毕业生知识库

近年来，web2.0和移动互联网的飞速发展催化了人工智能领域技术和应用地不断迭代进步和革新，越来越多的人工智能系统和产品被引入到我们的生活之中。作为人工智能领域代表性的应用，人机对话系统是一种将机器视为认知主体的人机交互系统，其具有广阔的适用性、高度的便利性，并且与图灵测试的形式高度契合。

随着计算机软硬件技术和移动互联网的迅猛发展，能够有效处理非精确信息交互的、符合人类自然交互习惯的认知型人机对话系统受到了广泛关注。目前，如苹果的Siri、微软的Cortana、谷歌的Google Now和百度的度秘等个人助理系统，Rokid家庭机器人和亚马逊的Echo智能家居服务系统等各种形式的对话系统已经为日常生活带来了更多的便捷、智能与改变。

本文针对目前对话系统在实际应用中的一个场景进行建模，对面向垂直领域的智能对话系统进行研究，同时也探索其基本的推理机制，以达到探索学术前沿、改进已有方法、推动实际应用的目的。

在深入调研了现有学术界的研究前沿，细致考虑了本任务的适应性之后，本文主要开展了三个方面的工作：（1）在对话系统的问答对话子领域提出一种基于融合语义和词汇特征的模型，并在NLPCC 2016开放领域答案句选择评测中用该方法获得第2名的成绩；（2）本文采用端到端神经网络对话模型作为基础框架，对神经对话系统中对话历史的短时记忆编码、类人的推理机制和信息提取与融合等关键问题进行深入研究，在尽量不需要人工参与特征和规则设计情况下，对对话系统中的历史信息进行向量化编码表示，通过用户当前时刻的输入文本自动激活神经记忆单元中的相关语义信息进行融合，并基于此融合向量进行动态序列化解码生成系统的响应文本；（3）在该神经网络的框架下，本文也就对话系统的未登录词的选择这一关键子问题进行了更加深入的探索，提出一种多层次的记忆网络模型，并在一系列实验中和与现有方法的对比中表现出非常优异的性能。

In recent years, the rapid development of web and mobile Internet has enriched constant progress and innovation of technology and application in the field of artificial intelligence. More and more artificial intelligence systems and products were introduced to our life. As a typical application of artificial intelligence, human-computer dialogue system is a human-computer interaction system in which machine serves as a cognitive subject. It has wide applicability, high level of convenience, and highly fit in the form of the Turing test.

With the rapid development of computer hardware and software technology and mobile Internet, cognitive human-machine dialogue system with ability to deal effectively with the accurate information interaction, and accord with human nature interaction habits attracts much attention. Just as apple's Siri, Microsoft's Cortana, Google's Now, Baidu's secret personal assistant system, Rokid home robot and Echo smart home service, different forms of dialog system and other various forms of dialogue system has brought more convenient, intelligence, and change for daily life.

This thesis builds the model in view of a practical scene in the application of present dialogue systems, conducts a research towards intelligent reasoning dialog system in specific domain to explore the academic frontier, improve existing methods, and promote the practical application in AI.

After in-depth research of the existing academic research front and careful consideration the adaptability of this task, this paper has carried out three aspects of work: (1) We propose a model combining lexical and semantic-based features for answer sentence selection and won the 2nd place in the NLPCC 2016 open domain evaluation. (2) We adopt the end-to-end neural networks to conduct the research in problems like short-term memory encoding, reasoning and information extraction in dialog history. This network aims at a model without human participation characteristics and rules of design cases as far as possible, activating related semantic information in the memory units by the user's text input, and decode dynamically to produce the response text. (3) Under the framework of this network, we proposes a hierarchical memory networks to explore a key problem in dialog system about the selection of unknown words. Experiments and comparation of existing approaches shows good performance of our model.