LLM-based Long-tail Query Rewriting in Taobao Search (WWW 2024)

Forum topic · 小凯 · 2026-07-05

Summary

This WWW 2024 industry paper from Taobao (Alibaba) presents a large language model (LLM) based approach to rewriting long-tail queries in e-commerce search. Long-tail queries are rare, ambiguous, or poorly phrased user queries that traditional query rewriting methods handle poorly due to sparse behavioral data. The authors leverage LLMs to rewrite such queries into clearer, more retrievable forms that improve recall and relevance in Taobao's production search engine. The work addresses practical industrial constraints such as hallucination control, latency, and online deployment, validating the approach through offline evaluation and online A/B testing on Taobao search traffic. The paper is a representative example of applying generative LLMs to query understanding in large-scale commercial search systems, and it connects to related research on query expansion, conversational query rewriting, and LLM-based retrieval-augmented generation. Readers interested in production-grade query rewriting, e-commerce search relevance, and the gap between academic benchmarks and real-world search systems will find the deployment lessons particularly useful. Full quantitative results are available in the original ACM paper (DOI: 10.1145/3589335.3648298).

LLM-based Long-tail Query Rewriting in Taobao Search (WWW 2024)

Overview

This paper, published at The Web Conference (WWW) 2024, describes how Taobao Search (Alibaba) uses large language models (LLMs) to rewrite long-tail queries — rare, ambiguous, or awkwardly phrased queries that conventional rewriting systems struggle with due to insufficient behavioral logs.

Source: ACM Digital Library — DOI 10.1145/3589335.3648298
Venue: WWW 2024 (industry track)
Domain: Query Understanding / E-commerce Search

Key Points

Problem: Long-tail queries suffer from sparse click and purchase data, so traditional statistical or learning-based query rewriting methods generalize poorly. Poor rewrites directly hurt retrieval recall and search relevance.
Approach: An LLM is used to rewrite long-tail queries into forms that better match the product catalog, leveraging the model's world knowledge and semantic understanding rather than relying solely on historical query–item behavioral pairs.
Industrial constraints: The system design accounts for hallucination control, deployment latency, and integration with Taobao's production retrieval and ranking stack.
Evaluation: Offline assessment of rewriting quality plus online A/B testing on live Taobao search traffic to measure impact on relevance and business metrics.

Positioning in the Literature

The paper sits at the intersection of several research threads:

Query expansion and rewriting (e.g., Aligned Query Expansion, LLM-trained query expansion)
Generative conversational query rewriting (e.g., Few-Shot Generative Conversational Query Rewriting, SIGIR 2020)
Hierarchical query classification in e-commerce search (WWW 2024)
Broader LLM-based retrieval-augmented generation (RAG) and agentic search, where rewriting is one step in a retrieve → rerank → generate pipeline

Takeaways for Practitioners

1. LLMs provide strong semantic priors for low-resource (long-tail) query understanding where behavioral data is absent. 2. Deployment requires guarding against hallucinated rewrites that could introduce irrelevant items. 3. Offline relevance gains must be validated with interleaving or A/B experiments, since offline metrics often diverge from online user satisfaction. 4. Latency and per-query cost are hard constraints in production search; distillation or small-model routing may be needed.

References

Original paper: *Large Language Model based Long-tail Query Rewriting in Taobao Search*, WWW 2024. https://dl.acm.org/doi/abs/10.1145/3589335.3648298

*Note: Detailed quantitative results should be verified against the original PDF; this page summarizes the paper based on its abstract and public metadata.*

LLM-based Long-tail Query Rewriting in Taobao Search (WWW 2024)

Summary

LLM-based Long-tail Query Rewriting in Taobao Search (WWW 2024)

Overview

Key Points

Positioning in the Literature

Takeaways for Practitioners

References

Tags