Neural Baselines for Word Alignment

The IWSLT 2019 Evaluation Campaign J. Niehues(1) R. Cattoni(2) S. St¨uker(3) M. Negri(2) M. Turchi(2) T. Ha(3) E. Salesky(4) R. Sanabria(5) L. Barrault(6) L. Specia(6) M. Federico(2) (1) DKE, Maastricht University, Netherlands (2) FBK - Via Sommarive 18, 38123 Trento, Italy (3) KIT - Adenauerring 2, 76131 Karlsruhe, Germany (4) Center for Language & Speech Processing, Johns Hopkins University, USA (5) School of Informatics, University of Edinburgh, U.K. (6) Department of Computing, Imperial College London, U.K. Abstract The IWSLT 2019 evaluation campaign featured three tasks: speech translation of (i) TED talks and (ii) How2 instruc- tional videos from English into German and Portuguese, and (iii) text translation of TED talks from English into Czech. For the first two tasks we encouraged submissions of end- to-end speech-to-text systems, and for the second task par- ticipants could also use the video as additional input. We received submissions by 12 research teams. This overview provides detailed descriptions of the data and evaluation con- ditions of each task and reports results of the participating systems. 1. Introduction This report informs on the evaluation campaign organized by the 16th International Workshop on Spoken Language Trans- lation (IWSLT). Spoken language translation (SLT) is the problem of translating speech input from a source to a tar- get language. In its most general variant, named speech- to-speech translation, the goal of SLT is to produce spo- ken output in the target language. However, since the be- gin, mainly due to the difficulty in developing and eval- uating speech-to-speech translation systems, the scope of IWSLT has been limited to speech-to-text translation sys- tems. Hence, IWSLT’s focus is on the integration of two problems: automatic speech recognition (ASR), i.e. the con- version of a speech signal into a transcript, and machine translation (MT), i.e. the translation of a transcript into an- other language. SLT is more difficult than the simple con- catenation of ASR and MT systems [1], for mainly two rea- sons: i) ASR systems are prone to make errors, ii) MT sys- tems, which are typically trained on written language, do not perform well on noisy transcripts of spoken language [2]. On the other hand, SLT is a very important problem to solve, given the vast amount of human-to-human communi- cation based on speech. Besides personal communication for business and traveling, important applications of SLT are for instance international teleconferencing and the subtitling of audiovisual content, just to mention two. Since 2004 [3], IWSLT has put effort in organizing chal- lenging SLT tasks, which were at the same time affordable for the currently available technology. Thus, IWSLT initially focused mainly on multilingual communication in the trav- eling domain [4, 5] a progressively moved to translation of speeches [6], university lectures [7], teleconference chats [8], dialogues in the health-care domain [9], etc. In parallel with the advance of technology, from statistical to neural models, also data collections prepared for the evaluations evolved, too. While for instance previous training data reflected the separation of ASR from MT, the recently collected …

Content truncated to 500 words