Attribute-Based and Value-Based Clustering: An Evaluation

Attribute-Based and Value-Based Clustering: An Evaluation Abdulrahman ALMUHAREB and Massimo POESIO Department of Computer Science and Centre for Cognitive Science University of Essex Colchester, United Kingdom, CO4 3SQ aalmuh@essex.ac.uk poesio@essex.ac.uk Abstract In most research on concept acquisition from corpora, concepts are modeled as vectors of relations extracted from syntactic structures. In the case of modifiers, these relations often specify values of attributes, as in (attr red); this is unlike what typically proposed in theories of knowledge representation, where concepts are typically defined in terms of their attributes (e.g., color). We compared models of concepts based on values with models based on attributes, using lexical clustering as the basis for comparison. We find that attribute-based models work better than value-based ones, and result in shorter descriptions; but that mixed models including both the best attributes and the best values work best of all. 1 Introduction In most recent research on concept acquisition from corpora (e.g., for lexicon construction), concepts are viewed as vectors of relations, or properties, extracted from syntactic structures (Grefenstette, 1993; Lin, 1998; Curran and Moens, 2002; Kilgarriff, 2003, and many others). These properties often specify values of attributes such as color, shape, or size: for example, the vector used by Lin (1998) for the concept dog includes the property (dog adj-mod brown). (We will use the term values here to refer to any modifier.) To our knowledge, however, no attempt has been made by computational linguists to use the attributes themselves in such vectors: i.e., to learn that the description of the concept dog includes elements such as (dog color) or (dog size). This is surprising when considering that most models of concepts in the AI literature are based on such attributes (Brachman and Levesque, 1985). Two problems need to be addressed when trying to identify concept attributes. The first problem is that values are easier to extract. We found, however, that patterns like the X of the dog, already used in (Berland and Charniak, 1999; Poesio et al, 2002) to find part-of relations (using techniques derived from those used in (Hearst, 1998; Caraballo, 1999) to find hyponymy relations) are quite effective at finding attributes. A second problem might be that instances of such patterns are less frequent than those used to extract values, even in large corpora such as the British National Corpus (BNC). But this problem, as well, is less serious when using the Web as a corpus (Kilgarriff and Schuetze, 2003; Keller and Lapata, 2003; Markert et al, submitted). We report on two experiments whose goal was to test whether identifying attributes leads to better lexical descriptions of concepts. We do this by comparing the results obtained by using attributes or more general modifiers – that we will simply call values – as elements of concept vectors used to identify concept similarities via clustering. In Section 2, we discuss how Web data were used to build attribute- and value- based concept vectors, and our clustering and evaluation methods. In Section 3, we discuss a first …

Content truncated to 500 words