ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2022, Vol. 59 ›› Issue (9): 2003-2014.doi: 10.7544/issn1000-1239.20210699

• 人工智能 • 上一篇    下一篇

面向小样本情感分类任务的弱监督对比学习框架

卢绍帅1,陈龙1,卢光跃1,管子玉2,谢飞3   

  1. 1(西安邮电大学通信与信息工程学院 西安 710121);2(西安电子科技大学计算机科学与技术学院 西安 710071);3(西安电子科技大学前沿交叉研究院 西安 710071) (199509@stu.xupt.edu.cn)
  • 出版日期: 2022-09-01
  • 基金资助: 
    国家自然科学基金项目(61936006,61876144,61876145,62103314);陕西省重点研发计划项目(2021ZDLGY02-06);陕西省自然科学基础研究计划项目(2020JQ-850)

Weakly-Supervised Contrastive Learning Framework for Few-Shot Sentiment Classification Tasks

Lu Shaoshuai1, Chen Long1, Lu Guangyue1, Guan Ziyu2, Xie Fei3   

  1. 1(School of Communications and Information Engineering, Xi’an University of Posts & Telecommunications, Xi’an 710121);2(School of Computer Science and Technology, Xidian University, Xi’an 710071);3(Academy of Advanced Interdisciplinary Research, Xidian University, Xi’an 710071)
  • Online: 2022-09-01
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (61936006, 61876144, 61876145, 62103314), the Key Research and Development Program of Shaanxi (2021ZDLGY02-06), and the Natural Science Basic Research Program of Shaanxi (2020JQ-850).

摘要: 文本情感分类是自然语言处理领域的挑战性研究课题.基于词典的方法和传统基于机器学习方法分别依赖高质量的情感词典和鲁棒的特征工程,而多数深度学习方法的性能则依赖大规模人工标注数据集.幸运的是,不同社交平台用户生成了大量带标签的舆情文本,这些文本可以作为弱标注数据集被用于情感分类任务,但是弱标注数据集中的噪声样本会对训练过程产生负面影响.提出了一种用于小样本情感分类任务的弱监督对比学习(weakly-supervised contrastive learning, WCL)框架,旨在学习海量带噪声的用户标记数据中的情感语义,同时挖掘少量人工标注数据中潜在的类间对比模式.该框架包括2个步骤:首先,设计了一种弱监督预训练策略来削弱噪声数据的影响;其次,在有监督微调阶段引入对比学习策略来捕获少量有标注数据的对比模式.在亚马逊评论数据集上评估了所提出的方法,实验结果表明所提出的方法显著优于其他同类对比方法.在仅使用0.5%(即32个样本)比例的有标注数据集进行微调的情况下,所提出方法的性能依然超出其他深度方法.

关键词: 情感分类, 弱监督学习, 有监督对比学习, 小样本学习, 迁移学习

Abstract: Text sentiment classification is a challenge research topic in natural language processing. Lexicon-based methods and traditional machine learning-based methods rely on high-quality sentiment lexicon and robust feature engineering respectively, whereas most deep learning methods are heavily reliant on large human-annotated data sets. Fortunately, users on various social platforms generate massive amounts of tagged opinioned texts which can be deemed as weakly-labeled data for sentiment classification. However, noisy labeled instances in weakly-labeled data have a negative impact on the training phase. In this paper, we present a weakly-supervised contrastive learning framework for few-shot sentiment classification that learns the sentiment semantics from large user-tagged data with noisy labels while also exploiting inter-class contrastive patterns hidden in small labeled data. The framework consists of two steps: first, we design a weakly-supervised pre-training strategy to reduce the influence of the noisy labeled samples, and then the contrastive strategy is used in supervised fine-tuning to capture the contrast patterns in the small labeled data. The experimental results on Amazon review data set show that our approach outperforms the other baseline methods. When fine-tuned on only 0.5% (i.e. 32 samples) of the labels, we achieve comparable performance among the deep baselines, showing its robustness in the data sparsity scenario.

Key words: sentiment classification, weakly-supervised learning, supervised contrastive learning, few-shot learning, transfer learning

中图分类号: