权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

Fundamental Research of Speech Translation Based on High Accurate Speech Recognition and Language-Concept Understanding

基于高精度语音识别和语言概念理解的语音翻译基础研究

基本信息

批准号：
62420052
负责人：
DOSHITA Shuji
金额：
$ 14.02万
依托单位：
Kyoto University
依托单位国家：
日本
项目类别：
Grant-in-Aid for General Scientific Research (A)
财政年份：
1987
资助国家：
日本
起止时间：
1987 至 1989
项目状态：
已结题

项目摘要

With respect to recognition and understanding of spoken language intended for speaker-independent Japanese sentences, we have performed experimental evaluation and establishment of methods in each stages of speech, language, conceptual information processing as follows:1. Speech analysis and recognition We have proposed pair-wise discrimination method to catch their opposite characteristics of consonants, and shown the possibility of recognition with higher accuracy. We have expected the effect of combination of discriminant analysis and HMPI to improve accuracy of analysis, and verified that by experiments. And we have proposed a method to construct neural networks to discriminate phonemes based on piece-wise linear discriminant functions.2. Language analysis and translation We have developed a system analyzing conversational sentences which is more patient of errors included in results of speech recognition by applying the method that in analyzing spoken sentences we construct directly concepts of their meanings using conceptual information without the stage of syntactic analysis.3. Conceptual information processing We have designed a schedule management method which can be used for erroneous inputs by establishing the task of schedule management for individuals and by applying assumption-based reasoning using default for insufficient information of abbreviations and ambiguous expressions. And we have developed a system reconstructing geometric models for scenes from sentences with description about them.4. Integrated information processing mechanism We have developed integrated parser, which is a mechanism to implement integrated information processing, and considered counterplans for erroneous inputs in speech recognition for integrated natural language understanding systems which output most reliable interpretations with maintaining consistency.

针对独立于说话人的日语句子对口语的识别和理解，我们在语音、语言、概念信息加工的各个阶段进行了实验评价和方法建立，具体如下：语音分析与识别我们提出了双分辨方法来捕捉辅音的相反特征，并展示了识别精度更高的可能性。我们期望判别分析与HMPI相结合可以提高分析的准确性，并通过实验验证了这一点。我们提出了一种基于分段线性判别函数的神经网络识别音素的方法。我们开发了一种分析会话句子的系统，该系统对语音识别结果中包含的错误更有耐心。在分析口语句子时，我们使用概念信息直接构建其含义概念，而不需要句法分析阶段。我们设计了一种针对错误输入的进度管理方法，通过建立个人的进度管理任务，对缩略语和歧义表达式的信息不足采用默认的假设推理。并且我们开发了一个系统，从带有描述的句子中重建场景的几何模型。集成信息处理机制我们开发了集成解析器，这是一种实现集成信息处理的机制，并考虑了集成自然语言理解系统语音识别中错误输入的对策，该系统输出最可靠的解释并保持一致性。