Automatic Learning of Parallel Dependency Treelet Pairs

Induction of synchronous grammars from empirical data has long been a problem unsolved; despite that generative synchronous grammars theoretically suit the machine translation task very well. This fact is mainly due to pervasive structural divergences between languages. This paper presents a statistical approach to learn dependency structure mappings from parallel corpora. The algorithm introduced in this paper extends the dependency tree word alignment algorithm in (Ding, 2003). The new algorithm automatically learns parallel dependency treelet pairs from loosely matched non-isomorphic dependency trees while keeping computational complexity polynomial in the length of the sentences. A set of heuristics are introduced and specifically optimized for the parallel treelet learning purpose using Minimum Error Rate training. As learning parallel syntactic structures is the key step in the automatic learning of a synchronous grammar, the learnt parallel dependency treelet pairs by our approach serve as an important first step of any lexicalized synchronous grammar induction.