Briefly Noted: Verbmobil

Verbmobil is a project sponsored by the German government, whose aim is to develop a translation system for face-to-face speech dialogue. This book is the report on a preliminary study of the project conducted by CSLI, Stanford University. The contents of the book are as follows: Chapter 1 contains introductory material; Chapter 2 and Chapter 3 are devoted to describing the current state-of-the-art of machine translation and speech recognition and speech synthesis technologies, respectively; and Chapter 4 reports a technical recommendation for the Verbmobil project based on the discussions in the preceding chapters. Chapter 2, which would be the most interesting for readers of Computational Linguistics, exhaustively covers various aspects of machine translation. First, the authors argue why machine translation is so difficult. They discuss the indeterminacy of language by referring to a number of examples: situatedness, mismatches between two languages in translation, and ambiguities of various levels. From these considerations, they suggest that good translation is not defined as preserving the meaning but as preserving the intention of the original (p. 27), sometimes by adding or deleting information (p. 26). Then, they review current technologies, historical perspectives, and theoretical issues such as syntax and grammar formalism. The most interesting and important discussion in this chapter is the section "Translation Strategy." First, the authors refer to the controversial argument concerning the comparison between the transfer approach and the interlingual approach, but they claim that "the issue of interest is not whether to pursue a transfer or an interlingual approach; the issue is which levels of analysis are necessary, and how to arrive at a representation suitable for generation of a target text" (p. 82). From this point, and by considering many examples of translation mismatch that imply the impossibility of a naive interlingual scheme, the authors propose a new architecture called "translation by negotiation." In this architecture, some interlingual representation is still supposed, but it does not require that an invariant representation that will be the same for the translation of a source sentence, or text, into any other language can be found at any time. There are three components: an analyzer, a generator, and a negotiator. The analyzer delivers to the negotiator an interlingual representation. The negotiator hands the interlingual representation to the generator. In some (or rare) cases, the generator can find the phrase in the target language that has exactly the same interlingual representation. In other cases, the generator reports to the negotiator error information such as the overspecification or underspecification of the input interlingual representation. Now, it is up to the negotiator to decide whether the error is fatal or not. When it is not so serious, the translation is accepted, but when it is serious, the negotiator has to solve it by some other means, such as referring to the context. In Chapter 2, the authors also discuss other topics: an overview of nonlinguistic translation approaches such as the stochastic approach and the example-based approach, and a comprehensive survey of current machine translation systems. These would also be quite informative to a reader. Chapter 3 is devoted to introducing speech-recognition technology and speechsynthesis technology, and Chapter 4 contains recommendations for the Verbmobil project. These might be less interesting for a reader of Computational Linguistics, so only a brief description is given below. In Chapter 3, general characteristics of speech are given, and current major speechrecognition technologies (the knowledgebased approach, the stochastic-based [HMMbased] approach, and the connectionismbased approach) are introduced, as well as a traditional template-based approach. These descriptions would be a good introduction for a reader who wants to see a general overview of speech-recognition technologies. As for speech synthesis, most pages are given