Semantic Clustering of Adjectives and Verbs Using their Syntactic Patterns

The idea that the syntactic behaviour of words is connected with their meaning has been the assumption behind research in elds such as lexical semantics and automatic clustering of words based on statistical methods. In particular much work has been done to describe the relation between the semantic characteristics of verbs and their syntactic patterns, among many Fillmore (1970) and Levin (1993), and to identify semantically similar words from large text corpora on the basis of linguistic and distributional properties, i.a. Brown et al. (1992), Pereira et al. (1993). Some research has also been done to extract the semantic meaning of adjectives on the basis of their co-occurrence with nouns We follow the assumption that there is a connection between the syntactic behaviour and the meaning of words. Although we agree with Levin (1993) that "verb meaning is a key to verb behaviour", in this paper we go the other way round: from the syntactic behaviour of words we derive some of their semantic characteristics. We have investigated to what extent it is possible to use the syntactic encodings of a corpus-based NLP lexicon to extract clusters of semantic related words. In particular we have extracted adjectives and verbs with the same syntactic pattern in a corpus-based Danish NLP lexicon, the LE-PAROLE lexicon, and evaluated the obtained clusters according to the "similarity" of the elements in each group. We deene "similarity" of meaning for adjectives by semantic parameters used in the literature to identify semantically related adjectives. Adjectives have "similar" meaning if they are synonymous or antonymous (Miller et al., 1993 1990) and if they belong to a linguistic scale. Linguistic scales, according to the deenition provided by Levinson (1983))133], are "sets of linguistic alternates, or contrastive expressions of the same grammatical category, which can be arranged in a linear order by degree of informativeness or semantic strength". When adjectives belong to a scale, independently of their orientation, we relate them to a common "superordinate" concept. Also verbal linguistic scales exist, but they are not so frequent as adjectival scales, and only few verbs have "real" opposites. We adopt the troponymy relation as deened in (Miller et al., 1993 1990) where verbs are said to be troponyms if they are connected to a superordinate along more semantic dimensions. Thus we consider verbs to be "similar" if they belong to a 1