Dynamic Q&A multi-label classification based on adaptive multi-scale feature extraction

In community question answering (CQA), questioners use labels for question and answer (Q&A) classification when asking questions. Since the answerers do not have the same understanding and perspective of the question, the original labels cannot accurately reflect the Q&A categories with constantly given answers. Therefore, this paper proposes a dynamic Q&A multi-label classification approach based on adaptive multi-scale feature extraction. First, global and local semantic features of Q&As are extracted based on bidirectional long short-term memory network and convolutional neural network models, respectively. Second, the label features extraction and fusion method is proposed. The semantic features of the labels are extracted, the label structure graph based on horizontal and vertical dependencies is constructed, and the label structure and semantic features are fused using the graph attention network integrating multi-head self-attention mechanism. Afterward, the label-aware local features of Q&As are constructed using the attention mechanism and fused with global features of Q&A using the multi-head self-attention, thereby multi-scale fusion classification features of Q&A are established. Then, to adaptively extract the core multi-scale fusion features, a multi-objective feature selection model is established and an improved binary multi-objective Sinh Cosh optimizer algorithm is proposed to solve the model. Finally, a classification prediction layer based on a multilayer perceptron is constructed to obtain the multi-label classification results of Q&A documents. The experimental results based on real Q&A data show the superior performance of the proposed method and validate the effectiveness of the proposed four modules.