Multi-label text classification task can be modeled as a text sequence to label sequence mapping task.However,existing sequence-to-sequence (Seq2Seq)models only extract coarse-grained text-level representations from noisy texts,ignoring fine-grained interaction cues between labels and words,leading to class understanding bias.In this regard,a label semantic interaction Seq2Seq model based on encoder-decoder structure is proposed.In the text semantic extraction stage,a gating mechanism is used to fuse coarse-grained text-level representations and fine-grained interaction cues,and finally,a class understanding corrected text representation is obtained.On two standard datasets,the experimental results compared with six algorithms such as LEAM,LSAN and SGM,and the results show that the model are significantly improved in two main evaluation indicators.
关键词
多标签文本分类/序列到序列/自适应门/多头注意力/标签嵌入
Key words
multi-label text classification/sequence to sequence (Seq2Seq )/adapative gate/multi-head attention/label embedding
维普
万方数据