Word classification based on combined measures of distributional and semantic similarity

Word classification based on combined measures of distributional and semantic similarity Viktor Pekar Bashkir State University, 450000 Ufa, Russia vpekar@ufanet.ru Steffen Staab Institute AIFB, University of Karlsruhe http://www.aifb.uni-karlsruhe.de/WBS & Learning Lab Lower Saxony http://www.learninglab.de Abstract The paper addresses the problem of automatic enrichment of a thesaurus by classifying new words into its classes. The proposed classification method makes use of both the distributional data about a new word and the strength of the semantic relatedness of its target class to other likely candidate classes. 1 Introduction Today, many NLP applications make active use of thesauri like WordNet, which serve as background lexical knowledge for processing the semantics of words and documents. However, maintaining a thesaurus so that it sufficiently covers the lexicon of novel text data requires a lot of time and effort, which may be prohibitive in many settings. One possibility to (semi-) automatically enrich a thesaurus with new items is to exploit the distributional hypothesis. Ac- cording to this approach, the meaning of a new word is first represented as the totality of textual contexts where it occurs and then assigned to that semantic class which members exhibit similar occurrence patterns. The distributional approach was shown to be quite effective for tasks where new words need to be as- signed to a limited number of classes (up to 5; e.g., Riloff and Shepherd, 1997; Roark and Charniak, 1998). However, its application to numerous classes, as would be the case with a thesaurus of a realistic size, proves to be much more challenging. For ex- ample, Alfonseca and Manandhar (2002) attain the learning accuracy' of 38% when assigning new words to 46 WordNet concepts. In the present paper we propose a method that is particularly effective for the task of classifying words into numerous classes forming a hierarchy. The po- sition of a class inside the hierarchy reflects the de- gree of its semantic similarity to other classes. Besides distributional data, our method integrates this semantic information: the classification decision is a function of both (1) the distributional similarity of the new word to the target class and (2) the strength of the semantic relatedness of the target class to other likely candidates. Thus, using the thesaurus as background knowledge we aim to make up for pos- sible insufficient quality of the distributional data. 2 Similarity measures We evaluate our approach on the task where nouns are classified into a predefmed set of semantic classes. Thereby, the meaning of each noun n is rep- resented as a distributional feature vector, where features are verbs vc V linked to the noun by predi- cate-object relations. The values of the features are conditional probabilities P(vIn) estimated from the frequencies observed in the corpus. To measure the similarity between vectors of nouns n and m, we used the L 1 distance metric 2 : Li (n ,� = El P(vI n)— P(v I m) I� (1) ecV To assign a noun to a class, we use the k near- est neighbors algorithm …

Content truncated to 500 words