English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Aligning Query Representation with Rewritten Query and Relevance Judgments in Conversational Search (CIKM 2024)

Forum topic · 小凯 · 2026-07-05

Summary

This CIKM 2024 paper addresses query representation learning in conversational search, where user queries are often context-dependent, ambiguous, or underspecified. The work proposes aligning the representation of the current conversational query with both the rewritten (self-contained) query and relevance judgments, so that dense retrievers can better capture user intent across multi-turn interactions. By treating query rewriting signals and relevance feedback as alignment targets during training, the approach aims to improve retrieval effectiveness on conversational benchmarks without relying solely on expensive rewrite generation at inference time. The paper situates itself within the broader shift toward LLM-era conversational IR, where multi-turn context modeling, retrieval efficiency, and intent understanding must be balanced. Reported evaluation follows standard conversational search protocols with ranking metrics such as nDCG and MRR against BM25, dense retrieval, and rewrite-based baselines; exact numerical results should be verified against the original PDF. Source: https://dl.acm.org/doi/abs/10.1145/3627673.3679534

Aligning Query Representation with Rewritten Query and Relevance Judgments in Conversational Search (CIKM 2024)

  • Venue: CIKM 2024
  • Link: https://dl.acm.org/doi/abs/10.1145/3627673.3679534
  • Topic area: Multi-turn / conversational search, dense retrieval
  • Key points

  • Targets a core problem in conversational search: user queries in later turns are often elliptical and context-dependent, making direct dense retrieval unreliable.
  • Proposes aligning the current query's representation with the rewritten query (a self-contained reformulation) and with relevance judgments, using these as supervision/alignment signals for the query encoder.
  • The goal is to make dense retrievers robust to conversational context without depending entirely on an explicit query rewriting module at inference time.
  • Background

    In large-scale conversational search systems, traditional pipelines separate retrieval, ranking, and generation, which complicates multi-turn intent understanding. This work addresses how to encode conversational queries so that retrieval remains effective across turns, in line with the broader move toward LLM-assisted conversational IR.

    Method outline

    The typical pipeline follows: problem formulation → model design → training procedure → inference.

    1. Input & representation: encode the conversational query (with dialogue history) into a dense embedding. 2. Alignment objectives: train the representation to be close to the rewritten query representation and consistent with relevance labels. 3. Learning strategy: supervised/contrastive training using rewrite supervision and relevance judgments. 4. Inference: standard dense retrieval over a prebuilt index, avoiding costly per-turn rewriting.

    Evaluation

  • Datasets: standard conversational search benchmarks (per the original paper's tables).
  • Metrics: ranking quality measures such as nDCG, MRR, Recall@k.
  • Baselines: BM25, dense retrieval without conversational modeling, and rewrite-based pipelines.
  • Exact numerical results are not reproduced here; consult the original PDF before citing quantitative claims.

    Takeaways

    1. Alignment signals from query rewrites and relevance judgments can compensate for noisy conversational context in dense retrieval. 2. Reducing reliance on explicit rewriting improves latency and system simplicity in multi-turn search. 3. Open questions include cross-domain generalization, evaluation of multi-turn consistency, and the trade-off between rewrite quality and retrieval cost.

    Related entries

  • CHIQ: Contextual History Enhancement for Improving Query Rewriting
  • Few-Shot Conversational Dense Retrieval (SIGIR 2021)
  • Beyond Whole Dialogue Modeling: Contextual Disentanglement for Conversational Understanding
  • Reference

  • Original paper: *Aligning Query Representation with Rewritten Query and Relevance Judgments in Conversational Search*, CIKM 2024. ACM DL

Tags

#conversational-search#dense-retrieval#query-rewriting#query-representation#information-retrieval#cikm-2024#multi-turn-interaction

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208778