基于位置特征与语义分割的方面级情感三元组抽取方法  

Aspect-Level Sentiment Triple Extraction based on Location Features and Semantic Segmentation

作　　者：杨爱琳 吴震 王燚 黄超 杨渭平 YANG Ailin;WU Zhen;WANG Yi;HUANG Chao;YANG Weiping(College of Cyberspace Security,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学网络空间安全学院,四川成都610225

出　　处：《成都信息工程大学学报》2025年第2期143-150,共8页Journal of Chengdu University of Information Technology

基　　金：四川省科技计划资助项目(2023YFG0292、2021ZYD0011);国家社会科学基金资助项目(23BSH061);四川省社会科学资助项目(SC21B034)。

摘　　要：方面情感三元抽取任务定义为识别句子中的方面术语、情感极性和意见术语。近期一种端到端的网格标注方法有效缓解了流水线框架中误差传播的问题,却忽略了字符间的位置关系以及局部上下文信息,导致模型无法充分挖掘文本中的局部情感特征,影响性能的进一步提升。针对上述问题,提出一种融合位置特征与语义分割的方面情感三元抽取模型。该模型首先通过BERT编码层学习每个单词的上下文表达,同时加入位置编码丰富模型对位置信息的感知。在此基础上,使用语义分割网络捕获字符间的局部依赖,充分建模文本的上下文信息,加强模型的局部建模能力。在Res14、Lap14、Res15和Res16标准数据集上的实验结果表明,相较于基准模型,提出模型的F1指标分别提升2.82、3.8、3.59、3.77个百分点,均取得最优性能,有效证明了所提方法的优越性。Aspect-sentiment triple extraction task is defined as recognizing aspectual terms,sentiment polarity,and opinion terms in a sentence.Recently,researchers have proposed an end-to-end grid labeling method that effectively mitigates the problem of error propagation in pipelined approaches.However,the method ignores the positional relationship between characters and local contextual information,which results in the model not being able to fully mine the local sentiment features in the text,affecting the further improvement of performance.To address the above problems,this study proposes an aspectual sentiment Triple extraction model that fuses positional features with semantic segmentation.In this study,we first learn the contextual expression of each word through the BERT coding layer,and at the same time,we add location coding to enrich the model's perception of location information.On this basis,the local dependencies between characters are captured using semantic segmentation network to fully model the contextual information of the text and strengthen the local modeling ability of the model.The experimental results on four standard datasets,Res14,Lap14,Res15 and Res16,show that compared with the benchmark model,the model proposed in this paper improves the F1 metrics by 2.82,3.8,3.59 and 3.77 percentage points,respectively,and all of them achieve optimal performance,which effectively proves the superiority of the method proposed in this paper.

关 键 词：三元抽取 语义分割 位置编码 情感分析 网格标注 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]