基于深度学习的对话系统主题分配技术研究

发布时间：2018-04-10 16:24

本文选题：主题分配 + 对话系统　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：随着人工智能的迅速发展,理解人类语言并能够与人类对话,给出相应的信息反馈的机器人成为了大多数人的需求。在这样的背景下,智能聊天机器人慢慢走入人们的视野。在这股机器人热潮中,能够帮助用户解决日常生活中各种事情的私人助理机器人应运而生。它能够帮助用户解决一系列生活中遇到的问题,如打车、预定餐厅等。与机器人对话的最大挑战就是要把人的自然语言翻译成机器可以听得懂的指令,从而给出相应的正确反馈。机器人能够给出正确反馈的第一步是理解人类需求,所以将用户输入理解为正确的主题,即对话系统中的主题分配起着非常重要的作用。本文的研究任务是将用户的输入分配到这个语句对应的主题下,以保证接下来的反馈方向正确。本文主要介绍了三种主题分配的方法:基于传统分类方法的主题分配模型、基于LDA主题模型特征扩展的主题分配方法以及基于深度学习的对话系统主题分配模型。基于传统分类方法的主题分配模型可以看做是文本分类任务,本文利用有监督学习的方法,在学习的过程中利用学习算法从训练语料中以特征的方式学习有用信息,从而得到主题分配的模型。该方法的效果高度依赖于人工选择的特征。基于LDA主题模型特征扩展的短文本分类方法考虑到了短文本词语稀疏性的特点,加入了扩展词后,主题特征被加入到了原来的短文本中,以达到语义扩展的效果,避免了短文本传统的文本表示方法特征稀疏的问题。实验表明,引入LDA主题词扩展特征后,主题分配模型取得了更好的效果。深度学习方法的避免了人工选取特征对实验结果的影响,使机器自动学习文本中的特征,增加了文本中隐藏的词与词之间的语义联系。本文利用基于卷积神经网络的句子分类方法以及基于循环神经网络的的方法作为主题分配的模型进行实验,实验结果表明基于深度学习的主题分配模型相比于传统方法取得了更好的效果。
[Abstract]:With the rapid development of artificial intelligence, the robot that can understand the human language and communicate with human, giving the corresponding information feedback, has become the demand of most people.In this context, the intelligent chat robot slowly walked into people's view.In this boom of robots, personal assistant robots, which can help users solve all kinds of things in their daily life, come into being.It can help users solve a range of life problems, such as taxi, restaurant reservations and so on.The biggest challenge in conversation with robots is to translate human natural language into instructions that machines can understand and give the correct feedback.The first step for robots to give correct feedback is to understand human needs, so the user input is understood as the correct topic, that is, topic assignment plays a very important role in the dialogue system.The task of this paper is to assign the user's input to the topic corresponding to the statement to ensure the correct direction of the following feedback.This paper mainly introduces three methods of topic assignment: the topic assignment model based on the traditional classification method, the topic assignment method based on the feature extension of the LDA topic model and the topic assignment model of the dialogue system based on in-depth learning.The topic assignment model based on traditional classification method can be regarded as the task of text classification. In this paper, we use supervised learning method and learning algorithm to learn useful information from training corpus in the way of feature.The model of topic assignment is obtained.The effect of this method is highly dependent on the characteristics of manual selection.The short text classification method based on the feature extension of LDA topic model takes into account the sparsity of the short text. After the extension word is added, the theme feature is added to the original short text to achieve the effect of semantic expansion.It avoids the problem of sparse features of traditional text representation in short text.The experimental results show that the topic assignment model is more effective when the extended feature of LDA theme words is introduced.The depth learning method avoids the influence of the artificial selection of the features on the experimental results, makes the machine automatically learn the features in the text, and increases the semantic relation between the hidden words and the words in the text.In this paper, the method of sentence classification based on convolution neural network and the method based on cyclic neural network are used as the model of topic assignment.The experimental results show that the topic assignment model based on deep learning is more effective than the traditional method.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1;TP18

【参考文献】