Consider first the use of features in semantic description. It is not radical to propose that words are categorized in terms of general classes. Nor is it novel to consider the fact that these classes bear particular relations to each other. Consider, for example, the English words in the matrix presented in (1):

In this chart words are represented in terms of particular binary dimensions. In each case the particular semantic dimension, or ‘semantic feature’, is the marking which specifies particular aspects of the semantic patterning of the lexical item. In (2) the positive value for each feature is associated with some sample sentence frames which indicate the kind of systematic restrictions each feature imposes. Thus in (2a) any noun marked ‘+ human’ can, among other things, appear in the sentence ‘The noun thought the problem over’, but any noun marked ‘ — animate’ cannot appear in the frame ‘the noun sensed the danger’; any noun marked ‘ + plant’ can appear in the frame ‘the noun has damaged clorophyll’. The privileges of occurrence in these kinds of frames of the different nouns in the figure are indicated by the placement of the ‘ + ’ and ‘ — ’ markings on the separate features.

It has rarely been questioned that such classifications play a role in natural language.² But the problem has been to decide which of the many possible aspects of classification are to be treated as systematically pertinent to a semantic theory. We cannot claim to have discovered all and only the semantic features of natural language. However, it is absolutely necessary to assume that there is some universal set from which particular languages draw their individual stock.

Furthermore, if the semantic analysis of natural language is to achieve explanatory adequacy, there must be a principled and precise manner of deciding among competing semantic analyses. In this way the semantic analysis which is ultimately chosen can be said to be the result of a formal device, and not due to the luck of a formal linguist. If the particular analysis is chosen on the basis of precise, formal criteria then the reality of the predictions made by the device constitutes a confirmation of the general linguistic theory itself.

Lexical analysis of single words includes a specification in terms of a set of semantic features. We propose, furthermore, that the most highly valued semantic analysis which meets these constraints be the one which utilizes the smallest number of symbols in a particular form of semantic analysis. For a given natural language this will direct the choice of which features are drawn from the universal set as well as the assignment of predictable features in the lexicon itself.

¹ This work was supported by the MITRE corporation, Bedford, Mass., Harvard Society of Fellows, NDEA, A.F. 19(68)-5705 and Grant # SD-187,1 Rockefeller University and IBM. A preliminary version of this paper was originally written as part of a series of investigations at MITRE corporation, summer 1963, and delivered to the Linguistic Society of America, December 1964. We are particularly grateful to Dr D. Walker for his support of this research and to P. Carey and G. A. Miller for advice on this manuscript.

The paper is reprinted from R. A. Jacobs and P. S. Rosenbaum (eds.), Readings in English Transformational Grammar, Blaisdell, 1970.

² Modern discussions of such features and their integration within grammatical theory can be found in Katz and Fodor, 1963; Chomsky, 1965; Miller, 1967. In this article we will assume that the reader has a basic familiarity with the role of semantic analysis in current transformational linguistic theory.