Speech-Input Multi-Target Machine Translation

Proceedings of the Second Workshop on Statistical Machine Translation, pages 56–63, Prague, June 2007. c⃝2007 Association for Computational Linguistics Speech-input multi-target machine translation Alicia P´erez, M. In´es Torres Dep. of Electricity and Electronics University of the Basque Country manes@we.lc.ehu.es M. Teresa Gonz´alez, Francisco Casacuberta Dep. of Information Systems and Computation Technical University of Valencia fcn@dsic.upv.es Abstract In order to simultaneously translate speech into multiple languages an extension of stochastic finite-state transducers is pro- posed. In this approach the speech trans- lation model consists of a single network where acoustic models (in the input) and the multilingual model (in the output) are em- bedded. The multi-target model has been evaluated in a practical situation, and the results have been compared with those obtained using several mono-target models. Experimental results show that the multi-target one re- quires less amount of memory. In addition, a single decoding is enough to get the speech translated into multiple languages. 1 Introduction In this work we deal with finite-state models which constitute an important framework in syntactic pat- tern recognition for language and speech processing applications (Mohri et al., 2002; Pereira and Riley, 1997). One of their outstanding characteristics is the availability of efficient algorithms for both optimiza- tion and decoding purposes. Specifically, stochastic finite-state transducers (SFSTs) have proved to be useful for machine trans- lation tasks within restricted domains. There are several approaches implemented over SFSTs which range from word-based systems (Knight and Al- Onaizan, 1998) to phrase-based systems (P´erez et al., 2007). SFSTs usually offer high speed during the decoding step and they provide competitive re- sults in terms of error rates. In addition, SFSTs have proved to be versatile models, which can be easily integrated with other finite-state models, such as a speech recognition system for speech-input transla- tion purposes (Vidal, 1997). In fact, the integrated architecture has proved to work better than the de- coupled one. Our main goal is, hence, to extend and assess these methodologies to accomplish spo- ken language multi-target translation. As far as multilingual translation is concerned, there are two main trends in machine translation de- voted to translate an input string simultaneously into m languages (Hutchins and Somers, 1992): inter- lingua and parallel transfer. The former has his- torically been a knowledge-based technique that re- quires a deep-analysis effort, and the latter consists on m decoupled translators in a parallel architec- ture. These translators can be either knowledge or example-based. On the other hand, in (Gonz´alez and Casacuberta, 2006) an example based technique consisting of a single SFST that cope with multiple target languages was presented. In that approach, when translating an input sentence, only one search through the multi-target SFST is required, instead of the m independent decoding processes required by the mono-target translators. The classical layout for speech-input multi-target translation includes a speech recognition system in a serial architecture with m decoupled text-to-text translators. Thus, this architecture entails a decod- ing stage of the speech signal into the source lan- guage text, and m further …

Content truncated to 500 words