《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (S2): 28-33.DOI: 10.11772/j.issn.1001-9081.2023040426

• 人工智能 • 上一篇    下一篇

动态微调的模型集成算法Bagging-DyFAS

李龚林(), 范一晨, 米宇舰, 李明   

  1. 西安理工大学 经济与管理学院,西安 710054
  • 收稿日期:2023-04-14 修回日期:2023-06-25 接受日期:2023-06-30 发布日期:2024-01-09 出版日期:2023-12-31
  • 通讯作者: 李龚林
  • 作者简介:李龚林(2002—),男,湖南长沙人,主要研究方向:自然语言处理
    范一晨(2002—),男,陕西西安人,主要研究方向:自然语言处理
    米宇舰(2002—),男,陕西西安人,主要研究方向:自然语言处理
    李明(1973—),女,山东莱芜人,讲师,硕士,主要研究方向:人工智能、自然语言处理。
  • 基金资助:
    陕西省大学生创新创业训练计划项目(S202210700148)

Bagging-DyFAS: model ensemble algorithm with dynamic fine-tuning

Gonglin LI(), Yichen FAN, Yujian MI, Ming LI   

  1. School of Economics and Management,Xi’an University of Technology,Xi’an Shaanxi 710054,China
  • Received:2023-04-14 Revised:2023-06-25 Accepted:2023-06-30 Online:2024-01-09 Published:2023-12-31
  • Contact: Gonglin LI

摘要:

针对单一模型用于文本分类存在的模型体量大,难以适用于舆情信息文本的多元化非规范的表达等问题,提出基于Bagging训练思想的、动态微调和二次加权的模型集成算法(Bagging-DyFAS)。首先,使用自助采样构建的数据集训练弱分类器,使该分类器具有一定的先验知识;其次,依据该分类器在开发集的表现,进行一次动态加权和一次静态加权,并使用得到的一系列权重将模型泛化到无标注的数据上,进一步提升模型在文本分类任务的性能。在所构建的数据集上的实验结果表明,在训练一轮的情况下,相较于基线模型MiniBRT、BRT3和LERT(Linguistically-motivated bidirectional Encoder Representation from Transformer),所提算法的准确率、精确率、召回率和F1值分别至少提升3.6、3.8、1.3和3.2个百分点,实验结果验证了所提算法的有效性。

关键词: 文本分类, 模型集成, 二次加权, 动态加权, 舆情分析, 预训练语言模型

Abstract:

In view of the problems of large model size and difficulty in applying a single model for text classification to diverse and non-normative representations of public opinion information, a model ensemble algorithm based on Bagging-Dynamic Fine-tuning And Secondary weighting (Bagging-DyFAS) was proposed. First, weak classifiers were trained with a dataset constructed by self-sampling, so that some priori knowledge was in the classifiers. Then, they were dynamically weighted once and statically weighted once based on their performance in the development set. Using the obtained series of weights, the models were generalized to unlabeled data, which could further improve the performance of the models in text classification tasks. Experimental results on the constructed test dataset show that, after training for on round, compared to the baseline models MiniBRT,BRT3 and LERT (Linguistically-motivated bidirectional Encoder Representation from Transformer), the proposed algorithm improved the accuracy, precision, recall and F1 value by at least 3.6, 3.8, 1.3 and 3.2 percentage points respectively, validating the effectiveness of the proposed algorithm.

Key words: text classification, model ensemble, quadratic weighting, dynamic weighting, public opinion analysis, pre-training language model

中图分类号: