作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2024, Vol. 50 ›› Issue (5): 83-90. doi: 10.19678/j.issn.1000-3428.0066475

• 人工智能与模式识别 • 上一篇    下一篇

基于多粒度图与注意力机制的半监督短文本分类

游奔, 李晓红, 姚锦, 冯绍杰   

  1. 西北师范大学计算机科学与工程学院, 甘肃 兰州 730070
  • 收稿日期:2022-12-08 修回日期:2023-06-30 发布日期:2023-09-05
  • 通讯作者: 李晓红,E-mail:xiaohongli@nwnu.edu.cn E-mail:xiaohongli@nwnu.edu.cn
  • 基金资助:
    国家自然科学基金(61862058,61967013);甘肃省自然科学基金(20JR10RA076);甘肃省高校产业支撑项目(2022CYZC11)。

Semi-supervised Classification for Short Text Based on Multi-grained Graphs and Attention Mechanism

YOU Ben, LI Xiaohong, YAO Jin, FENG Shaojie   

  1. College of Computer Science and Engineering, Northwest Normal University, Lanzhou 730070, Gansu, China
  • Received:2022-12-08 Revised:2023-06-30 Published:2023-09-05
  • Contact: 李晓红,E-mail:xiaohongli@nwnu.edu.cn E-mail:xiaohongli@nwnu.edu.cn

摘要: 短文本语义稀疏模糊、蕴含信息不足、表达不规则等缺陷给短文本分类任务带来了极大的挑战,且现有短文本分类方法通常忽略词项间的交互信息,不能充分挖掘隐含的语义信息,导致分类效率低下。针对上述问题,提出一种基于多粒度图与注意力机制的半监督短文本分类模型MgGAt。该模型在词粒度和文本粒度基础上构建2种类型的图,通过充分挖掘语义信息完成分类任务。首先构建词级图,捕获词嵌入,进而学习得到文本特征表示。在词级图上引入跳内注意力和跳间注意力,从多种语义角度有效提取词项间隐含的高阶信息,捕获语义丰富的词嵌入。同时依据词级子图的特点设计池化策略,聚合词嵌入,学习文本表征。其次构建文本级图,借助部分已知的标签信息,利用图神经网络的优势,在图上执行标签传播和推理,完成半监督短文本分类任务。在4个公开数据集上的实验结果表明,与基线模型相比,MgGAt模型的短文本分类精确率平均提升了1.18个百分点,F1值平均提升了1.37个百分点,具有更好的分类性能。

关键词: 短文本分类, 半监督分类, 图神经网络, 注意力机制, 多粒度图

Abstract: Sparse and fuzzy semantics, insufficient information, and irregular expressions in short texts pose great challenges to short text classification tasks. Moreover, the existing short text classification methods ignore the interactive information between terms, and implicit semantics cannot be fully exploited; therefore, they are classified inefficiently. To address these problems, a semi-supervised short text classification method based on multi-grained graphs and attention mechanism, named MgGAt, is proposed. Two types of graphs are constructed based on word and text granularities, and semantic information is fully mined to perform classification task. First, the model builds a word-level graph, captures word embeddings, and learns the feature representations of a short text. Specifically, intra- and inter-hop attention are introduced on a word-level graph to effectively extract high-order information from various semantic perspectives that are hidden in word terms and obtain word embeddings with rich semantics. Simultaneously, a pooling strategy is designed according to the characteristics of the word embeddings, which are aggregated into text vectors. Thereafter, a text-level graph is constructed, and with the help of part of the labeled information, the advantage of the Graph Neural Network(GNN) is used to perform label propagation and reasoning on the graph to achieve semi-supervised short text classification. Experimental results on four public datasets demonstrate that, compared with baseline models, the classification accuracy and F1 value of the proposed MgGAt increased by 1.18 and 1.37 percentage points respectively, on average, resulting in better classification performance.

Key words: short text classification, semi-supervised classification, Graph Neural Network(GNN), attention mechanism, multi-grained graph

中图分类号: