Enhanced Product Embedding with Sememe for Product Search

Product representation plays a crucial role in downstream applications. Content-based approaches rely on language models to represent products. However, interpreting products based solely on their titles or properties, which are often brief with limited information, can be challenging. Language models may struggle to perform well in such sparse semantic contexts. To address the problem, this paper proposes a novel approach to enhance the product representation with the Sememe tree, which is constructed by the FP-Sememe-Tree construction method. Those sememe tree vectors are encoded with a conventional language model and fed into the Sememe-TreeLSTM model to reveal deeper meanings. Experimental results show that our approach to product representation learning significantly enhances the semantic features for product search.