Rescarch on Natural Language Dialogue System using Adverbs by Case-base

本研究的目的是实现一个无主题的自然语言对话系统，提供环境，人与计算机进行交流。由于我们的开发方法是使用范例库来产生计算机的话语，所以无主题自然语言对话系统可以通过嵌入实时修改范例库的功能来提供用户所需要的信息。许多副词反映了使用者对自己所处环境的主观想象。当一个副词用得不自然时，人们就不能正确地理解句子。本研究分析了时间副词和程度副词的用法。时间副词通常表达说话人对时间认知的主观性。我们使用了一个扩展的Hornstein模型，并在一个自然语言对话系统上进行了实验。等级副词通常修饰形容词，但有时也修饰动词。在这种情况下，形容词被省略了。本文分析了这一现象，建立了一个估计省略形容词的模型，通过关键词和句子结构来估计输入句子与实例数据的匹配程度。在关键词匹配中，我们的系统提取了重要的术语，如名词，动词和形容词。在结构匹配中，根据词在句法树上的位置来判断词的重要性。为了实现我们的系统，我们开发了一个形态分析器和一个句法分析器，这是有效的提取信息，我们的对话系统。此外，我们开发了一个句子生成器，使用slot-method.express用户意图，我们研究了面部标记，有一个相当于副词的字符。虽然人脸标记产生的机制还不清楚，我们模拟情绪的过渡和估计一个人脸标记，这是适合于当前的句子神经网络。

The aim of this research is to implement a topic-free natural language dialogue system that provides environments to communicate human with computers. Since our developing method uses case-database to generate computer's utterances, the topic-free natural language dialogue system can supply user's desirable information by embedding a function of revising the case-database in real-time to the system.In this research, we focus on the use of adverbs in such a natural language dialogue. Many adverbs reflect user's subjective images about his situation. When an adverb is used in unnatural way, human cannot interpret the sentence correctly. In this research, we analyzed usages of time-adverb and grade-adverb. Time-adverb usually express the subjectivity of a speaker for time recognition. We used an extended Hornstein model and experimented it on a natural language dialogue system. Grade-adverb usually modifies adjectives but sometimes modifies verbs. In this case, adjectives are omitted. We analyzed this phenomenon and made a model of estimating omitted adjectives.A matching between an inputted sentence and case-data is evaluated by keywords and structures of sentences. In the keyword-matching, our system extract important terms like a noun, a verb and an adjective. In the structural-matching, the importance of terms is judged by the position on the syntactic tree. To implement our system, we developed a morphological analyzer and a syntactic parser, which are efficient to extract information for our dialogue system. In addition, we developed a sentence generator using a slot-method.express user intention, we researched on face-marks that have an equivalent character to adverbs. Though the mechanism of generating face-marks has not been known yet, we modeled the transition of emotions and estimating a face-mark that is suitable for the current sentence by neural network.