The Microsoft Speech Language Translation (MSLT) Corpus for Chinese and Japanese: Conversational Test data for Machine Translation and Speech Recognition

The Microsoft Speech Language Translation (MSLT) Corpus for Chinese and Japanese: Conversational Test data for Machine Translation and Speech Recognition Christian Federmann chrife@microsoft.com William D. Lewis wilewis@microsoft.com Microsoft Translator Microsoft AI+Research, Redmond, WA, USA Abstract Recent years have seen unprecedented growth in the use of MT across industries and domains. Partly this is due to the ready availability of open source MT tools such as Moses or online or customizable services. It is also due to fundamental shifts in the technology, specifically the move to deep learning, which has dramatically improved the quality of MT engines, including those used by online services. Likewise, improvements in Speech Recognition (SR) technology, also driven by the move to deep learning, are showing significant improvements in quality driven by deep learning alone. The improvements of both of these technologies, MT and SR, increase the potential viability for speech translation, since the error cascade caused by daisy- chaining these technologies drops as the quality bar raises. MT is a crucial component in speech translation systems, yet developing conversational MT systems essential to speech translation is not a focus for many working in the Machine Translation discipline. Particularly problematic for many languages is the absence of test and dev data, not any less true for the Chinese and Japanese languages, where forays into conversational MT in and out of these languages are limited by the lack of publicly available conversational test data. In this paper, we seek to address this problem, by providing MT test and dev data that has been built from actual bilingual conversations between English and Japanese and Chinese, test data that can be useful to drive further research in this space for these two languages. Our plan is to make the data described in this paper available to the public by MT Summit. 1 Introduction The commoditization of MT, as evidenced by increased use of MT across industries and domains, results most significantly by the availability of open source MT tools such as Moses [1] and online tools for training and building customized systems (such as those offered by Microsoft, SDL, IBM, etc.). But it is also due to fundamental Proceedings of MT Summit XVI, vol.1: Research Track Nagoya, Sep. 18-22, 2017 | p. 72 shifts in the technology, specifically the move to deep learning, which has dramatically improved the quality of MT engines [2, 3, 4], including those used by online services (e.g., Google, Microsoft, Baidu, etc.). Likewise, improvements in speech recognition technology, also driven by the move to deep learning, are showing 25-50% improvements in quality driven by deep learning alone; for instance, [5] showed a 32% reduction in Word Error Rate when switching from Gaussian Mixture Models (GMMs) to Deep Neural Networks (DNNs), with no change in training data. The improvements of both of these technologies increase the potential viability for speech translation, since the error cascade caused by daisy-chaining these technologies drops as the quality bar in each component technology increases. MT is a crucial component in speech …

Content truncated to 500 words