Automatic Acquisition of Semantic Transfer Rules for Machine Translation
Speaker:Michael Jellinghaus
Institution:Saarland University
Abstract:
'Deep' transfer-based approaches to machine translation (MT) have the potential to excel in the production of output which is both grammatical ('fluency' criterion) and semantically close to the meaning of the input ('adequacy' criterion). However, systems of this kind suffer from a development bottleneck in that the necessary transfer rules have to be hand-crafted laboriously. By contrast, data-driven approaches like statistical MT (SMT) exhibit the advantage of a rather short development cycle achieved by learning translation information directly from bilingual corpora.
In this talk, I will present a method of how the automatic acquisition of semantic transfer rules from sentence-aligned bilingual corpora can integrate the advantages of data-driven techniques into MT systems based on semantic transfer.
This method involves creating a parallel corpus of semantic structures in Minimal Recursion Semantics (MRS) format by parsing both sides of a text corpus with Head-Driven Phrase Structure Grammars (HPSG). An algorithm then semantically aligns MRS substructures of the source language sentence with MRS substructures of the target language sentence. Finally, by making use of the correspondence information obtained in this manner, a set of semantic transfer rules can be extracted.