Word-Sense Classification by Hierarchical Clustering

Language, Information and Computation (PACLIC12), 18-20 Feb, 1998, 236-247 Word-Sense Classification by Hierarchial Clustering Ken Y .K. Lau and Robert W.P. Luk Department of Computing, Hong Kong Polytechnic University Email: csrluk@comp.polyu.edu.hk Abstract This paper investigates the use of clustering techniques in word-sense classification, which identifies different contexts that a word was used with the same or similar sense. For simplicity, we have used the hierarchical clustering techniques: single- and complete-linkage, and we showed that the latter is a more suitable technique from our performance measurements (i. e. recall and precision) compared with manually grouping different contexts of similar meaning. We found that the use of part-of-speech tags and fixed-length context has better clustering performance than without part-of-speech tags and sentence context, respectively. The differences between manually identified groups of different contexts are measured in terms of recall and precision at about 80%, which are not very different from the average recall and precision performance of complete-linkage clustering at 80% and 75%, respectively. 1. Introduction Lexicographers can access, collect and analyze a large volume of language data by computer. The processed data are often put in the form of concordances as show in Figure 1. Lexicographers use these concordances to identify the different word senses and usually representative word senses are registered in the dictionary or glossary. However, identification of different word senses is a formidable task of an objective lexicographer because (s)he has to browse through f(i) number of concordances if the word i has occurrence frequency f(i). Since at most there are f(i) different senses, the lexicographer has to perform at most fa/ amount of comparison between concordances. Although computers cannot write down the senses of each word in the concordance, computers can assist the lexicographer by identifying the different groups of concordances for which the word has similar meaning. This would save the lexicographer from browsing through a large list of concordances and the lexicographer can select examples from groups of concordances of similar meaning, and register and write down those representative word senses. tgigfagag , 4;1* n•� [ ] IT-34%RTEt• °� 1 2 Mtratt5tttiga4. 3 ATtult [ l l L IN-Fftittt ° eiattx 4 3 0 � 5:83..Mt� [ 1 JZaJdgilit_EttrlilifIsAtaCz ° � [ I ]� EtEn1111'gE3).Z.—±.1-PJArITO &Mgr °� [ I ERYPtiltitXiMaZ_E*Vriff Figure 1: A list of different concordances found in the PH corpus [1] for the character/word R. Apart from assisting lexicographers, automatically grouping different concordances of the word with similar meaning has found applications in document retrieval [2] since assigning words with different sense tags differentiates their meaning and therefore improve the precision performance of retrieval systems. There are already many work reported in grouping different words based on their concordances for the automatic construction of thesauri [3,4]. However, there are relatively few work reported in grouping concordances with similar meaning, particularly for Chinese. Yarowsky [5] was advocating one sense per collocation (or concordance) so that word senses registered are only a representative or generic meaning of the words extracted from their different occurrences. Schutze …

Content truncated to 500 words