Inductive inference of recognition rules for continuous speech recognition based on knowledge base

我一直在从事开发环境的知识为基础的系统建设的连续语音识别专家SPREX。本研究的最终目标是建立一种语音识别的知识工程方法论。具体来说，我开发了一个新的环境，知识库建设，它集成了学习机制和手工制作的规则建设。主要研究结果如下：1.新属性的生成：设计了一种通过合并已有属性生成新属性的方法.决策树的一般化：设计了两种决策树泛化算子，以提高决策树对未知样本的分类性能.方法的评价：进行了一项实验，以评估学习方法的性能。根据该学习方法得到的决策树的识别率为82%，人工规则的识别率为88%，证明了该学习方法在没有领域先验知识的情况下具有很高的性能.我们开发了一个最终的系统，它有一个新的体系结构，集成了学习和手工制作的规则构造方法。对上述相同数据，该方法的识别率达到88%，与人工规则的识别率相当。因此，我们得到了一个新的学习系统的非常高的性能。

I have been engaged in developing an environment for knowledge based system construction for continuous speech recognition expert SPREX. The final goal of this research is to establish a methodology of knowledge engineering approach to speech recognition. Specifically, I developed a new environment for knowledge base building which integrates both a learning mechanism and hand-made rule construction. The result is summarized as follows:1. Generation of new attributes: A method for generating some new attributes by combining existing attributes is designed.2. Generalization of decision tree: Two kinds of operators for generalization of decision tree are devised so as to augment the classification performance of the tree for unknown samples.3. Evaluation of the method: An experiment was done to evaluate the performance of the learning method. The recognition result of the decision tree obtained according to the learning method was 82% and that of man-made rule was 88%, which demonstrates the high performance of the learning method considering that it has no a priori knowledge about the domain.4. We developed a final system which has a new architecture integrating both learning and hand-crafting rule construction method. It attained 88% of recognition rate for the same data above, which is comparable to that of man-made rules. Thus, we obtained a new learning system of very high performance.