Intelligent pattern recognition and understanding by integrating probabilistic and symbolic reasoning

对于智能语音识别和理解，我们研究了多个知识层面的推理策略，并将其集成到语音理解系统中，具体如下：(1)音素识别我们首先改进了音素识别，这是整个系统的基础。提出了基于成对贝叶斯分类器的音素HMM， 27个音素识别率为83.1%，653个单词识别率为84.8%。(2)句法分析句法分析器是在词汇和句法层面将概率推理和符号推理相结合而开发出来的。启发式搜索是基于HMM的语法规则和概率预测进行的。提出了一种使用词对约束作为启发式方法的允许^*的上下文无关解析。(3)关键字识别：在没有语法规则的情况下，可以用多个关键字来理解句子。然而，传统的方法仅使用自己的分数提取关键词，因此不足。提出了一种新的识别算法，该算法假定输入是包含目标关键词的音素或词序列。(4)语义分析开发了基于网络的语义分析器，该分析器同时接受n个最优词序列和关键字格，并获得语义表示。在此基础上，将语义、语用和对话层面的知识进行整合，并结合候选词的概率得到似是而非的假设。(5)语音理解系统在语音理解系统上实现了两种推理策略。一种是语法驱动的，它集成了(1)、(2)和(4)。另一种是语义驱动的，它集成了(1)、(3)和(4)。我们在一个词汇量为244，单词困惑度为80的任务上评估了这两个系统。对于符合语法的话语，句法驱动的准确率为65.5%，而语义驱动的准确率仅为44.0%。然而，语义驱动的方法对于语法错误的话语是有效的。

For intelligent speech recognition and understanding, we have examined reasoning strategies on several knowledge-levels, and integrated them into speech understanding systems as follows:(1) Phoneme recognitionWe have firstly improved phoneme recognition, which is the base of the whole system. Phoneme HMM based on pair-wise Bayes classifiers is proposed with 27 phoneme recognition rate of 83.1% and 653 word recognition rate of 84.8%.(2) Syntactic analysisSyntactic analyzer is developed by integrating probabilistic reasoning and symbolic reasoning on vocabulary and syntax level. Here heuristic search is performed based on prediction by syntax rules and probabilities of HMM. A^*-admissible context-free parsing with word-pair constraints as heuristics is presented.(3) Keyword spottingIt is possible to make sense of sentences with multiple keywords, without syntax rules. However, conventional method extracts keywords using only the scores of their own, thus insufficient. A new spotting algorithm is presented with assumes logical constraint that the input is a phoneme or word sequence containing target keywords.(4) Semantic analysisNetwork-based semantic analyzer is developed which accepts both N-best word sequences and a keyword lattice and obtains a semantic representation. Here semantic, pragmatic and dialog-level knowledge is integrated and plausible hypothesis is obtained by combining probabilities of candidate words.(5) Speech understanding systemTwo reasoning strategies are implemented on speech understanding systems. One is syntactic-driven which integrates (1), (2) and (4). The other is semantic-driven which integrates (1), (3) and (4). We have evaluated both systems on a task whose vocabulary size is 244 and word perplexity is 80. For grammatical utterances, syntactic-driven approach got an accuracy of 65.5%, while semantic-driven achieved just 44.0%. However, semantic-driven approach is effective for out-of-grammar utterances.