DiAL: Diversity-Aware Listwise Ranking for Query Auto-Complete (EMNLP 2024)
Overview
DiAL (Diversity-Aware Listwise ranking) is a paper accepted at EMNLP 2024, published under Amazon Science's Search Assistance research area.
- Publication page: https://www.amazon.science/publications/dial-diversity-aware-listwise-ranking-for-query-auto-complete
- Venue: EMNLP 2024
- Topic: Ranking for query auto-complete, with an explicit diversity objective
- Asking Clarification Questions to Handle Ambiguity in Open-Domain QA
- Asking Clarifying Questions in Open-Domain Information-Seeking Conversations
- Enhancing Discoverability in Enterprise Conversational Systems with Proactive Suggestions
- Evaluating Auto-Complete Ranking for Diversity and Relevance, ECIR 2025
- Evaluation and Continual Improvement for an Enterprise AI Assistant
- Generating Query Recommendations via LLMs (May 2024, arXiv)
- Quantitative results: Verify against the original PDF; this entry is based on the abstract and public metadata.
- Engineering trade-offs: Auto-complete has a strict latency budget (typically p99 in the tens of milliseconds), so listwise re-ranking depth, caching of popular prefixes, and early stopping are key implementation considerations.
- Evaluation: Diversity gains should be validated with interleaving experiments and human audits, not just offline metrics.
- Original publication: *DiAL: Diversity aware listwise ranking for query auto-complete*, EMNLP 2024. See the Amazon Science link above for the authoritative version.
Problem and Motivation
Query auto-complete systems suggest completions as users type in a search box. Ranking these suggestions by relevance or popularity alone tends to produce lists of near-duplicate completions, limiting the range of user intents a single suggestion list can cover. DiAL tackles this by treating suggestion ranking as a listwise problem: the model optimizes the composition and ordering of the whole candidate list, jointly accounting for relevance and diversity rather than scoring each candidate in isolation.
Positioning in the Ranking Landscape
This work sits at the intersection of learning-to-rank (LTR) and large-scale search systems. In the classic search stack, recall covers coverage, ranking handles discrimination, and generation handles presentation. Neural ranking has evolved from BM25, to BERT cross-encoders and dense bi-encoders, to generative retrieval and LLM-based agentic search — with each generation balancing the efficiency–effectiveness–maintainability triangle. Diversity-aware listwise ranking adds a further constraint: the marginal value of a suggestion depends on what else is already in the list.