计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (4): 104-111.DOI: 10.3778/j.issn.1002-8331.2108-0352

• 模式识别与人工智能 • 上一篇    下一篇

基于混合注意力Seq2seq模型的选项多标签分类

陈千,韩林,王素格,郭鑫   

  1. 1.山西大学 计算机与信息技术学院,太原 030006
    2.山西大学 计算智能与中文信息处理教育部重点实验室,太原 030006
  • 出版日期:2023-02-15 发布日期:2023-02-15

Multi-Label Classification of Options Based on Seq2seq Model of Hybrid Attention

CHEN Qian, HAN Lin, WANG Suge, GUO Xin   

  1. 1.School of Computer and Information Technology, Shanxi University, Taiyuan 030006, China
    2.Key Laboratory of Computational Intelligence and Chinese Information Processing of Ministry of Education, Shanxi University, Taiyuan 030006, China
  • Online:2023-02-15 Published:2023-02-15

摘要: 选项多标签分类是高考文学类阅读理解选择题解答任务中的重要一环,对不同标签类型的选项调用不同的答题引擎,可以有效提高选择题答题准确率。由于选项类型复杂多样,一个选项可能有多个类别特征,将其看作多标签分类任务。传统多标签分类算法仅考虑到文本与标签间相关性,忽略了标签间相关性,且选项内部存在着强语义关联性,对最终的标签预测产生较大影响。为了充分利用选项内相关性,提出一种基于混合注意力的Seq2seq模型,同时考虑选项标签间相关性和选项内相关性。采用Bi-LSTM获得选项到标签的相互信息,利用多头自注意力获得选项内关联语义。为获取标签间语义相关性,使用标签嵌入方式进行隐式融合。在高考文学类阅读理解选择题数据集上的实验结果表明,对多种相关性建模能有效提升选项多标签分类精度。

关键词: 阅读理解, 多标签文本分类, 自注意力, 选项相关性

Abstract: The multi-label classification of options is an important part of the task of multiple-choice questions for reading comprehension of literature in college entrance examination (RCL-CEE). It can effectively improve the accuracy of multiple-choice questions by invoking different answering engines for different types of options. Option classification is regarded as a multi-label learning task since an option may have multiple characteristics for the complexity and variety of options. Traditional multi-label classification only considers the correlation between text and label, ignores the correlation between labels, and there exists strong semantic relevance within one option, which has great impact on label prediction. In order to handle these challenges, a hybrid attention based Seq2seq model is proposed, which considers the correlations from the option to the label and internal correlation of an option. Bi-LSTM is used to obtain the mutual information from the option to the label, and the multi-head self-attention is used to obtain the correlations semantics within one option. The label embedding is used to implicitly fuse semantic correlation between labels. Experimental results on the dataset of multiple-choice questions for RCL-CEE show that modeling above correlations can effectively improve the accuracy of options multi-label classification.

Key words: reading comprehension, multi-label text classification, self-attention, option correlation