计算机工程 ›› 2019, Vol. 45 ›› Issue (8): 315-320.doi: 10.19678/j.issn.1000-3428.0051831

• 开发研究与工程应用 • 上一篇    

基于深度学习的隐性评价对象识别方法

王仁武, 张文慧   

  1. 华东师范大学 经济与管理学部 信息管理系, 上海 200241
  • 收稿日期:2018-06-15 修回日期:2018-08-18 出版日期:2019-08-15 发布日期:2019-08-08
  • 作者简介:王仁武(1968-),男,副教授、博士,主研方向为数据分析、数据挖掘;张文慧,硕士研究生。
  • 基金项目:
    国家社会科学基金"基于数据驱动的图书馆资源发现系统平台研究"(16BTQ026)。

Implicit Evaluation Object Recognition Method Based on Deep Learning

WANG Renwu, ZHANG Wenhui   

  1. Department of Information Management, Faculty of Economics and Management, East China Normal University, Shanghai 200241, China
  • Received:2018-06-15 Revised:2018-08-18 Online:2019-08-15 Published:2019-08-08

摘要: 在线评论文本具有口语化的特点,其评价词缺少对应的评价对象,影响了细粒度情感分析的效果。为此,提出一种利用深度学习自动识别评价对象的方法。设计研究领域的文本序列标注规范,在对评论语料分词后,进行评价词与评价对象的命名实体标注,得到单词序列、词性序列和标注序列。将单词序列、词性序列转为神经网络语言模型的词向量,并用循环神经网络进行训练,采用条件随机场(CRF)输出评价对象标签,得到缺失的评价对象。实验结果表明,与单一CRF模型相比,BiLSTM+CRF模型和BiGRU+CRF模型的识别效果较好,BiGRU+CRF模型的F1值最高可达0.84。

关键词: 隐性评价对象, 隐性特征, 深度学习, 循环神经网络, 条件随机场, 命名实体识别

Abstract: Online review text has the characteristics of colloquialism,and its evaluation words lack corresponding evaluation objects,which affects the effect of fine-grained sentiment analysis.Therefore,a method of automatic identification of evaluation objects based on deep learning is proposed.The specification of text sequence annotation in the research field is designed,the Named Entity Recognition (NER) of the evaluation word,as well as the evaluation object after the comment word segmentation is annotated,and the word sequence,the word part sequence and the tagging sequence are achieved.Transfer the word sequence and the word part sequence to the word vectors of the neural network language model,and train them with Recurrent Neural Network (RNN).Conditional Random Field(CRF) is used to output the tag of the evaluation object,and the evaluation object is achieved as well.Experimental results show that compared with single CRF model,BiLSTM+CRF model and BiGRU+CRF model have better recognition effects and the F1 value of BiGRU+CRF model can reach up to 0.84.

Key words: implicit evaluation object, implicit feature, deep learning, Recurrent Neural Network(RNN), Conditional Random Field(CRF), Named Entity Recognition(NER)

中图分类号: