| 摘 要: 方面级情感三元组抽取(Aspect Sentiment Triplet Extraction,ASTE)旨在识别句子中的方面词、观点词及其对应的情感极性。针对现有模型语义理解能力和泛化性不佳的问题,提出了基于对抗训练和片段级别的双向情感三元组抽取模型,预测情感极性时考虑了片段之间的相互作用,使用双仿射分类分析它们之间的情感依赖关系。为了保证上下游任务的一致性,通过SpanBERT(Improving Pre-training by Representing and Predicting Spans)模型得到词向量表征,使用BiGRU(Bidirectional Gated Recurrent Unit)网络进行特征提取,提出使用FGM(Fast Gradient Method)对抗训练算法提高模型的鲁棒性和泛化能力。相较于基线模型,基于对抗训练和片段级别的双向情感三元组抽取模型在4个数据集上的F1分数分别提升了0.85百分点、1.42百分点、2.27百分点和2.85百分点,表明了本文所提方法的可行性。 |
| 关键词: 情感分析;情感三元组抽取;双仿射;片段;对抗训练 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 杭州市重大科技创新项目(2022AIZD0145);“尖兵”“领雁”研发攻关计划(2023C01041) |
|
| Bidirectional Aspect Sentiment Triplet Extraction Model Based on Adversarial Training and Span Leve |
|
ZHOU Yi1, MA Hanjie2, XU Yongen3, ZONG Jiamin3, LI Shaohua1
|
(1.School of Inf ormation Science and Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China;
2.School of Computer Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China;
3.Codvision Technology, Hangzhou 311199, China)
964007968@qq.com; mahanjie@zstu.edu.cn; 273994287@qq.com; zongjiamin@codvision.com; lshua98@163.com
|
| Abstract: Aspect Sentiment Triplet Extraction (ASTE) aims to identify aspect terms, opinion terms, and their corresponding sentiment polarity in sentences. To address the issues of poor semantic understanding and generalization in existing models, this paper proposes a bidirectional ASTE model based on adversarial training and span level. This model takes into account the interactions between fragments when predicting sentiment polarity and utilizes a biaffine classification to analyze the emotional dependencies between them. To ensure consistency across upstream and downstream tasks, the model employs SpanBERT (Improving Pre-training by Representing and Predicting Spans) for word vector representation and uses a BiGRU (Bidirectional Gated Recurrent Unit) network for feature extraction. Furthermore, Fast Gradient Method (FGM) adversarial training algorithm is proposed to enhance the robustness and generalization ability of the model. Compared to baseline models, the proposed model achieves F1 score improvements of 0.85 percentage points, 1. 42 percentage points, 2. 27 percentage points, and 2. 85 percentage points across four datasets, respectively, which demonstrates the feasibility of the proposed model. |
| Keywords: sentiment analysis; sentiment triplet extraction; biaffine; span; adversarial training |
