CTR-Guided Generative Query Suggestion in Conversational Search (EMNLP 2025 Industry Track)
Overview
This paper appears in the EMNLP 2025 Industry Track (ACL Anthology) and tackles query suggestion in conversational search — generating follow-up queries that anticipate user intent across multi-turn interactions.
Problem and Motivation
Conversational search systems must suggest next queries that are not only semantically coherent with the dialogue context, but also likely to be clicked and satisfy the user. Traditional pipelines separate retrieval, ranking, and generation, and often optimize only for semantic relevance, ignoring behavioral signals such as click-through rate (CTR). This work proposes using CTR signals as a guiding objective for generative query suggestion with large language models (LLMs).
Key Contributions
- A generative query suggestion framework for conversational search guided by CTR feedback from real search logs.
- Alignment of LLM-generated suggestions with both dialogue context and actual user engagement, rather than semantic plausibility alone.
- An industrial deployment perspective, reflecting the constraints of production search systems (latency, cost, safety, and measurable online impact).
- Generative Information Retrieval (Gen-IR): replacing cascaded retrieve-then-rank pipelines with generation.
- Conversational search / query reformulation: extending prior work such as generative query rewriting (e.g., ConvGQR) toward engagement-aware suggestion.
- LLM alignment with business metrics: using click and session logs as reward or supervision signals, similar in spirit to RLHF-style tuning but grounded in production behavioral data.
- Original paper: https://aclanthology.org/2025.emnlp-industry.178/
- Related topics: conversational search surveys, generative query reformulation (ConvGQR), agentic conversational search, LLM-based retrievers (e.g., ChatRetriever).
Context in the Field
The work sits at the intersection of:
Practical Takeaways
For practitioners building conversational or generative search systems:
1. Behavioral signals matter: semantic quality of suggestions does not guarantee user engagement; CTR guidance helps close the offline/online gap. 2. Pipeline integration: suggestion generation should be evaluated as part of the full retrieve → rank → generate loop, not in isolation. 3. Deployment constraints: latency budgets, hallucination risk, and output filtering remain hard requirements for industrial systems.