Improving Semantic Matching via Multi-Task Learning in E-Commerce

Semantic matching plays a critical role in an e-commerce search engine, while one of the biggest challenges is the language gap between queries and products. Traditionally, some auxiliary functions, such as the category navigation, are designed to help buyers to clarify their intent. Recently, the advances in deep learning provide new opportunities to bridge the gap, however, these techniques suffer from the data sparseness problem. To address this issue, in addition to the click-through data from buyers, we exploit other types of semantic knowledge from the product category taxonomy and sellers’ behavior. We investigate the correlation between query intent classification and semantic textual similarity, and propose a multi-task framework to boost their performance simultaneously. Moreover, we design a Progressively Hierarchical Classification (PHC) network architecture with the taxonomy to solve the category imbalance problem . We conduct extensive offline and online A/B experiments on a real-world e-commerce platform, and the results show that the proposed method in this paper significantly outperforms the baseline and achieves higher commercial value.