《计算机应用》唯一官方网站

• •    下一篇

自适应跨度特征的细粒度意见元组提取

陈林颖1,2,刘建华1,2,孙水华1,2,郑智雄1,2,林鸿辉1,2,林杰1,2   

  1. 1. 福建工程学院 计算机科学与数学学院,福州 350118;
    2. 福建省大数据挖掘与应用技术重点实验室(福建工程学院),福州 350118

  • 收稿日期:2022-04-14 修回日期:2022-06-06 接受日期:2022-06-09 发布日期:2022-07-01 出版日期:2022-07-01
  • 通讯作者: 刘建华
  • 基金资助:
    国家自然科学基金;福建省自然科学基金项目;福州市科技创新平台项目

Aspect-oriented fine-grained opinion extraction with adaptive span feature#br#

  • Received:2022-04-14 Revised:2022-06-06 Accepted:2022-06-09 Online:2022-07-01 Published:2022-07-01
  • Contact: jianhua jianhuaLiu

摘要: 面向方面的细粒度意见提取(AFOE)以意见对的形式从评论中提取方面词和意见词,或者再提取方面词的情感极性,形成意见三元组。针对现有研究方法忽略了意见对与上下文的相关性问题,提出一种自适应跨度特征的网格标记方案(ASF-GTS)模型。首先,利用BERT模型获得句子特征表示;然后,采用自适应跨度特征策略方法加强意见对局部上下文的联系;其次,通过网格标记方案将意见对提取转化为统一的网格标记任务;最后,使用特定的解码策略生成对应的二元组或三元组。在适应于意见元组提取任务的AFOE的四个基准数据集上进行了实验,与GTS-BERT模型相比,所提模型在意见对和意见三元组任务的F1分值分别提高了7.30%、5.10%、4.41%、2.42%和4.17%、4.27%、6.61%、2.62%。实验结果表明,所提模型能够有效保留意见对与上下文的情感联系,更为精确提取意见对及其情感极性。

关键词: 网格标记方案, 方面词, 意见词, 意见对提取, 意见三元组提取, 面向方面的细粒度意见抽取

Abstract: Aspect-oriented Fine-grained Opinion Extraction (AFOE) extracts aspect terms and opinion terms from review in the form of opinion pairs or additionally extracting sentiment polarity of aspect term to form opinion triplet. Aiming at the problem of neglect correlation between the opinion pair and context, an Adaptive Span Feature for Grid Tagging Scheme (ASF-GTS) model was proposed. First, Bidirectional Encode Representation from Transformers (BERT) model was used to obtain the feature representation of the sentences. Then, the correlation between the opinion pair and local context was enhanced by the way of adaptive span feature. Next, opinion pair extraction was transformed into a uniform grid tagging task by GTS. Finally, opinion pair or opinion triplet were generated by the designed decoding algorithm. Experiments are launched on the four standard of AFOE datasets which adapt to the task of opinion tuple extraction. Compare with GTS-BERT model, F1 score of the proposed model on the opinion pair and opinion triplet tasks are improved by 7.30%,5.10%,4.41%,2.42% and 4.17%,4.27%,6.61%,2.62%, respectively. Experimental results show that the proposed model can effectively reserve sentiment correlation between the opinion pairs and context, and more accurately extract opinion pair and their sentiment polarity.

Key words: Keywords: Grid Tagging Scheme(GTS), aspect terms, opinion terms, Opinion pair extraction(OPE), Opinion Triplet Extraction(OTE), Aspect-oriented Fine-grained Opinion Extraction(AFOE)

中图分类号: