聊天机器人中短文本分类技术研究.docxVIP

PAGE 25 聊天机器人中短文本分类技术研究 摘 要 本文探讨了聊天机器人中进行短文本分类的两种机器学习方法，分别是最大熵模型和卷积神经网络。其中，最大熵模型作为基线方法进行参照，而卷积神经网络则分别考虑单层神经网络、单连接神经网络、多连接神经网络结构，进行探究。对基线方法最大熵模型的实验显示，对于有明确主题词的类别（如饮食推荐、出行消费）有较好效果，而对于其他类别则会出现准确率、召回率都偏低的情况。对于卷积神经网络分类方法的探讨显示，在数据规模较小的情况下，网络结构的复杂程度越高，分类的性能在不同类别就越不均衡，具体表现就是相比基线方法，一些类别的准确率、召回率显著更高，而一些类别的准确率、召回率显著更低。因此，在数据量较小的情况下，更复杂的神经网络方法的表现未必能比传统方法表现得更好，而且随着模型复杂程度的增加，分类性能的不均衡性也会越来越突出。 关键词：人机对话；人工智能；最大熵模型；卷积神经网络 Abstract: This paper discusses two kinds of machine learning methods, which are the maximum entropy model (ME) and the convolution neural network (CNN). The maximum entropy model is used as the baseline method, and the convolutional neural network is considered in multi structures, including a single layer neural network, a single connected neural network and a multi connected neural network structure. The baseline method experiment shows that, for a category with clear theme word (such as dietary recommendations, travel consumption) classifier can perform well, but for other categories, accuracy rate and recall rate are low. For convolutional neural network classification method, in the small data set, the higher the complexity of the network structure, the more unbalanced the classification performance in different categories, which bring significantly higher accuracy and recall rate in some classes, as well as significantly lower accuracy and recall rate in other classes. Therefore, in the case of small amount of data, more complex neural network method may not be able to perform better than the traditional method, and with the increase of model complexity and unbalanced classification performance will become increasingly prominent. Keywords: human computer dialogue, artificial intelligence, maximum entropy model, convolutional neural network  目 录 TOC \o 1-3 \h \z \u 摘 要 I Abstract: II 目 录 III 第1章 绪论 1 1.1 课题来源 1 1.2 研究的目的和意义 2 1.3 国内外在该方向的研究现状及分析 3 1.4 主要研究内容 7 1.5 研究方案 8 1.6 本章小结 9 第2章 分类方法研究 10 2.1 最大熵模型 10 2.2 卷积