Research On Multi-Label Classification Method Based On Deep Learning

With the onset of the big data era and the swift evolution of its nearly related artificial intelligence technology,the amount of text information not only shows a trend of leapfrog increment,but also gradually presents multi-labeling,multi-granularity and high complexity.In order to classify and manage text information more efficiently and achieve effective retention and accurate filtering of text content,researchers have started to focus on the most universal multi-label classification techniques in the field of natural language processing in recent years.Deep learning-based multi-label classification methods can automatically assign labels to significant information in text sequences,thus achieving efficient utilization and classification management of text data.In this paper,we provide a comprehensive and specific analysis of the multi-label classification task and propose the following two solutions to address the shortcomings of traditional multi-label classification models:(1)Traditional multi-label classification models either ignore the local semantics or discard the global dependencies of sequences when capturing semantic information in text sequences,which makes the information in the text sequences not fully exploited and thus causes the degradation of label prediction efficiency.We propose a novel sequence-to-sequence learning strategy called "parallel encoding,serial decoding" and utilize it to construct a hierarchical sequence-to-sequence multi-label text classification model.The model employs convolutional neural networks and self-attention as encoders in parallel to extract fine-grained local neighborhood information and global interaction information from the source text.We design a hierarchical decoder to decode and predict labeled sequences.Our approach not only fully considers the interpretable fine-grained information in the source text,but also efficiently uses this information to generate label sequences.We have conducted extensive comparison experiments on three datasets.The results indicate that the proposed model has considerable superiority to state-of-the-art baselines.Moreover,our analysis shows that the model is competitive with the RNN-based Seq2 Seq model and more robust in handling datasets with high label/sample ratios.(2)Since traditional multi-label classification models based on sequence-tosequence(Seq2Seq)architecture,decoders predict labels sequentially in temporal order when making predictions,resulting in a significant loss in their time efficiency.For the MLC task,the output labels are typically unordered from each other.Seq2Seq-based models have the label order fixed during training,which leads to often unstable predictions during testing.We propose a novel semantic-label multigranularity attention(SLMA)model for solving multi-label classification tasks.The model constructs a multi-granularity semantic feature representation of text sequences in terms of local correlation and long-term dependency by stacked expansive convolutional structures in a unified module.Meanwhile,the label representation is updated by directly modeling the correlation between labels using graph attention network(GAT).Subsequently,the semantic feature representations at different levels of granularity are weighted by the designed multi-granularity attention to the labels.Finally,the correct prediction of labels is achieved through a fully connected layer that shares the weights with the label embedding matrix.Experiments shows that the model achieves good performance on all three benchmark datasets.In further exploration experiments,SLMA is found to have better robustness for classification of both high and low frequency labels.