Multilingual Generation for Translation in Speech-to-Speech Dialogues and its Realization in Verbmobil

This paper presents the generation module of the speech-to-speech dialogue translation system Verbmobil. Spontaneous speech, large multilingual vocabulary, difficulty of the translation task, robustness and real-time constraints make the design of such a module very challenging. In order to overcome these difficulties, we have developed a system based on a general kernel and the declarativity of the knowledge sources. This fully implemented system proves the practical relevance of several techniques such as constraint-solving for the microplanning task or HPSG-to-TAG compilation for syntactic realization. In addition to the successful deployment of our module into the Verbmobil system, the kernel has been adapted to other domains and tasks. 1 Speech–to–Speech Dialogue Translation Speech-to-speech dialogue translation systems try to exploit the recent advances in speech and natural language processing to go beyond the language barrier. Their task is the multilingual translation of spontaneous speech input in order to allow several speakers to discuss freely in their own languages. In practice the dialogues are limited to restricted domains. In contrast to a text translation system, the processing of spontaneous speech requires extended functionalities in almost every module because the system has to be able do deal with, e.g., ill–formed and disfluent (hesitations, repetitions, repairs) speech input. These additional functionalities have a strong impact on the generation task and raise to several specific requirements. 1.1 Requirements on Generation While the macroplanning task is important and mandatory in text generation, it is limited in dialogue translation. Most of the related problems, for instance sentence segmentation and sentence mood choice, have been solved in the source language. On the contrary, the spontaneous speech translation task leads to specific requirements: Robustness: Incomplete and non-wellformed input for the generation module can result from spoken disfluencies, from faults inside previous processing level (speech recognition module for example) or from the translation task. Time constraints: The usability of a dialogue translator depends on its acceptability with respect to speed. Portability: The generator has to cover a large vocabulary and different languages. DFKI GmbH, Stuhlsatzenhausweg 3, D-66123 Saarbrücken, email: becker kilger lopez poller @dfki.de Although the generation system presented in this paper tackles these constraints without being domain-dependent, they have been integrated first in the real-world speech-to-speech dialogue translation system Verbmobil. 1.2 The Verbmobil system The Verbmobil project is a long term research effort: Started in 1993 the system is reaching completion this year. The Verbmobil System [12] aims to translate in spontaneously-spoken dialogues robustly and bidirectionally for German/English and German/Japanese. The current Verbmobil system is speaker-independent with the ability to process large vocabularies (about 10,000 words) for German and English. Verbmobil works on specific domains (negotiations about scheduling appointments, travel planning and hotel reservations) and can support different voice media from face-to-face dialogue to cellular phones. Due to the high complexity of this task, the system is subdivided into 24 separate subtasks (implemented modules). Moreover, Verbmobil combines different approaches to machine translation. A simplified overall architecture of the Verbmobil system is presented figure 1. User 2 Language B Speech Synthesis Robust Semantics Statistical Translator Generator Transfer Translator Translator Case-based Dialog act-based User 1 Language A Speech Recognizer