基于BERT的端到端方面级情感分析  

End-to-end aspect-level sentiment analysis based on BERT

作　　者：曾凡旭 李旭 姚春龙[1] 范丰龙[1] ZENG Fanxu;LI Xu;YAO Chunlong;FAN Fenglong(School of Information Science and Engineering, Dalian Polytechnic University, Dalian 116034, China;Innovation and Entrepreneurship Education Center, Dalian Polytechnic University, Dalian 116034, China)

机构地区：[1]大连工业大学信息科学与工程学院,辽宁大连116034 [2]大连工业大学工程训练中心,辽宁大连116034

出　　处：《大连工业大学学报》2022年第3期228-234,共7页Journal of Dalian Polytechnic University

基　　金：国家重点研发计划专项项目(2017YFC0821003-3);辽宁省自然科学基金项目(20180550395);辽宁省教育厅青年科技人才“育苗”项目(J2020113).

摘　　要：方面词提取和情感分类是方面级情感分析的两个子任务。传统的方法是将这两个任务以流水线方式进行,这种流水线式工作模式会导致错误的累积而且无法利用两个任务之间的联合信息。传统模型大都使用了Word2vec、GloVe等词向量,无法捕获语义的上下文相关性。为此设计了一种同时包含方面词位置信息与情感极性的标签,将方面级情感分析转化为序列标注问题。利用预训练语言模型BERT自动学习带有上下文信息的特征表示,再通过多种分类器输出预测结果并进行对比。在SemEval-2014 Task4数据集上,本研究模型F1比传统流水线模型高8.07%,比现存最先进的联合模型高5.63%。Aspect word extraction and sentiment classification are two subtasks of aspect-level sentiment analysis.The traditional method is to carry out these two tasks in a pipelined manner,which results in the accumulation of errors and the inability to take advantage of the joint information between the two tasks.In addition,traditional models mostly use word vectors such as Word2vec and GloVe,which cannot capture the contextual relevance of semantics.A label containing both aspect word location information and sentiment polarity was designed,and aspect-level sentiment analysis was transformed into a sequence labeling problem.The pre-trained language model BERT was used to automatically learn feature representations with contextual information,and then the predicted results were outputted by various classifiers for comparison.On the SemEval-2014 Task4 data set,the F1 value of the model is 8.07%higher than the traditional pipeline model,and 5.63%higher than the existing state-of-the-art joint model.

关 键 词：方面级情感分析 端到端 深度学习 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]