Development of a Speech Understanding system and a Spoken Dialog system

我们开发了日语口语对话系统。这个对话系统是在封闭的世界观光导游。该系统对旅游信息进行引导，用户可以通过自然语言语音输入系统。该系统由语音识别部分、句子理解部分、对话处理部分、用户话语预测部分等部分组成，语音识别部分采用音节Hacker（隐马尔可夫模型）对语音的音节进行建模，对输入语音进行识别。CFG在句子理解部分，利用日语词典和KAKARIUKE规则对语音识别得到的文本进行处理在对话处理部分，进行省略补语和代词指称，则通过对对话规则的解释来进行对话。这种对话规则可以很容易地根据不同的情况进行调整。对话的澄清和验证部分是为了消除它们的歧义。该系统引导用户并向用户提出积极的问题，以获得消歧的信息。然而，在这样一个有限的任务域上，用户往往会说各种各样的句子类型，因此很难正确地识别语音。用户话语预测部分预测用户话语的单词/语法以供系统响应，以提高系统与用户之间口语对话的可靠性。在系统评估中，我们获得了足够的语音识别率来进行对话，对话系统可以与用户自然匡威。

We developed the spoken Japanese dialog system. This dialog system is in the closed world of sightseeing guide. The system guides the information about singhtseeing, and user can input to the system through natural language speech. This sysem consists of speech recognition part, sentence understanding part, dialog proessing part, user utterance prediction part, and so on.The speech recognition part recognized the input speech using syllable HMMs (Hidden Markov Model) that model the syllables of speech. CFG (Context Free Grammar) is used for modeling the linguistical restriction of user utterances.In the sentence understanding part, the text obtained form the speech recognition is processed using Japanese lexicon and KAKARIUKE rules (dependency grammar), then transformed to the semantic network using case frames.In the dialog processing part, the ellipsis complement and pronoun reference are performed, then the dialog is proceeded by the interpretation of the dialog rules. This dialog rules can easily adjusted to the various situations.In the dialog, ambiguities of meanings of input sentences often occur. The part of dialog for clarification and verification is performed to disambiguate them. The system leads the user and asks the user a question positively to get the information for the disambiguation. There process can make the dialog certainly.On such a limitative task domain, however, user tends to speak various sentence types, so it is difficult to recognize the speech correctly. The user utterance prediction part predicts the word/syntax of user's utterance for the system's response to improve the reliability of spoken dialog between the system and user.On the system evaluation, we got the enough speech recognition rate for progressing the dialog, The dialog system could converse with a user naturally.