A Study of Punctuation Handling for Speech-to-speech Translation

This paper is devoted to finding out proper methods of handling punctuation when translating unpunctuated text output from automatic speech recognition systems. Three methods of handling punctuation and two machine translation (MT) systems were studied on a Japanese-to-English parallel corpus of about 5 million sentence pairs. BLEU calculated without punctuation was employed as measurement of translation quality in order to reflect the quality of synthesized speech output. Experimental results show that, for both phrase-based MT systems and pre-ordering phrase-based MT systems, methods of predicting punctuation with either a hidden n-gram model or a monolingual machine translation systems yield translation systems which are close in quality to the one using oracle punctuation, and better than ignoring punctuation.