Augmenting Lexicons Automatically: Clustering Semantically Related Adjectives

Our work focuses on identifying various types of lexical data in large corpora through statistical analysis. In this paper, we present a method for grouping adjectives according to their meaning, as a step towards the automatic identification of adjectival scales. We describe how our system exploits two sources of linguistic knowledge in a corpus to compute a measure of similarity between two adjectives, using statistical techniques and a clustering algorithm for grouping. We evaluate the significance of the results produced by our system for a sample set of adjecfives. 1. I N T R O D U C T I O N A linguistic scale is a set of words, of the same grammatical category, which can be ordered by their semantic strength or degree of informativeness [1]. For example, "lukewarm," "wa rm " , "ho t " fall along a single adjectival scale since they indicate a variation in the intensity of temperature of the modified noun. Linguistic properties of scales derive both from conventional logical entailment on the linear ordering of their elements and from Gricean scalar implicature [1]. Despite these properties and their potential usefulness in both understanding and generating natural language text, dictionary entries are largely incomplete for adjectives in this regard. Yet, if systems are to use the information encoded in adjectival scales for generation or interpretation (e.g. for selecting an adjective with a particular degree of semantic strength, or for handling negation), they must have access to the sets of words comprising a scale. While linguists have presented various tests for accepting or rejecting a particular scalar relationship between any two adjectives (e.g., [2], [3]), the common problem with these methods is that they are designed to be applied by a human who incorporates the two adjectives in specific sentential frames (e.g. " X is warm, even hot") and assesses the semantic validity of the resulting sentences. Such tests cannot be used computationally to identify scales in a domain, since the specific sentences do not occur frequently enough in a corpus to produce an adequate description of the adjectival scales in the domain [4]. As scales vary across domains, the task of compiling such information is compounded. In this paper we describe a technique for automatically grouping adjectives according to their meaning based on a given text corpus, so that all adjectives placed in one group describe different values of the same property. Our method is based on statistical techniques, augmented with linguistic information derived from the corpus, and is completely domain independent. It demonstrates how high-level semantic knowledge can be computed from large amounts of low-level knowledge (essentially plain text, part-ofspeech rules, and optionally syntactic relations). While our current system does not distinguish between scalar and non-scalar adjectives, it is a first step in the automatic identification of adjectival scales, since the scales can be subsequently ordered and the non-scalar adjectives filtered on the basis of independent tests, done in part automatically and in part by hand in a post-editing phase. The result is a semi-automated system for the compilation of adjectival scales. In the following sections, we first describe our algorithm in detail, present the results obtained, and finally provide a formal evaluation of the results.