Hierarchical Phrase-based Stream Decoding

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1089–1094, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Hierarchical Phrase-based Stream Decoding Andrew Finch and Xiaolin Wang and Masao Utiyama and Eiichiro Sumita Advanced Speech Translation Research and Development Promotion Center Advanced Translation Technology Laboratory National Institute of Information and Communications Technology Kyoto, Japan {andrew.finch,xiaolin.wang,mutiyama,eiichiro.sumita}@nict.go.jp Abstract This paper proposes a method for hierar- chical phrase-based stream decoding. A stream decoder is able to take a contin- uous stream of tokens as input, and seg- ments this stream into word sequences that are translated and output as a stream of tar- get word sequences. Phrase-based stream decoding techniques have been shown to be effective as a means of simultaneous in- terpretation. In this paper we transfer the essence of this idea into the framework of hierarchical machine translation. The hi- erarchical decoding framework organizes the decoding process into a chart; this structure is naturally suited to the process of stream decoding, leading to an efficient stream decoding algorithm that searches a restricted subspace containing only rel- evant hypotheses. Furthermore, the de- coder allows more explicit access to the word re-ordering process that is of crit- ical importance in decoding while inter- preting. The decoder was evaluated on TED talk data for English-Spanish and English-Chinese. Our results show that like the phrase-based stream decoder, the hierarchical is capable of approaching the performance of the underlying hierarchi- cal phrase-based machine translation de- coder, at useful levels of latency. In ad- dition the hierarchical approach appeared to be robust to the difficulties presented by the more challenging English-Chinese task. 1 Introduction Statistical machine translation traditionally oper- ates on sentence segmented input. This technol- ogy has advanced to the point where it is becom- ing capable enough to be useful for many applica- tions. However, this approach may be unsuitable for simultaneous interpretation where the machine translation system is required to provide transla- tions within a reasonably short space of time after words have been spoken. Under this type of con- straint, it may not be possible to wait for the end of the sentence before translating, and segmenta- tion at the sub-sentential level may be required as a consequence. This segmentation process is difficult, even for skilled human interpreters, and presents a major challenge to a machine since in addition to the translation process, decisions need to be made about when to commit to outputting a partial translation. Such decisions are critical since once such an output is made it can be dif- ficult and highly undesirable to correct it later if it is in error. 2 Related Work In order to automatically perform segmentation for interpretation, two types of strategy have be proposed. In the first, which we will call pre- segmentation, the stream is segmented prior to the start of the machine translation decoding process, and the machine translation system is constrained to translate using the given segmentation. This approach has the advantage that it can …

Content truncated to 500 words