Abstract: In order to realize automatic classification of commodities by leveraging text titles of commodities, a commodity classification model high-level feature fusion (HFF) based on high-level feature fusion is proposed. Firstly, a char embedding and word embedding based low-level feature representation method for the text title is proposed. Then a stronger feature expression of the commodity title structure can be obtained. Secondly, a joint self-attention mechanism, convolutional neural network, and channel attention are proposed to enhance the low-level features and obtain high-level enhancement features of the text title. Finally,by fusing the high-level enhancement features of the word embedding and the char embedding of the text, a comprehensive feature of the text title of the commodity is finally obtained and used for the commodity classification. Experiments are conduct on the dataset of the commodity titles. The experiments show that the classification accuracy of HFF for the third-level commodity can reach 84.348%. In addition, the recall and the F1 value of the HFF reach 47.8% and 49.4%, respectively, which is superior to the existing advanced short text classification method that can be used for the commodity text titles classification.

Key words: commodity classification, short text classification, feature fusion, feature enhancement, attention