基于深度神经网络的嵌入式向量及话题模型  

Embedded vector and topic model based on deep neural networks

作　　者：何永强[1] 秦勤[1] 王俊鹏[2] 

机构地区：[1]河南工程学院计算机学院,河南郑州450007 [2]北京理工大学计算机学院,北京100081

出　　处：《计算机工程与设计》2016年第12期3384-3388,3399,共6页Computer Engineering and Design

基　　金：河南省高等学校重点科研基金项目(15A520054);河南省科技厅科技计划课题基金项目(112102310550)

摘　　要：在对文档集合进行话题分析的过程中,为描述文档中单词间的依赖关系,提高话题分析的效果,提出一种基于反馈递归神经网络的嵌入式向量生成及话题模型。在将单词表示为One-hot向量后,采用递归神经网络将文档嵌入在低维的向量空间。在文档的嵌入式向量计算过程中,采用LSTM(long short-term memory)描述单词间的前向依赖关系,提出一种反馈神经网络用于描述单词间的后向依赖关系。在话题分析模型中,采用原文档和变异文档对作为正样本,采用原文档和随机文档对作为负样本进行模型的训练。实验结果表明,该方法时间和空间复杂度低,具有更好的话题分析效果。While mining topics in a document collection,to describe the relationships between words and further improve the effectiveness of discovered topics,a feedback recurrent neural network based embedded vector and topic model was proposed.After representing a word as One-hot vector,a document was embedded into low dimensional space.During the process of document embedding,LSTM（long short-term memory）was applied to describe the forward relationships between words,and a feedback neural network was proposed to describe the backward relationships between words.In topic model,the original and muted docu-ment pairs were used as positive samples,and the original and random document pairs were used as negative samples to train the model.Experimental results show that,the proposed model has not only lower running time and memory,but also better effectiveness during topic analyzing.

关 键 词：话题模型 递归神经网络 深度学习 反馈机制 嵌入式 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]