Few-Shot Conversational Dense Retrieval (SIGIR 2021)

Forum topic · 小凯 · 2026-07-05

Summary

This forum entry indexes the SIGIR 2021 research paper 'Few-Shot Conversational Dense Retrieval', published in the ACM Digital Library at https://dl.acm.org/doi/abs/10.1145/3404835.3462856. The paper addresses conversational dense retrieval, where a search system must resolve multi-turn context (such as coreference and ellipsis in follow-up questions) before matching queries to documents. A central challenge is that conversational search datasets are small, making fully supervised training of dense retrievers difficult; the work therefore targets few-shot settings where labeled conversation pairs are scarce. The surrounding forum post places the paper in a curated reading list section on multi-turn interaction, alongside surveys on LLM multi-turn capabilities, query rewriting methods such as CHIQ, and empirical studies of conversational recommender systems. Note that the post itself contains mostly template scaffolding rather than detailed results, so quantitative findings should be verified against the original PDF. This entry is useful for readers tracing the lineage from early conversational dense retrieval toward modern LLM-based agentic search and RAG systems, and for understanding how few-shot learning was framed in conversational IR prior to the LLM era.

Few-Shot Conversational Dense Retrieval (SIGIR 2021)

Overview

This forum entry indexes the SIGIR 2021 paper "Few-Shot Conversational Dense Retrieval", available in the ACM Digital Library:

Source: https://dl.acm.org/doi/abs/10.1145/3404835.3462856
Venue: SIGIR 2021
Category: Multi-Turn / Conversational Information Retrieval

Problem Setting

Conversational dense retrieval sits at the intersection of dialogue systems and neural information retrieval. In multi-turn search, follow-up queries are often incomplete on their own — they rely on coreference resolution and context from earlier turns ("What about the second one?"). A conversational dense retriever must encode dialogue history into query representations that can be matched against documents via dense (embedding-based) retrieval.

A key difficulty is data scarcity: large-scale conversational search datasets with per-turn relevance labels are rare, which motivates few-shot approaches that reduce dependence on extensive labeled conversations.

Context in the Reading List

The post places this paper in a section on multi-turn interaction, cross-referenced with related entries including:

Surveys on multi-turn interaction capabilities of LLMs and LLM-based conversation agents
Query rewriting and contextual history enhancement methods (e.g., CHIQ)
Empirical analyses of multi-turn conversational recommender systems
Work on contextual disentanglement for conversation modeling

Editorial Notes

The forum post is largely template scaffolding; it does not reproduce the paper's method details or quantitative results. Readers should consult the original PDF at the ACM link above for:

The exact few-shot training strategy and supervision signals used
Benchmark datasets (e.g., CAsT-style conversational search collections) and metrics (nDCG, MRR, Recall@k)
Baselines and ablation results

Why It Matters

This paper represents an early-stage effort to make dense retrieval work in conversational settings before the LLM era. It is relevant to anyone tracing the evolution from classical conversational IR toward modern retrieval-augmented generation (RAG) and agentic search, where multi-turn context modeling and efficient retrieval under limited supervision remain core challenges.

Few-Shot Conversational Dense Retrieval (SIGIR 2021)

Summary

Few-Shot Conversational Dense Retrieval (SIGIR 2021)

Overview

Problem Setting

Context in the Reading List

Editorial Notes

Why It Matters

Tags