Fusing Attribute Type Features for Attribute Value Extraction from Product via Question Answering
Miaobo Hu, Jun Xiao, Yunfei Liu, Weihu Guo, Xipeng Fan · 2022
Extracting attribute values from product titles is a crucial e-commerce task. Previous attribute extraction method was insufficient because the exist dataset lacked attribute type information. To overcome these obstacles and promote product attribute research, we first improve the AliExpress dataset with appending entity type information to better facilitate research on product attribute value extraction. Then, we propose an Attribute Type Attention via Question Answering (ATAQA) approach to fully utilize the attribute type information for attribute value extraction, which captures semantic interaction of attribute and attribute type with a sentence and get the vectors containing attributes, attribute types and sentence features. We combine the attribute type vector and the attribute vector to get the processed vector, and then we use the self-attention layer connect the attribute vector and the processed vector, finally export a sentence-attribute-comprehension representation to the CRF for final tagging. The proposed approach outperforms previous best methods by a significant margin, as shown by the experimental results. Our Data is available at https://github.com/wandugu/AE-improved.