Building a Paraphrase Corpus for Speech Translation
Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC 2004)
Abstract
When a machine translation (MT) system receives input sentences of spoken language, the following two types of sentences are difficult to translate: (1) long sentences and (2) sentences having redundant expressions often seen in spoken language. To reduce these difficulties, we are developing methods to paraphrase input sentences into more translatable ones. In this paper, we report a preliminary Japanese paraphrase corpus. The corpus consists of original sentences derived from travel conversation and versions of them paraphrased by humans. We use three paraphrasing methods: plain, segment, and summary paraphrasing. Plain paraphrasing is applied to short sentences, where redundant expressions are replaced with plain ones. Segment and summary paraphrasing is applied to long sentences, where long sentences are converted into one or several short sentences. We also report a comparison of machine translation quality between the original sentences and the paraphrased sentences. We use two corpus-based machine translation systems in the experiment.