基于层次嵌入的方面抽取模型  

Aspect Extraction Model Based on Hierarchical Embedding

作　　者：刘漳辉[1,2,3] 肖顺鑫 郑建宁 郭昆 LIU Zhanghui;XIAO Shunxin;ZHENG Jianning;GUO Kun(College of Mathematics and Computer Sciences,Fuzhou University,Fuzhou 350116;Fujian Provincial Key Laboratory of Network Computing and Intelligent Information Processing,Fuzhou 350116;Key Laboratory of Spatial Data Mining&Information Sharing,Ministry of Education,Fuzhou 350116;Power Science and Technology Corporation State Grid Information&Telecommunication Group,Fuzhou 350003)

机构地区：[1]福州大学数学与计算机科学学院,福州350116 [2]福建省网络计算与智能信息处理重点实验室,福州350116 [3]空间数据挖掘与信息共享教育部重点实验室,福州350116 [4]国网信通亿力科技有限责任公司,福州350003

出　　处：《计算机与数字工程》2020年第5期1121-1130,共10页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:61300104,61300103,61672158);福建省高校杰出青年科学基金项目(编号:JA12016);福建省高等学校新世纪优秀人才支持计划项目(编号:JA13021);福建省杰出青年科学基金(编号:2014J06017,2015J06014);福建省科技创新平台计划项目(编号:2009J1007,2014H2005);福建省自然科学基金项目(编号:2013J01230,2014J01232);福建省高校产学合作项目(编号:2014H6014,2017H6008)资助。

摘　　要：方面抽取旨在抽取评论文本中观点持有者所评价的实体属性,是细粒度情感分析的一项重要基本任务。现有的研究大多基于规则或传统机器学习模型,具有简单易行和较高性能等优点,但需要花费较多的精力来人工构建规则模板或者特征工程。为了提高模型自动化,提出一种基于层次嵌入的方面抽取模型。首先,对原始语料执行多阶段的预处理操作;然后,使用字符层次的嵌入和双向循环神经网络获得词的高层次特征;最后,通过级联词嵌入与字符嵌入特征以作为词层次双向循环神经网络的输入,获得最终标注结果。实验结果表明,该模型明显优于基于规则和传统机器学习模型的方法,也优于单层神经网络模型。Aspect extraction aims to extract the aspects of the entity that the opinion holder has expressed in the opinioned text.It’s an important task for fine-grained sentiment analysis.Most of the existing researches are based on rules or traditional machine learning models which has the advantages of simple and high performance.But these methods require much effort to build rule templates or feature engineering artificially.In order to improve model automation,an aspect extraction model based on hierarchical embedding is proposed.First,a multistage preprocessing operation is performed on the original corpus.Second,the high-level features of the word are obtained by using the character-level embeddings and bidirectional recurrent neural network.Finally,the input of word-level bidirectional recurrent neural network by cascading the word embeddings and character embeddings feature are obtained and the label of words are tagged.The experimental results show that the model is not only superior to the rule-based methods and the methods based on traditional machine learning models but also superior to the single layer neural network model.

关 键 词：方面抽取 层次嵌入 循环神经网络 情感分析 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]