Learning the Optimal use of Dependency-parsing Information for Finding Translations with Comparable Corpora

摘要

Using comparable corpora to find new word translations is a promising approach for extending bilingual dictionaries (semi-) automatically. The basic idea is based on the assumption that similar words have similar contexts across languages. The context of a word is often summarized by using the bag-of-words in the sentence, or by using the words which are in a certain dependency position, e.g. the predecessors and successors. These different context positions are then combined into one context vector and compared across languages. However, previous research makes the (implicit) assumption that these different context positions should be weighted as equally important. Furthermore, only the same context positions are compared with each other, for example the successor position in Spanish is compared with the successor position in English. However, this is not necessarily always appropriate for languages like Japanese and English. To overcome these limitations, we suggest to perform a linear transformation of the context vectors, which is defined by a matrix. We define the optimal transformation matrix by using a Bayesian probabilistic model, and show that it is feasible to find an approximate solution using Markov chain Monte Carlo methods. Our experiments demonstrate that our proposed method constantly improves translation accuracy.