Incremental Segmentation and Decoding Strategies for Simultaneous Translation

International Joint Conference on Natural Language Processing, pages 1032–1036, Nagoya, Japan, 14-18 October 2013. Incremental Segmentation and Decoding Strategies for Simultaneous Translation Mahsa Yarmohammadi†, Vivek K. Rangarajan Sridhar◦, Srinivas Bangalore◦, Baskaran Sankaran‡ †Center for Spoken Language Understanding, Oregon Health & Science University ◦AT&T Labs - Research ‡School of Computing Science, Simon Fraser University {yarmoham}@ohsu.edu, {vkumar,srini}@research.att.com, {baskaran}@cs.sfu.ca Abstract Simultaneous translation is the challeng- ing task of listening to source language speech, and at the same time, produc- ing target language speech. Human inter- preters achieve this task routinely and ef- fortlessly, using different strategies in or- der to minimize the latency in produc- ing target language. Toward modeling the human interpretation process, we propose a novel input segmentation method using the phrase alignment structure of the lan- guage pair. We compare and contrast three incremental decoding and two different in- put segmentation strategies, including our proposed method, for simultaneous trans- lation. We present accuracy and latency tradeoffs for each of the decoding strate- gies when applied to audio lectures from the TED collection. 1 Introduction In simultaneous speech translation, it is important to keep the delay between a source language chunk and its corresponding target language chunk (re- ferred to as ear-voice span) minimal in order to continually engage the listeners. Simultaneous hu- man interpreters are able to generate target speech incrementally with very low ear-voice span by us- ing a variety of strategies (Chernov, 2004) such as anticipation, cognitive and linguistic inference, paraphrasing, etc. However, current methodolo- gies for simultaneous translation are far from be- ing able to exploit or model such complex phe- nomena. Quite often, models trained for consec- utive translation are repurposed for incremental translation. One of the first attempts at incremental text translation was presented by Furuse and Iida (1996) using a transfer-based MT approach and more recently by Sankaran et al. (2010) us- ing a phrase-based approach. On the other hand, incremental speech translation has been ad- dressed in simultaneous translation of lectures and speeches (Hamon et al., 2009; F¨ugen et al., 2007). Some previous work (Cettolo and Federico, 2006; Rao et al., 2007; Matusov et al., 2007) addressed source text (reference or ASR hypothesis) seg- mentation strategies in speech translation. Con- straining the search process during decoding to be monotonic (Tillmann and Ney, 2000) is one way of reducing latency and promoting incrementality. However, finding the optimal segmentation of the complete source string using dynamic program- ming is still slow. By shifting the focus of the task to appropriate segmentation of incoming text, consecutive trans- lation models have been used with good success to simulate incremental translation, such as incre- mental speech-to-speech translation (Bangalore et al., 2012) which focuses on translating the par- tial hypotheses generated based on the silences detected by a speech recognizer. However, stud- ies on human interpreters show that in only a few cases the interpreters encode the chunks of speech as uttered in the source: the mean proportion of silence-based chunking by interpreters is 6.6% when the source is …

Content truncated to 500 words