基于BERT和句法分析的汽车评论属性级观点挖掘  

Aspect-Level Opinion Mining from Car Reviews Based on the BERT and Syntactic Parsing

作　　者：白云龙 张耿耿 Bai Yunlong;Zhang Genggeng(SAIC Volkszwagen Automotive Co.,Ltd.Shanghai 201805,China)

机构地区：[1]上汽大众汽车有限公司,上海201805

出　　处：《传动技术(中英文)》2024年第4期39-47,共9页Drive System Technique

摘　　要：网络媒体与电商平台的兴起产生了海量的针对汽车的主观评论。对这些评论文本进行观点挖掘对车企改进产品设计和制定竞争策略有重要参考意义。如何在评论中实现细粒度的属性级观点挖掘已成为一个亟待解决的问题。针对属性级观点挖掘中观点抽取和类别分类两个子任务,提出了基于BERT的挖掘方法。首先将BERT序列标注模型和基于依存句法、词性标注的句法分析模型并联,融合生成<评价对象,评价观点>二元组。在此基础上,通过BERT模型对二元组进行属性分类和情感分类。对比试验结果表明,采用BERT的挖掘方法效果显著优于单纯的句法分析方法,而融合了句法分析的BERT模型在召回率的表现上优于单纯的BERT模型,验证了所提方法的可行性和有效性。The rise of online media and e-commerce platforms has generated a large number of subjective comments on cars.The opinion mining of these comments is of great reference significance for automobile enterprises to improve product design and formulate competitive strategies.How to achieve fine-grained aspect-level opinion mining in comments has become an urgent problem.Aiming at the two subtasks of opinion extraction and category classification in aspect-level opinion mining,a mining method based on BERT is proposed.First,the BERT sequence tagging model and the syntactic analysis model based on dependency syntax and part of speech tagging are connected are paralleled to generate the binary<opinion targets,opinion words>.On this basis,BERT model is used to perform attribute classification and sentiment m classification of the binary.The comparative test results show that the methods with BERT work significantly better than syntactic anaiysis,while the performance of the BERT model combined with syntactic analysis is better than that of pure BERT analysis.That validates the feasibility and efficacy of the proposed method.

关 键 词：汽车评论观点挖掘 属性级观点挖掘 BERT 词性标注 依存句法分析 

分 类 号：U462[机械工程—车辆工程]