ASR Normalization for Machine Translation

In natural spoken language there are many meaningless modal particles and dittographes, furthermore ASR (automatic speech recognition) often has some recognition errors and the ASR results have no punctuations. Therefore, the translation would be rather poor if the ASR results are directly translated by MT (machine translation). In this paper, an ASR normalization approach was introduced for machine translation which based on maximum entropy sequential labeling model. Before translation, the meaningless modal particles and dittograph were deleted, and the recognition errors were corrected, and ASR results were also punctuated. Experiments show that the MT BLEU of 0.2465 is obtained, that improved by 17.3% over the MT baseline without normalization. The positive experimental results confirm that ASR normalization is effective for improvement of translation quality for spoken language machine translation.

I. Introduction

Compared to traditional text machine translation whose input text has normative style of writing, spoken language machine translation is more difficult. The main challenges are: (1) there are many meaningless modal particles and dittograph in natural spoken language; (2) ASR results have no punctuations; (3) even the state-of-art ASR often has some recognition errors, etc. al. These all deteriorate badly the effect of the machine translation. The following examples extracted from a practical bilingual phone question answering system demonstrate the depressed phenomena.