基于Tree LSTM+CRF的属性级观点挖掘  

Tree LSTM+CRF for aspect-level opinion mining

作　　者：赵华 邹若飞 ZHAO Hua;ZOU Ruofei(College of Computer Science and Engineering,Shandong University of Science and Technology,Qingdao,Shandong 266590,China)

机构地区：[1]山东科技大学计算机科学与工程学院,山东青岛266590

出　　处：《山东科技大学学报（自然科学版）》2020年第6期115-122,共8页Journal of Shandong University of Science and Technology(Natural Science)

基　　金：青岛市哲学社会科学规划研究项目(QDSKL1901124);教育部人文社会科学研究青年基金项目(16YJCZH154)。

摘　　要：评价对象与观点内容的提取是观点挖掘中非常重要的任务。本研究提出了一个树结构长短期记忆网络(Tree LSTM)结合条件随机场(CRF)的联合模型抽取评价对象和观点内容。首先对评论句进行依存句法分析,根据句子的依存分析树构建Tree LSTM,并设计树结构下LSTM单元的计算方法;接着将Tree LSTM的输出作为CRF的输入进行序列标注,实现评价对象与观点内容的抽取。最后在SemEval Challenge 2014任务4的数据集上对模型性能进行了验证,评价对象和观点内容抽取结果的平均F1值在餐馆和笔记本电脑领域分别为86.76%、83.22%和79.86%、80.42%,优于现有的评价对象和观点内容抽取方法。实验结果表明,设计的Tree LSTM能很好地学习词语之间的层次关系,同时联合模型有效避免了传统CRF需要构建特征工程的弊端。The extraction of aspect terms and opinion terms is a significant task in opinion mining.In this paper,a tree-structured long short-term(Tree LSTM)memory network combined with Conditional Random fields(CRF)is proposed to extract aspect terms and opinion terms.Initially,dependency parsing of commentary sentences is carried out,Tree LSTM is built according to the dependency parsing tree of sentences,and calculation method of LSTM unit under the tree structure is designed.Afterwards,the output of Tree LSTM is tagged as the input of CRF to realize the extraction of aspect terms and opinion terms.This paper validates the performance of the model on the data set of SemEval Challenge 2014 Task 4.The average F1 scores of aspect terms and opinion terms extraction results are 86.76%,83.22%and 79.86%and 80.42%respectively in restaurants and laptops,which are superior to the existing aspect terms and opinion terms extraction methods.The experimental results show that the Tree LSTM design presented in this paper can learn the hierarchical relationship between words well,and the joint model effectively avoids the drawbacks of traditional CRF which needs to construct feature engineering.

关 键 词：观点挖掘 评价对象抽取 观点内容抽取 树结构长短期记忆网络 条件随机场 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]