作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2018, Vol. 44 ›› Issue (7): 199-204,211. doi: 10.19678/j.issn.1000-3428.0047729

• 人工智能及识别技术 • 上一篇    下一篇

基于多特征融合与双向RNN的细粒度意见分析

郝志峰 1,2,黄浩 1,蔡瑞初 1,温雯 1   

  1. 1.广东工业大学 计算机学院,广州 510006; 2.佛山科学技术学院,广东 佛山 528000
  • 收稿日期:2017-06-27 出版日期:2018-07-15 发布日期:2018-07-15
  • 作者简介:郝志峰(1968—),男,教授,主研方向为机器学习、人工智能;黄浩(通信作者),硕士研究生;蔡瑞初,教授;温雯,副教授。
  • 基金资助:

    国家自然科学基金-广东联合基金(U1501254);广东省自然科学基金(2014A030306004,2014A030308008);广东省科技计划项目(2015B010108006,2015B010131015);广东特支计划项目(2015TQ01X140);广州市珠江科技新星专项(201610010101);广州市科技计划项目 (201604016075)。

Fine-grained Opinion Analysis Based on Multi-feature Fusion and Bidirectional RNN

HAO Zhifeng  1,2,HUANG Hao  1,CAI Ruichu  1,WEN Wen  1   

  1. 1.School of Computers,Guangdong University of Technology,Guangzhou 510006,China; 2.Foshan University,Foshan,Guangdong 528000,China
  • Received:2017-06-27 Online:2018-07-15 Published:2018-07-15

摘要:

文本细粒度意见分析主要有属性抽取和基于属性的情感分类2个任务,现有方法完成上述任务采用条件随机场(CRF)训练属性抽取模型,并运用循环神经网络(RNN)训练基于属性的情感分类模型。但同时完成2个任务则无法找到属性和情感倾向的对应关系。针对该问题,提出利 用双向RNN构建基于序列标注的细粒度意见分析模型。通过融合文本的词向量、词性和依存关系等语言学特征,学习文本的修饰和语义信息,并设计一个时间序列标注模型,同时抽取属性实体判断文本的情感极性。在真实数据集上的实验结果表明,与CRF、TD-LSTM、AE-LSTM 等模型相比,该模型情感分类效果提升明显。

关键词: 特征融合, 词向量, 循环神经网络, 属性抽取, 细粒度意见分析

Abstract:

Text fine-grained opinion analysis mainly includes attribute extraction and attribute-based sentiment classification.The existing methods accomplish the above tasks by adopting Conditional Random Field(CRF) training attribute extraction model and using a Recurrent Neural Network(RNN) to train attribute-based emotion classification model.However,the completion of two tasks at the same time can not find the corresponding relationship between attributes and emotional tendencies.To solve this problem,a two-dimensional RNN is proposed to build a fine-grained opinion analysis model based on sequence annotation.By merging the linguistic features of the text,such as word vectors,part of speech and dependence,it learns the text’s modification and semantic information,designs a time series annotation model,and extracts attribute entities to determine the sentiment polarity of the text.Experimental results on real datasets show that compared with CRF,TD-LSTM,AE-LSTM and other models,the emotional classification effect of this model obviously improves.

Key words: feature fusion, word vector, Recurrent Neural Network(RNN), attribute extraction, fine-grained opinion analysis

中图分类号: