Learning to Parse Bilingual Sentences Using Bilingual Corpus and Monolingual CFG

Abstract We present a new method for learning to parse a bilingual sentence using Inversion Transduction Grammar trained on a parallel corpus and a monolingual treebank. The method produces a parse tree for a bilingual sentence, showing the shared syntactic structures of individual sentence and the differences of word order within a syntactic structure. The method involves estimating lexical translation probability based on a word-aligning strategy and inferring probabilities for CFG rules. At runtime, a bottom-up CYK-styled parser is employed to construct the most probable bilingual parse tree for any given sentence pair. We also describe an implementation of the proposed method. The experimental results indicate the proposed model produces word alignments better than those produced by Giza++, a state-of-the-art word alignment system, in terms of alignment error rate and F-measure. The bilingual parse trees produced for the parallel corpus can be exploited to extract bilingual phrases and train a decoder for statistical machine translation.