Some association-based techniques for lexical disambiguation by machine
Philip J. Hayes · Infoscience (Ecole Polytechnique Fédérale de Lausanne) · 1977
Natural Languages are pervaded by the phenomenon of lexical ambiguity, whereby a given word may be interpreted in radically different ways depending on the context in which it appears. Such ambiguities rarely cause problems for humans; this work investigates what is required to endow Computer Natural Language Processing systems with a similar ability. Several such computer systems, which perform lexical disambiguations, are examined, and the way in which they tacitly define this task is shown to be inadequate. On the basis of these criticisms, a more adequate view is formulated; this view, which, in particular, sees the concept of word-sense as inherently vague, is used throughout the remainder of the work. A number of English examples, in which ambiguous words have unambiguous interpretations, are examined, and from them, rules are generalized which account for a class of examples much larger than those examined. These rules form the basis for the disambiguating techniques of CSAW, a working computer system for Choosing the Senses of Ambiguous Words; CSAW accepts sequences of simple English sentences as input, and produces semantic analyses of these sentences, which, in particular, resolve any lexical ambiguities the sentences contain. Several of the disambiguating techniques of CSAW are based on association, that is, the finding of a (possibly unspecified) relationship between two entities. The real-world knowledge necessary to find such associations is represented by CSAW in a large semantic net on which a higher-level structure of frames has been superimposed. This higher-level structure aids in associative searches by reducing the number of search steps required, through chunking together of closely related information; the search can also be controlled more tightly by making it sensitive to the different types of relations defined between the frames. This semantic net system is also able to represent the real-world knowledge required for CSAW's other disambiguating techniques. In addition, the representation system embodies some new approaches to several general representational problems, including multiple subparts, and property inheritance, particularly among closely related, but different, entities. An attempt is made to delimit the disambiguating powers of CSAW, by describing, in some detail, what it cannot do.