Curriculum Contrastive Context Denoising for Few-shot Conversational Dense Retrieval (SIGIR 2022)

Forum topic · 小凯 · 2026-07-05

Summary

This SIGIR 2022 paper addresses few-shot conversational dense retrieval, a setting where a dense retriever must be trained with limited in-domain data while conversational query rewriting introduces noisy, irrelevant context that harms retrieval quality. The forum entry summarizes the work's placement within conversational search research: the system tackles the problem of multi-turn context denoising, using a contrastive learning framework combined with a curriculum strategy that progresses from easier to harder denoising examples. This design aims to teach the retriever to distinguish informative turns from distracting ones in a dialogue history when forming dense query representations. The original ACM record is available at https://dl.acm.org/doi/abs/10.1145/3477495.3531961. Note: this forum post is a structured stub built from metadata; quantitative results, exact datasets, and ablations should be verified against the original PDF. Related topics include conversational query rewriting, dense retrieval, contrastive representation learning, and LLM-era retrieval-augmented systems.

Curriculum Contrastive Context Denoising for Few-shot Conversational Dense Retrieval (SIGIR 2022)

Overview

This entry covers the SIGIR 2022 paper "Curriculum Contrastive Context Denoising for Few-shot Conversational Dense Retrieval", part of the forum's Conversational Search section.

Venue: SIGIR 2022
Source: https://dl.acm.org/doi/abs/10.1145/3477495.3531961
Topic area: Conversational search / few-shot dense retrieval

Problem Setting

Conversational dense retrieval requires encoding multi-turn dialogue context into a single query representation. Two core challenges discussed in this entry:

1. Few-shot supervision: In-domain training data is scarce, so the retriever must generalize from limited examples. 2. Context noise: Dialogue histories contain turns that are irrelevant to the current information need; naively concatenating all turns degrades the dense representation.

Approach (as described by the entry)

The paper's title indicates a combination of two techniques:

Contrastive context denoising: A contrastive learning objective that teaches the retriever to distinguish informative context from distractor turns when building the query embedding.
Curriculum learning: Denoising tasks are ordered from easy to hard, progressively training the model to handle noisier contexts.

Experimental Notes

The forum entry is a metadata-based stub. Specific datasets, baselines, and metric values (nDCG@10, MRR, Recall@k, etc.) are not reproduced here; consult the original PDF via the ACM link above before citing quantitative results.

Context in the Field

Conversational dense retrieval sits within the broader evolution of neural IR: from BM25 and BERT cross-encoders, through bi-encoder dense retrieval, to retrieval-augmented generation (RAG) and agentic search with LLMs. This work targets the retrieval stage of that pipeline, where multi-turn intent understanding and robust representation learning remain critical.

Related Entries

A Survey of Conversational Search (Sep 2025, ACM)
Engineering Conversational Search Systems: A Review of Applications
ChatRetriever: Adapting Large Language Models for Generalized and Robust Retrieval
CoSearchAgent: A Lightweight Collaborative Search Agent

Disclaimer

This post summarizes the paper based on its title, venue, and public metadata. Readers should verify all technical details and results against the original publication.

Curriculum Contrastive Context Denoising for Few-shot Conversational Dense Retrieval (SIGIR 2022)

Summary

Curriculum Contrastive Context Denoising for Few-shot Conversational Dense Retrieval (SIGIR 2022)

Overview

Problem Setting

Approach (as described by the entry)

Experimental Notes

Context in the Field

Related Entries

Disclaimer

Tags