Better Translation with Knowledge Extracted from Source Text
Satoshi Kinoshita, Miwako Shimazu, Hideki Hirakawa · 2005
This paper presents a framework in which a source text is translated using not only given knowledge but knowledge which is extracted from the text. First, co-occurrence relations between words are extracted prior to translation, and then are used in translation to resolve ambiguities in the source text. The important feature of the system is that extracted relations are categorized as either reliable or semi-reliable according to ambiguities in the analysis. Reliable relations are highly likely to be correct and therefore can be used for translating other texts. By contrast, semi-reliable ones may contain some errors and therefore should be restricted to texts from which the relations have been extracted. However, semi-reliable relations prove valuable when the system has limited knowledge; in our experiment the accuracy of disambiguating verb-noun dependencies has improved from 91.5% to 93.0% with a use of both reliable and semi-reliable knowledge, while with only reliable knowledge the accuracy recorded 92.3%. The effectiveness was furthermore examined when the system was provided with manually collected co-occurrence relations which are essentially equivalent to a complete set of co-occurrence relations in the domain.