Research On Multi-label Text Classification For E-commerce Reviews

With the arrival of the era of big data,The platform of e-commerce has facilitated people's life.Meanwhile,with the development of e-commerce platforms,a large amount of user review information has exploded.These reviews contain the real quality of a large number of products,user experience and logistics service efficiency.Therefore,mining the critical information in these reviews is of great importance to the supervision of manufacturers.Multi-label text classification technology is introduced to extract label information from e-commerce reviews so as to provide data basis for product analysis and supervision.Multi-label text classification technology associates text to the most relevant set of label sets that can be used to quickly categorize,search,and analyze text.There are many difficulties and challenges in the research of multi-label text classification in e-commerce review scenario : review data is prone to semantic complexity and information redundancy because the text is too long,so how to extract effective feature representation from the text is crucial;Product labels contain rich semantic information,and how to use this information to guide model interaction is a major difficulty.There are relationships between product labels,and how to model and leverage these relationships is also a challenge.In view of the difficulties and challenges in the above research,the main work of this thesis is as follows:1.To solve the problems of semantic complexity and information redundancy in long texts that cannot be effectively solved by existing researches,a method based on inter-word relations is proposed.Firstly,word co-occurrence information is extracted by gated graph neural network,and word order information is mined by bidirectional short and long time memory network.Finally,the self-attention mechanism is used to capture keywords with different labels and form text feature representations.Experimental results show that the proposed method outperforms the benchmark model in performance,and the effectiveness of each module is verified through ablation,sensitivity analysis and visualization of sample.2.To solve the problem that existing researches fail to effectively use label semantic information for interaction,a method based on label semantics is proposed.First by message passing mechanism from label co-occurrence matrix to enrich the co-occurrence information to the expression of vector,and then through the attention mechanism with supervision from the text learn each label keywords,combined with labels semantic actively focusing on each label keywords,finally combining these two keywords generated text characteristic representation of each label.Experimental results show that the proposed method outperforms the benchmark model in performance,and the effectiveness of each module is verified and demonstrated through ablation experiments and visualization.3.To solve the problem of the failure of existing researches to effectively model and utilize the relationship between labels in e-commerce reviews,a method based on the label selection mechanism is proposed.Firstly,the most discriminant labels are selected,and the relationship between labels is utilized by the model reuse mechanism.Finally,all classifiers are integrated according to the effect of the mechanism and the generation of discriminant labels is influenced.Experimental results show that the proposed method outperforms the benchmark model in performance,and the effectiveness of each module is verified by sensitivity analysis.