【作者】 王玉美；

【导师】 阮晓钢；

【作者基本信息】 北京工业大学 ， 模式识别与智能系统， 2003， 硕士

【摘要】 论文以汉语文法分析为应用背景，采用专家系统的体系结构，构造了一种基于神经网络的汉语文法分析系统，其核心是构造存储和管理文法知识的知识库和具有语言专家智能行为和文法分析能力的推理机。 论文取得了以下研究成果： 1．提出了一种基于神经网络的汉语文法分析专家系统。论文给出了系统的总体结构，详细介绍了该系统的各结构模块及其功能。并分析了将专家系统与神经网络相结合的智能化系统用于汉语文法分析的可行性。 2．论文根据现代汉语语法简表提炼出文法知识和语法规则，并对产生式知识表达方法进行了系统的分析。为了便于知识的管理和利用神经网络存储知识，将产生式表达的语法规则进行标准化处理，即使之二元化。 3．根据知识描述的特点，提出了采用人工神经网络存储文法分析知识并构造知识库的方法。训练样本来源于根据现代汉语语法简表总结的汉语文法知识的规则库，经训练后的神经网络的连接权值和阈值矩阵即为神经网络专家系统的知识库。 4．根据文法规则的特点，分析了与／或树的搜索策略，推理机制采用了与／或树深度优先搜索策略和自顶向下的推理策略，并提出了相应的文法分析推理算法。其特色是专家系统与神经网络有机结合，从不同的层次进行知识推理，并以神经网络来完成自学习功能，提高了文法分析的效率。 5．在实验部分，对BP网络的知识存储能力进行了研究。对不同的BP网络模型的知识存储能力进行了比较分析，确定了最优的即网络模型。在系统实现部分，运用VC++进行了系统程序和界面的设计。推理机制采用了两个BP神经网络同时进行推理的方法。系统运行结果表明，该系 北京工业大学工学硕士学位论文 统能够快速的记忆和存储知识，进行有效的文法推理。 6．论文研究了汉语文法分析中采用启发式搜索的方法，通过启发函数值的 性质引入了“概率语法”的概念，推导了概率语法中文法分析的方法及 相关性质。论文针对文法分析中的知识扩充问题，提出了确定编码的论 域和采用多个神经网络同时进行推理的方法。并进一步分析了模糊系统 和神经网络的相似之处，提出了进一步改进的方法。 7、论文设计、实现的汉语文法分析系统在运行中取得了较理想的效果，其 研究成果对汉语文法分析方法的研究具有参考价值。更多还原

【Abstract】 The thesis develops an Chinese syntax analysis system based on neural network with expert system frame on the ground of Chinese syntax analysis, in which constructing the knowledge database for saving and managing syntax knowledge and developing the reasoning machine with the linguist intelligent behavior and the ability of syntax analysis is the most important.The main research results are clarified as below.1. The Chinese syntax analysis expert system based on neural network is proposed. The thesis presents the whole system framework, and introduces the frame and function of each construct module in detail. What’s more, the feasibility of the intelligent system for Chinese syntax analysis, which is united by expert system and neural network, is discussed in the thesis.2. In the thesis, syntax knowledge and grammar rules are abstracted from the modern Chinese grammar library, and the knowledge express method by product rule is anal sized systematically. In order to make knowledge manageable and exploit network to save knowledge, the grammar rules with product express are standardized, that is dualistic.3. The method of saving syntax knowledge and constructing knowledge database with neural network is proposed in the thesis according to the special feature of knowledge description. The train samples come from the rule database which is summarized the by modern Chinese grammar library. The join weight value and threshold value matrix of neural network after training form the knowledge database of the expert system based on neuralnetwork.4. The thesis discusses the search strategy of and/or tree according to the express feature of the grammar rule, and adopts the and/or tree deep first search strategy and the top down inference strategy in the reasoning mechanism. The corresponding Chinese syntax analysis inference algorithm is developed and is constructed experiment research, which feature is that the expert system and the neural network are united organically and the knowledge inference is carried out by different steps, besides, the function of self-study is implemented by neural network, which improves the efficiency of syntax analysis greatly.5. The ability of saving knowledge of BP network is researched in system experiment. And the ability of saving knowledge of different BP network is compared and anal sized so as to establish the optimization BP network model. In the portion of system implement, the design of system program and interface is completed with Visual C++. The reasoning mechanism adopts the method of two BP network inference at the same time. The result of system running shows that the system not only can fleetly remember and save knowledge but also can conduct syntax inference effectively.6. The thesis explores the method of introducing heuristic search in Chinese syntax analysis, and by means of the properties of heuristic function value, the thesis introduces the concept of probability grammar, and also deduces the Chinese syntax analysis method and relative properties in probability grammar. In the thesis, being aimed at the knowledge expansion problem, the methods of determining the code’s demonstration fields and adoptingmultiple neural networks inference are presented. In addition, the thesis analyzes the similarity of the vague system and the neural network, and also proposes the method for further improvement.7. The Chinese syntax analysis system, which is designed and implemented in the thesis, obtains good results in running. The research findings are worthy to be referenced in Chinese syntax analysis methods study field.更多还原