Development of corpus-based technologies for robust spoken dialogue systems

我们开发了一种利用人类对话示例的对话处理技术，旨在实现用户友好的对话系统。该方法以数据驱动的方式执行语音理解、语音生成和语音规划等大部分模块。以真实的对话实例为数据，使系统能够实现以下功能：利用相似的案例进行鲁棒的语音理解，通过自然表情输出类似人类的自发语音，通过添加实例对对话系统进行扩展，如下所示：(1)对话实例的构建和利用：将大规模对话数据库构建为对话系统能够立即使用的数据。以CIAIR车载语音对话语料库为目标数据。对话语单位进行了语素分析、句法分析、言语行为分析等标记。(2)鲁棒性语音理解技术的发展：针对信息检索领域开发了对话理解技术。这是基于从口语对话语料库中选择与输入话语句子最相似的句子，并将输入话语句子的意图视为赋予句子的意图的思想。实际上，测量话语句与话语句相似度的方法被设计为一种利用语素、句法、关键词、句尾形式、语境等信息的相似度计算方法。(3)语音对话系统的实现：设计了基于实例的对话处理，并以运行中的餐厅导览系统为例实现。整个系统由语音理解、对话控制、语音生成和实例库等部分组成。通过实验验证了我们的口语对话处理技术的有效性和可实现性。少

We have developed a technique for dialogue proceesing of utilizing human conversation examples for aiming at the achievement of user-friendly dialogue systems. This method executes most of the modules such as speech understanding, speech generation, and speech planning in a data-driven manner. To use the real dialogue examples as data made the system become possible the following functions : the robust speech understanding byusing a similar case, the human-like spontaneous speech output bynatural expressions, and the extension of the dialogue system by adding the examples, as shown in the below :(1) Construction of the dialogue examples and their utilization : The large-scale dialogue database was constructed as a data for which the conversation system was able to be used immediately. The CIAIR in-car speech dialogue corpus was used as target data. The tags for morpheme analysis, syntactic analysis, and speech-act analysis, etc.were given to the utterance units.(2) The development of a … More technique for robust speech understanding : The technique for dialogue understanding was developed for information retrieval domain. This is based on the idea of choosing the most similar sentence to the input utterance sentence from among the spoken dialogue corpus, and regarding the intention of the input utterance sentence as the intention given to the sentence. Actually, the method of measuring the similarity between the utterance sentence and the utterance sentence was designed as a similarity calculation method of using information on the morpheme, the syntax, the key word, the end of sentence form, and the context, etc.(3) Implementation of the speech dialogue system : The example-based dialogue processing was designed, and implemented as a system of the restaurant guide in the running car. The entire system is composed of the speech understanding, the dialogue control, the speech generation, and the examples database. The effectiveness and realizability of our spoken dialogue processing technique were confirmed by the experiments. Less