Mind your language! Controlled language for inference purposes

Controlled Generation for Speech-to-Speech MT Systems Arendse Bernth IBM T.J. Watson Research Center P.O. Box 218 Yorktown Heights NY 10598 arendse@us.ibm.com Abstract In spoken dialog systems, a well- crafted prompt is important in order to get the user to respond with an expected type of utterance. We identify a new, important area for research in speech- to-speech translation, which focuses on the fact that the output of the MT sys- tem serves as the prompt for the user on each end. The MT engine used in speech-to-speech translation must pay special attention to its generation com- ponent, to such an extent that it makes sense to talk about controlled genera- tion. Some rules for controlled gener- ation are given. 1 Introduction It is valuable to combine controlled language (CL) and machine translation (MT); see e.g. (Huijsen, 1998; Mitamura, 1999; Bernth and Gdaniec, 2001). Often text that is input to the MT system is constrained to conform to a specific CL that will make it easier for the MT system to per- form well. In other words, CL is applied to the in- put. In this paper we propose a different scenario where CL is applied to the output. This seems of particular use for speech-to-speech MT systems, but other uses are not hard to imagine. In spoken dialog systems (SDSs), a well- crafted prompt is important in order to get the user to respond with an utterance that the system can handle. In this paper we identify a new, important area for research in speech-to-speech translation, which focuses on the fact that the output of the MT system serves as the prompt for the user on each end. Since prompt design is extremely im- portant in eliciting expected types of responses, the MT engine used in speech-to-speech trans- lation must pay special attention to its genera- tion component, to such an extent that it makes sense to talk about controlled generation. Thus this paper stands at the intersection of three im- portant natural language technologies: Controlled language, machine translation, and speech pro- cessing. We will first, in Section 2, give some examples demonstrating the importance of a good prompt for SDSs. In Section 3 we relate MT out- put to dialog prompts, and in Section 4 we give some suggestions for CL rules in the context of speech-to-speech MT. 2 Importance of a Good Prompt There are at least two reasons that a well-crafted prompt is important. One is to ensure that the system does not get a user response that it can- not handle, and the other is to ensure that the user does not get confused. Since the users of speech- to-speech MT typically come from different lin- guistic backgrounds with different linguistic con- ventions, it is particularly important to pay atten- tion to the problem of potential confusion. In this section we illustrate both cases with ex- amples taken from the world of travel. For the target language parts, just the gloss in English is given.1 1Some …

Content truncated to 500 words