Learning Source-Target Surface Patterns for Web-based Terminology Translation

摘要

This paper introduces a method for learning to find translation of a given source term on the Web. In the approach, the source term is used as a query and part of patterns to retrieve and extract translations in Web pages. The method involves using a bilingual term list to learn source-target surface patterns. At runtime, the given term is submitted to a search engine then the candidate translations are extracted from the returned summaries and subsequently ranked based on the surface patterns, occurrence counts, and transliteration knowledge. We present a prototype called TermMine that applies the method to translate terms. Evaluation on a set of encyclopedia terms shows that the method significantly outperforms the state-of-the-art online machine translation systems.