Estimating Point-of-View-based Similarity Using POV Reinforcement and Similarity Propagation

Language, Information and Computation(PACLIC 11), 1996, 373-382 Estimating Point-of-View-based Similarity using POV Reinforcement and Similarity Propagation Kenji Nagamatsu� Hidehiko Tanaka The University of Tokyo {naga,tanaka}emtl.t .u-tokyo.ac . jp Abstract This paper . proposes a similarity measure which takes account of point- of-views (abbreviated to POV, hereafter) in the calculation of similarity values. So far many researches on similarity measures have been per- formed but none takes account of POVs. The similarity measure pro- posed in this paper is based on co-occurrence probabilities of words and this makes it possible to obtain preferable precision even if POVs are not given. This method consists of two parts of processes, POV reinforce- ment and similarity propagation. First, the POV reinforcement process, which affects the similarity between words, modifies the weights of links according to the relatedness between the link and the POV word. Sec- ond, the similarity propagation process propagates the weights of links and defines a similarity value for word pairs which do not actually co- occur in the corpus. Using those two processes this method becomes capable both to take POVs into consideration and to cope with the sparseness of corpora to some degree. This paper, however, focuses on the POV reinforcement and evaluates the effectiveness of the method.. 1 Introduction Rapid growth of computer networks has increased the number of machine-readable texts and also made it possible for us to use various search engines to get desired documents. They are, however, keyword-based and strict. Even if some docu- ments are related to a user's interest, he or she cannot obtain them as long as they do not contain the keywords given in advance. Otherwise he or she will be handed too many documents which contain just the keywords but are not necessarily related to his or her interest. To solve these problems, it is necessary to take account of similarity between words or concepts and to make use of the measure in search processing. However, because there are many similar words in a text, employing a similarity measure alone only will expand the range of relatedness and produce more and more results. When considering meanings of words, we, human beings, do not consider the whole meanings at a time, rather some interesting aspects of their concepts just the same as we look at a landscape from some point-of-view. Hence in similarity of words it is required to take account of their POVs. This makes it possible both to expand the range of matching in some situations and to restrict the range in others. Expectation is that employing valid POVs has both the expansion and the restriction be suitable and search processing produces more appropriate results. This paper proposes a similarity measure between words which measure takes account of the effect of POVs. So far many researches on concept (or word) sim- ilarity have been performed but none handles POVs in their similarity measures. The proposed method utilizes co-occurrence probability-based similarity as a basis 373 and extends this fundamental measure by …

Content truncated to 500 words