Reproduce “DSIRM: Learning Query-Bridged Discrete Semantic Identifiers for E-commerce Relevance Modeling” on OpenResearch

About this paper

Bokang Wang, Xing Fang, Mingmin Jin, Jing Wang, Zhentao Song, Guangxin Song, Jianbo Zhu

Alibaba researchers developed the Discrete Semantic Identifier Relevance Model (DSIRM) for e-commerce search, which integrates query-bridged discrete semantic identifiers (SIDs) into ranking models. The framework leverages a supervised hierarchical quantization process for items and generative Large Language Models for query-side SID prediction, leading to improvements of +0.13% in User Click-Through Rate and +0.25% in User Click-To-Conversion Rate in Tmall's production environment.

Key insights

• Explicit query-item interaction supervision, via contrastive learning applied to quantized embeddings, is crucial for learning discriminative discrete semantic identifiers that capture relevance-aware semantic partitions.
• Decoupling item representation learning from the discrete quantization process, by using static pre-trained embeddings, allows the RQ-VAE to focus on discovering more effective hierarchical discrete structures.
• Leveraging generative LLMs for multi-cluster decoding of query SIDs effectively approximates conditional distributions of relevant item clusters, thereby explicitly addressing user intent ambiguity.

Summary by alphaXiv.