Automatic Sense Disambiguation for Target Word Selection

Automatic Sense Disambiguation for Target Word Selection Kwon Yang Kim� Se Young Park� Sang Jo Lee Dept.of Computer Science� Computer Dept.� Kyungpook Kyungpook Sanup University� ETRI� National University Hayang, Korea� Taejon, Korea�Taegu, Korea Abstract This paper describes a method of automatic sense disambiguation for target word selection in Korean to English machine translation. At first, we define the concept of cluster for each sense of given verb according to corresponding target word. And then, we propose a method which selects the sense combination of words as the correct sense that has the greatest number of overlaps between input case slots and the predefined clusters for the given verb. 1 Introduction The problem of word sense disambiguation is one which has received increased at- tention in recent work on natural language processing applications such as machine translation and information retrival. Given an occurrence of a polysemous word in text, it is needed to examine a set of senses in the dictionary and detect the clues for deciding intended sense in the context. Recently, several researches have been experimented with target word selection based on example-based machine translation[1,2,3}. There is no failure of selecting target word because the method is to select most similar one in a thesaurus than exact correspondent. However, the similarity methods based on thesaurus have some problem that it is difficult to make good thesaurus which provides accurate distance measure between words. In order to solve the imperfection problems of disambiguation methods based on similarity, some researches using the exact matching with collocation extracted from bilingual corpus were experimented by [4,5]. But, these methods have a drawback that they can not select a target word when input sentences have nouns which were not in- cluded in collocation list because they need exact matching of collocation. And it is not clear how large the bilingual corpus would have to be become authoritative for disam- biguation as an application independent collocation. The well-known attempts to utilize information in MRD(Machine Readable Diction- aries) for lexical disambiguation are that of [6,7,8] which select the correct sense of a word by counting the overlaps between a dictionary sense definition and the definition of the nearby in the phrase. The sense of a word with greatest number of overlaps with senses of other words in the sentence is chosen as the correct one. These methods based on a small number of overlap between sense definitions is weak as a clue for disam- biguation. As a result, these methods could not have certainty about disambiguation, because the relation is wholly dependent on word usage in the sense definition of a par- ticular dictionary. Another variations of this idea are based on co-occurrence statistics, which is meant the preference two words appear together in the same context [9,10,11] . But, their method has same problems of MRD approaches. Our method is based on the idea that it selects the sense combinations of words as 109 the correct sense with the greatest number of overlaps between input …

Content truncated to 500 words