A two-stage e-commerce search matching model incorporating contrastive learning and course-based hard negative example sampling

Abstract

Text matching is a fundamental task in natural language processing. To address the short and ambiguous search statements in e-commerce domain, the complexity of headlines and the expensive manual annotation samples, this paper proposes a two-stage "vectorized retrieval + refined ranking" text matching model with a mixture of contrastive learning and course-based hard negative example sampling. By using supervised learning data augmentation, domain pre-training, comparative learning and hard case sampling to assist in ranking, this work achieves an MRR@10 value of 0.3890 in the test set of the 2022 "Ali Lingjie" E-Commerce Search Algorithm Competition, ranking second, demonstrating the effectiveness of the model.