Make Word Sense Disambiguation in EBMT Practical

Make Word Sense Disambiguation in EBMT Practical Feiliang REN Tianshun YAO School of Information Science & Engineering, Northeastern University, Shenyang 110004, China renfeiliang@ise.neu.edu.cn Abstract. In an EBMT system, we will meet the word sense disambiguation problem. The disambiguation methods used at present can’t be used easily in EBMT. We propose a new method for word sense disambiguation in EBMT: it is based on a language mode of the target language. Its main idea is that a proper word sense can make the whole sentence fluent. We use a language model to measure this fluency, and use dynamic programming method to compute the proper words sense sequence in EBMT. It has the virtues of easily being used, and doesn’t need extra lingual knowledge, besides, the general performance of it can be improved by using more target language resource to train. And experiment shows our method works well. Keywords: EBMT, word sense disambiguation, language model, N-gram, dynamic programming 1. Introduction Example based machine translation (EBMT[1]) is a method of translation by the principle of analogy. When given an input sentence, the EBMT system first retrieval the bilingual aligned corpora and select one or some example sentences as translation templates whose source sentence parts are similar to the given input sentences. And system then modifies the target parts of these translation templates to get finial translation result. These modification operations can be classified as: delete, replace, and insert. There are many methods for retrieving some translation templates whose source parts are similar to the input sentences. When we have finished this step and begin to the second step: modify the translation templates, the ambiguity problem comes. For example, for a Chinese-Japanese EBMT system, the input Chinese sentence is and what’s we retrieved translation template is: , we can see there is a Chinese word doesn’t exist in the source part of the translation template, to get the right translation of the input sentence, we must insert the sense of in a proper place of the target part of the translation template (the proper insert place can be gotten by comparing two string). If from a bilingual dictionary we know, there are different senses for the Chinese word , which sense should we choose? The same situation will occur also in replacement operation. This is the problem of word sense disambiguation EBMT. 1 2 3 4 5... m S c c c c c c = 1 2 4 5 1 2 3 4 ... ... n c c c c c j j j j j ↔ l 3c 3c k 3c 414 Traditional word sense disambiguation methods don’t work here. They will neither meet a knowledge bottleneck [3-5], nor lower performance [2]. And it can’t satisfy the requirement in an EBMT system. Here we propose a new word sense disambiguation method, which is based on the N-gram language model. We think a proper word sense must be the sense that makes the whole sentence looks frequent, and we use N-gram language model[6,7] …

Content truncated to 500 words