Acquiring Semantic Transfer Rules from Bilingual Corpora
Speaker: Michael Jellinghaus
Abstract:
The work I am going to present is based on the observation that the two prevailing paradigms in Machine Translation (MT) exhibit complementary strengths and weaknesses: Statistical MT (SMT) systems are robust and quick to develop but often fail to deliver correct translations for syntactically or semantically more complex input. Rule-based MT (RBMT) systems, on the other hand, are capable of covering linguistically difficult phenomena but generally suffer from decreased robustness and the development bottleneck of hand-crafting transfer rules.
In this talk, I will present a hybrid approach combining 'deep' RBMT based on semantic transfer with SMT-like data-driven rule acquisition, thus integrating both worlds' major advantages. The method involves parsing a sentence-aligned bilingual corpus with Head-Driven Prase Structure Grammars (HPSG) and the extraction of transfer rules by aligning the resulting semantic representations.
