Curriculum Contrastive Context Denoising for Few-shot Conversational Dense Retrieval (SIGIR 2022)
Overview
This entry indexes a SIGIR 2022 research paper on few-shot conversational dense retrieval:
- Title: Curriculum Contrastive Context Denoising for Few-shot Conversational Dense Retrieval
- Venue: SIGIR 2022
- DOI: https://dl.acm.org/doi/abs/10.1145/3477495.3531961
- Topic area: Conversational Search
- A Survey of Conversational Search, Sep 2025, ACM
- Engineering Conversational Search Systems: A Review of Applications
- Agentic Conversational Search with Contextualized Reasoning via Reinforcement Learning
- CTR-Guided Generative Query Suggestion in Conversational Search, EMNLP
- ChatRetriever: Adapting Large Language Models for Generalized and Robust Dense Retrieval
- CoSearchAgent: A Lightweight Collaborative Search Agent with Large Language Models
- Original paper: Curriculum Contrastive Context Denoising for Few-shot Conversational Dense Retrieval, SIGIR 2022. <https://dl.acm.org/doi/abs/10.1145/3477495.3531961>
Problem Setting
Conversational dense retrieval requires retrieving relevant documents for the user's current query while resolving coreferences and omissions using multi-turn conversation history. In the few-shot setting, only limited labeled data is available, which makes it hard to train dense retrievers that correctly understand which parts of the dialogue context are relevant and which are noise.
Proposed Approach
Based on the paper title, the method combines two ideas:
1. Context denoising — filtering out noisy, irrelevant utterances from the conversational history so that the query representation is not polluted by unrelated context. 2. Contrastive learning with a curriculum — organizing training examples from easy to hard and applying contrastive objectives to learn better dense representations of conversational queries under few-shot supervision.
Experimental Details
The original post does not reproduce the paper's abstract, dataset details, or result tables. Specific benchmark numbers (e.g., datasets, nDCG/MRR/Recall figures, and baselines) should be verified directly against the official ACM PDF at the DOI link above.