Simultaneous Translation using Optimized Segmentation

Simultaneous Translation using Optimized Segmentation Maryam Siahbani maryam.siahbani@ufv.ca Department of Computer Information System, University of Fraser Valley, Abbotsford, V2S 7M7, Canada Hassan S. Shavarani sshavara@sfu.ca Ashkan Alinejad aalineja@sfu.ca Anoop Sarkar anoop@sfu.ca School of Computing Science, Simon Fraser University, Burnaby, V5A 1S6, Canada Abstract Previous simultaneous translation approaches either use a separate segmentation step followed by a machine translation decoder or rely on the decoder to segment and translate without training the segmenter to mini- mize delay or increase translation quality. We integrate a segmentation model and an incremental decoding algorithm to create an automatic simultaneous translation framework. Oda et al. (2014) propose a method to provide annotated data for sentence segmentation. This work uses this data to train a segmentation model that is integrated with a novel simultaneous translation decoding algorithm. We show that this approach is more accurate than previously proposed segmentation models when integrated with a translation decoder. Our results on the speech translation of TED talks from English to German show that our system can achieve translation quality close to the offline translation system while at the same time minimizing the delay in producing the translations incrementally. Our approach also outperforms other comparable simultaneous translation systems in terms of translation quality and latency. 1 Introduction In simultaneous translation the incoming speech stream is segmented and translated incrementally to re- duce the latency. There are two approaches for simultaneous translation task: sentence segmentation and incremental decoding, also called stream decoding. In incremental decoding, incoming words are fed into the decoder one-by-one, and the decoder updates its internal state. The decoder is responsible to decide when to begin the translation process and when to output the translation. Incremental decoding algorithms have been proposed for phrase-based (Kolss et al., 2008; Sankaran et al., 2010) translation, hierarchical phrase-based (Finch et al., 2015) and syntax-based (Oda et al., 2015) translation systems. Real-world speech translation systems estimate the sentence boundaries using punctuation insertion methods (Rangarajan Sridhar et al., 2013). As a result, recent work in simultaneous machine translation assume the input is already segmented into sentences, and focus on splitting the sentences into shorter sub- sequences of words (segments). This approach is called sentence segmentation. As soon as a segment is recognized, it is given to a decoder to generate and output the translation for that segment. Proceedings of AMTA 2018, vol. 1: MT Research Track Boston, March 17 - 21, 2018 | Page 154 Different methods have been proposed for sentence segmentation. Some use prosodic boundaries for segmentation (F¨ugen et al., 2007; Bangalore et al., 2012), while others use classification models. For exam- ple Rangarajan Sridhar et al. (2013) train a classifier to predict punctuation marks. The other approaches rely on the reordering probabilities of phrases to predict the segment boundaries (Fujita et al., 2013; Yarmo- hammadi et al., 2013; Siahbani et al., 2014). Oda et al. (2014) propose a method to provide annotated data for sentence segmentation which can be used in training a segmentation model. This method which later have …

Content truncated to 500 words