Prospects for advanced speech translation

MT Summit VII_______________________ Sept. 1999 Prospects for Advanced Speech Translation Hitoshi Iida Speech and Language Information Processing Laboratory Sony Computer Science Laboratories Tokyo, Japan iida@csl.sony.co.jp Abstract Speech communication includes many impor- tant issues on natural language processing and they are related with desirable advanced speech translation systems. Advanced systems need to be able to handle the interaction for speech communication, pragmatics in speech, and arbitrariness of speech usage. General characteristics of speech communication are discussed. Also the various viewpoints regarding interaction, pragmatics, and arbitrary usage are discussed. Some of the present speech translation approaches are explained and new basic technologies are introduced. In this paper, a synthetic NLP technology such as a composite art form is proposed for speech communication and speech translation. 1 Introduction When we are doing cross-language communication in daily life, a conventional machine translation system that translates written language texts does not always give the correct translation because the utterances depend on the situation and the real meaning can differ from the literal meaning. Translation results suitable for speech communication cannot be absolutely decided, but relatively dependent on each dialogue situation. Figure 1 shows the state of affairs regarding the differences between a conventional MT world and a speech translation world. In case 1, the source language grammar, target language grammar, and transfer rules might establish an ideal machine translation where the translation result can be determined uniquely and its meaning must be absolutely correct for the readers and the recipients. On the other hand, in case 2, any utterance meaning is situation depended and the translation result is one candidate of potential utterances in the target language. In this case what an MT system requires is just a necessary and sufficient expression suitable for the dialogue situation and the context in the target language. General characteristics of speech communication are discussed in the next section. Also, various viewpoints regarding interaction, pragmatics, and arbitrary usage are discussed in the section 3, 4, and 5, respectively. Some of the present speech translation approaches are explained in the section 6. And requirements for advanced speech translation and desirable new basic technologies are discussed in sections 7 and 8. 2 Speech Communication Almost all the current continuous speech recognition systems output word sequences [Lavie 96]. And they gen- erally do not make any spelling errors because wrong phoneme sequences are handled by modifying them into certain word sequences. Thus, a method to eliminate incorrect word sequences including - 107- MT Summit VII______________________________________________________________Sept. 1999 unnecessary or unrelated words is different from an error correction method and is expected to avoid errors for speech-to-speech systems. Three types of 'ill-natured' speech recognition outputs can at least be observed. 1) a word sequence that includes an incorrect subse- quence that may be analyzed by a kind of syntactic parser, 2) an incorrect subsequence that can be interpreted as a correct meaning by a kind of semantic analysis, and 3) a word sequence that is difficult to divide into sen- tences or clauses only …

Content truncated to 500 words