3ed on intrinsic and normally unchangeable properties. Srs were first described systematically by Katz and Fodor (1...