Few-Shot Generative Conversational Query Rewriting (SIGIR 2020)
Overview
- Venue: SIGIR 2020
- DOI / Source: https://dl.acm.org/doi/abs/10.1145/3397271.3401323
- Topic area: Query Understanding / Conversational Search
- Query expansion with aligned or LLM-based techniques
- Query decomposition for complex / tip-of-the-tongue retrieval
- Retrieval-augmented generation (RAG) and agentic search, where reformulation and retrieval become iterative, model-driven processes
- The core idea — synthesize supervision from structured signals in a large corpus (anchor/relation patterns) instead of manual labels — remains applicable when bootstrapping query rewriting for new domains or languages.
- When reproducing or citing quantitative results, consult the original PDF; this entry is based on the paper metadata and public source listing.
- Aligned Query Expansion: Efficient Query Expansion for Information Retrieval
- Beyond the Limitation of a Single Query: Train Your LLM for Query Expansion
- Decomposing Complex Queries for Tip-of-the-tongue Retrieval
- Hierarchical Query Classification in E-commerce Search (WWW 2024)
- Hypothetical Documents or Knowledge Leakage? Rethinking LLM-based Query Expansion
- Original paper: Few-Shot Generative Conversational Query Rewriting, SIGIR 2020. https://dl.acm.org/doi/abs/10.1145/3397271.3401323
Problem Statement
In conversational search, users issue follow-up queries that are often underspecified when read in isolation (e.g., "and what about its population?"). Retrieval systems, however, typically match each query against documents independently, so the query must first be rewritten into a standalone form that resolves coreferences and omitted context from the dialogue history.
Prior approaches to query rewriting generally require substantial labeled training data of (contextual query → rewritten query) pairs, which is expensive to produce and domain-specific.
Approach
The paper proposes a generative few-shot solution:
1. Pre-trained generative model: A GPT-2 style language model is used to generate rewrites, rather than training a task-specific rewriter from scratch. 2. Automatically constructed training signal: Training examples are derived from Wikipedia by exploiting anchors of salient relation types — for instance, entities linked alongside a given entity (locations, dates, other entities). The surrounding text serves as context, and the anchored expression serves as the target mention, approximating the structure of conversational queries (where an entity is implied by context and referred to elliptically). 3. Few-shot usage: Because training pairs are mined from Wikipedia rather than manually annotated conversational logs, the method avoids in-domain supervision and transfers to conversational search settings with minimal adaptation.
At inference time, the model conditions on the conversation context plus the current (underspecified) query and generates the rewritten standalone query, which is then fed to the retrieval pipeline.
Positioning in the Literature
This work sits in the query understanding stage of the search stack, upstream of retrieval and ranking. It is closely related to subsequent research directions: