Translation Knowlegde Acquisition for Pattern-based Machine Translation
Mihoko Kitamura · Institutional Repositories DataBase (IRDB) · 2004
The quality of machine translation is strongly dependent on the quantity and quality of the translation knowledge available to the system.Constructing translation knowledge by hand has inherent limitations, which begs for techniques to construct translation knowledge automatically or semi-automatically, and to integrate this translation knowledge easily.This thesis deals with pattern-based machine translation and translation knowledge acquisition from parallel corpora, in order to fulfill the above demand.The first work advocates the use of complex patterns in machine translation.In previous pattern-based machine translation, writing new patterns was difficult due to the lack of flexibility .We have built a pattern-based machine translation system with an emphasis on pattern readability.Patterns can be constructed by hand or automatically from parallel corpora.The second work proposes a translation pattern extraction method that greedily extracts translation patterns based on co-occurrence of original and target word sequences in parallel corpora.This method can acquire translation patterns combining good coverage and accuracy, without any preliminary translation dictionary.The third work extends the second work by combining it with extra linguistic resources, such as chunking information and translation dictionaries.Additionally we allow manual confirmation of extracted translation patterns.Experimental results show both higher accuracy and coverage.The above proposal is a