Long Sentence Partitioning using Structure Analysis for Machine Translation

in machine translation, long sentences are usually assumed to be difficult to treat. The main reason is the syntactic ambiguity which increases explosively as a sentence become longer. Especially, in the machine translation using sentence patterns, a long sentence causes a critical coverage problem. In this paper, we present a method of sentence partitioning which recognizes sub-sentence ranges by structure analysis, reducing the length of a sentence for translation. For the analysis of the clausal structure, phrase-level sentence patterns which have only a little syntactic ambiguities are employed. The structure analysis is conducted by the recognition of starting points of all clauses, dependency analysis, and depth analysis. Then, the ranges of sub-sentences are extracted based on the depth by stages. Our method was evaluated on 108 sentences extracted from CNN transcripts. It showed 85.2% accuracy in the detection of simple sentences.