Statistical analysis of target language corpus for word sense disambiguation in a machine translation system

129 Statistical analysis of target language corpus for word sense disambiguation in a machine translation system Tayebeh Mosavi miangah1 and Ali Delavar khalafi2 1English Language Department, Shahre Kord University, Shahre Kord, Iran. E-mail: mousavi-t @lit.sku.ac.ir 2Mathematics Department, Shahre Kord University, Shahre Kord, Iran. E-mail: delavar@sci.sku.ac.ir Abstract This article studies different aspects of a new approach for word sense disambiguation using statistical information gains from a target language monolingual corpus. Here, the source language is English and the target one is Persian, and this disambiguation method in those aspects which gives desirable results can be directly applied in the system of English-to-Persian machine translation for solving lexical ambiguity problems in this system. Unlike the other disambiguation programs using corpora for handling the problem, which use probabilistic Model in their statistical works, this paper uses Simulation Model. We believe that this model is more reasonable from the scientific point of view with the most precise and accurate results. This method has been tested for a selected set of English texts having some multiple-meaning words in respect to Persian language and the results are encouraging. 1. Introduction Ambiguity is the most considerable problem in natural language processing systems, among which machine translation systems suffer this problem in a high degree. The problem of ambiguity in translating texts by the machine is different from one by the human. Human is a complex machine so that he can choose the suitable target equivalent(s) of any source language forms, sometimes without becoming aware of the irrelevant alternatives, based on his understanding in the context. He can also automatically consider a group of the words, rather than individual word to understand the meaning of a sentence, even if the words of the group are not relevant. But a machine cannot think at all. Since only written texts are presented to the computer, it can not mechanically use the relevant text. However, this problem has been somehow solved by the field of discourse analysis which is not within the scope of this study. Nowadays the application of statistical approaches and studying statistical-based methods in natural language processing as well as in machine translation has been rapidly increasing. Statistical linguistics basically relies on the studying of various linguistic units occurrences frequencies including word-forms, lexemes, morphemes, letters, etc. in a sample corpus, and solving various linguistic problems as ambiguity with reference to these certain frequencies and calculating the probability of them. Statistics- based approaches factor out the need for computational mechanisms and high linguistic knowledge for solving linguistic problems. So computational cost of a statistics-based approach is much lower than a knowledge-based or a rule- based approach (Su and Chang, 1990). A statistics- based system needs a large database or corpus to guarantee its reliability, however, with availability of many tagged and untagged text corpora, acquiring linguistic knowledge from a large sample corpus is no longer an impossible task (Garside et al, 1987) . Lexical ambiguity refers to a case in which either a lexical unit belongs to different lexical …

Content truncated to 500 words