计算机工程与应用 ›› 2018, Vol. 54 ›› Issue (11): 145-152.DOI: 10.3778/j.issn.1002-8331.1701-0337

• 模式识别与人工智能 • 上一篇    下一篇

基于用户和产品Attention机制的层次BGRU模型

郑雄风,丁立新,万润泽   

  1. 武汉大学 计算机学院,武汉 430072
  • 出版日期:2018-06-01 发布日期:2018-06-14

User and product Attention mechanism based hierarchical BGRU model

ZHENG Xiongfeng, DING Lixin, WAN Runze   

  1. School of Computer Science, Wuhan University, Wuhan 430072, China
  • Online:2018-06-01 Published:2018-06-14

摘要: 文本情感分类的核心问题是如何有效地表示文本的情感语义,然而,目前的大多数方法只考虑到了文本内容中的情感语义,忽略了与文本内容相关的用户信息以及文本内容所描述的产品信息。已有的包含用户和产品信息方法也存在着以下两个问题:(1)不能有效地表示用户和产品信息,而且模型复杂度过高导致训练速度满。(2)文本情感语义表示模型过于简单,不能有效地表示文本中的上下文语义信息。针对以上两个问题,提出了相应的解决方案:(1)针对用户和产品的评价数据,利用奇异值分解(Singular Value Decomposition,SVD)的方法得到用户和产品的语义准确的先验信息,同时避免了用户和产品信息等相关参数的训练,缓解了模型复杂度高的问题。(2)利用双向的门循环单元(GRU)模型代替原有的简单模型,更加有效地结合了文本中的上下文语义信息。实验结果表明:相比传统的文本分类方法,提出的方法有更好的分类效果,在部分实验数据中达到了最好的分类准确度。同时模型的训练速度也得到了提升。

关键词: 深度学习, 文本情感分类, 注意力机制, 奇异值分解, 双向门循环单元

Abstract:

The core issue of text emotion classification is how to represent the emotion semantics of the text effectively. However, most current methods only consider the emotion semantics in the text content, ignoring the user and product information related to the text content. Existing methods that incorporate user and product information still exist two problems:(1)The methods can not effectively represent the user and product information. (2)The semantic representation model is too simple and can not effectively represent the context semantic information in the text. To address these issues, this paper puts forward the corresponding solutions:(1) According to the user and the product data, the Singular Value Decomposition(SVD) method is used to obtain the prior information of user and product. This method avoids the training of user and product parameters, which solves the problem that the iterative speed of the model is very slow. (2)Using the bidirectional Gated Recurrent Unit(GRU) model instead of the original simple model, the context semantic information in the text is more effectively combined in the document feature representation. The experimental results show that the proposed method has good classification performances, even achieves the best state-of-the-art result at some experiment datasets. Meanwhile it improves the training speed of the model.

Key words: deep learning, text sentiment classification, Attention, Singular Value Decomposition(SVD), Bidirectional Gated Recurrent Unit(BGRU)