Abstract:

For the existing multi-label classification algorithm has ignored the endogenous relationship between the labels, In this paper, the multi-label classification problem is converted into a sequence generation problem, and the symbiotic relationship between the labels is fully considered. Based on the Seq2Seq model, text features are extracted from two dimensions：word level and semantic level. By improving the feature extraction module, encoder structure, mixed attention mechanism, and decoder prediction part, a multi-label classification algorithm based on multi-level features and mixed attention mechanism is proposed. The effectiveness of the algorithm is verified on the three data sets of Zhihu, RCV1-V2 and AAPD and compared with existing algorithms. The proposed algorithm is superior to other algorithms in F1 value, recall rate and Hamming loss.

Key words: multi-label classification, multi-level features, mixed attention

摘要：

针对现有多标签分类算法忽略了标签之间的内生关系，将多标签分类问题转化为序列生成问题，充分考虑标签之间的共生关系，以Seq2Seq模型为基础，从词语级别和语义级别两个维度提取文本特征，通过对特征提取模块、编码器结构、混合注意力机制、解码器预测部分的改进，提出了基于多级特征和混合注意力机制的多标签分类算法。在Zhihu、RCV1-V2和AAPD三个数据集上进行算法有效性验证并与现有算法对比，提出的算法在F1值、召回率和汉明损失三个指标上均优于其他算法。

关键词: 多标签分类, 多级特征, 混合注意力