A generic multi-lingual open source platform for limited-domain medical speech translation

50 EAMT 2005 Conference Proceedings A Generic Multi-Lingual Open Source Platform for Limited- Domain Medical Speech Translation Pierrette Bouillon1, Manny Rayner1,2, Nikos Chatzichrisafis1, Beth Ann Hockey2, Marianne Santaholma1, Marianne Starlander1, Yukie Nakao3, Kyoko Kanzaki3, Hitoshi Isahara3 (1) University of Geneva, TIM/ISSCO/ETI {pierrette.bouillon, nikolaos.chatzichrisafis}@issco.unige.ch, {marianne.santaholma, marianne.starlander}@eti.unige.ch (2) ICSI/UCSC/NASA Ames Research Center mrayner@riacs.edu, bahockey@email.arc.nasa.gov (3) NICT, Kyoto, Japan {isahara, kanzaki}@nict.go.jp, yukie-n@khn.nict.go.jp Abstract. We present an overview of MedSLT, an Open Source platform for developing limited-domain medical speech translation systems. We focus in particular on the speech understanding architecture, which uses grammar-based language models derived using cor- pus-based specialisation methods from a single linguistically motivated grammar, and summarise the results of two evaluations which investigate the appropriateness of these de- sign choices. Other sections describe the interlingua and its relationship with the recogni- tion architecture, and the current demo system. 1. Introduction Medical domains are an attractive area for spo- ken language translation systems. They offer not only potential for interesting applications that can be of real use, but can also be suffi- ciently constrained to permit reasonable per- formance. A recent high-profile example is the hand-held Phraselator translator1, which is cur- rently being used by the US military in Iraq. Discussions with physicians suggest that doctor/patient examination dialogues are both useful and manageable as a task, and have sev- eral advantageous properties from the point of view of building a spoken language translation system. As interactions are highly constrained, the input to be recognised is limited. Examina- tions can also be divided into smaller subdo- mains based on symptom types, for example, headaches, chest pains, gastric complaints, and so on. This gives the possibility of further con- straining the range of utterances that needs to be 1 http://www.phraselator.com recognised at any specific moment in the dia- logue. Another important point is that the dialogue can be mostly initiated by the doctor, with the patient giving only non-verbal responses. Speech recognition in the doctor to patient di- rection is much easier than the reverse, since doctors who use the system regularly will have had time to acclimatise to it; patients, on the other hand, will in general have had no previous exposure to speech recognition technology, and may be reluctant to try it. With this type of design, it is also possible to improve accuracy by displaying a back-trans- lation into source language via interlingua, and to let the physician-user accept or reject the recognition before invoking translation. In this way, the doctor can check if the content of the sentence has been correctly understood, provid- ing a safeguard against possible parsing and recognition errors. The MedSLT project is currently engaged in developing a medical speech translation frame- work based on these general ideas. In the rest of A generic multi-lingual open source platform for limited-domain medical speech translation EAMT 2005 Conference Proceedings 51 the paper, we describe the architecture of the system, summarise two recent evaluations we have carried out investigating our design choi- ces, and describe the actual prototype that will …

Content truncated to 500 words