Parsing Spontaneous Dialogue

语音识别领域的最新进展表明，在未来十年内，实用的口语界面大有希望。但由于缺乏处理自发语音的有效方法，这项技术尚未准备好应用于许多领域。这个项目的重点是能够在语音识别系统的输出之外为自然语言解析器提供额外信息的技术。这些额外的信息将用于处理自发对话中出现的以下问题：将话语分割成有意义的单元，识别和纠正语音修复，检测和纠正语音识别错误，以及从表面结构，词汇线索和语调中解释预期的言语行为。该方法包括使用随机技术来识别可能的识别错误、语音修复、发音单位边界和韵律特征，然后将这些信息传递给解析器，解析器在给定所有信息的情况下确定最佳解释。该项目的结果将有助于开发用于自发对话的准确语音识别系统。将对所开发的技术进行评价，方法是对人类对话语料库进行一系列经验测试，并将这些技术整合到一个功能齐全的对话系统中。

Recent progress in the speech recognition field suggests great promise for practical spoken language interfaces within the next decade. But this technology is not yet ready for many applications because of the lack of effective methods for handling spontaneous speech. This project focuses on techniques that can provide additional information to a natural-language parser beyond the output of the speech recognition system. This extra information will be used in handling the following problems that arise in spontaneous dialogue: the segmentation of utterances into meaningful units, the identification and correction of speech repairs, the detection and correction of speech recognition errors, and the interpretation of the intended speech act from surface structure, lexical cues and intonation. The approach involves using stochastic techniques to identify likely recognition errors, speech repairs, utterance unit boundaries and prosodic features and then passing this information on to a parser which determines the best interpretation given all the information. The results of this project will help the development of accurate speech recognition systems for spontaneous dialogue. The techniques developed will be evaluated using a series of empirical tests on a corpus of human-human dialogues, and by integrating the techniques into a fully functional dialogue system.