Decomposing Complex Queries for Tip-of-the-tongue Retrieval (arXiv 2305.15053)

Metadata

| Field | Value | |-------|-------| | Title | Decomposing Complex Queries for Tip-of-the-tongue Retrieval | | Authors | Kevin Lin, Kyle Lo, Joseph E. Gonzalez, Dan Klein | | Published | May 2023 (arXiv) | | Link | https://arxiv.org/abs/2305.15053 | | Type | Academic paper | | Section | Query Understanding |

Overview

This post discusses an arXiv paper on tip-of-the-tongue (TOT) retrieval: the task of retrieving a sought item (e.g., a movie or book) from a vague, circumlocutory natural-language description rather than its title — for example, describing "a movie where a guy tosses a coin" instead of naming the film.

Core Idea

• TOT descriptions are long, complex queries that describe multiple facets of an item, making direct retrieval difficult.
• The paper proposes using large language models (LLMs) to decompose a complex user description into simpler sub-queries, each targeting one facet or property of the sought item.
• Candidates are retrieved independently for each sub-query and then aggregated/reranked to produce the final ranking.
• Decomposition improves coverage of the item's attributes and, combined with LLM-based reranking, yields stronger end-to-end retrieval performance than querying with the raw description alone.

Context in Modern IR

The post situates the paper within broader trends:

• Retrieval pipeline evolution: from BM25 and dense bi-encoders to cross-encoder rerankers, generative retrieval, and LLM agents.
• RAG / Agentic Search: treating retrieval as an iterative, plannable process rather than a single-shot lookup; query decomposition is a key primitive in this paradigm.
• Evaluation: standard IR metrics such as nDCG@10, MRR, Recall@k, plus task success rate, latency, and token cost.

Engineering Checklist (from the post)

| Concern | Recommendation | |---------|----------------| | Latency | Cascaded retrieval with early stopping; cache frequent queries | | Quality | Verify offline gains translate to online satisfaction; human audit | | Safety | Source whitelisting, adversarial detection, output filtering | | Cost | Route to smaller models where possible; hybrid sparse-dense retrieval |

Related Entries

• Aligned Query Expansion (arXiv 2507.11042)
• Train Your LLM for Query Expansion (arXiv 2510.10009)
• Few-Shot Generative Conversational Query Rewriting (SIGIR 2020, DOI: 10.1145/3397271.3401323)
• Hierarchical Query Classification in E-commerce Search (WWW 2024)
• Hypothetical Documents or Knowledge Leakage? (arXiv 2504.14175)

Takeaways

1. For researchers: query decomposition with LLMs is a strong baseline for complex, multi-facet descriptions; report statistical significance and compute cost. 2. For engineers: decomposition and reranking modules can be plugged into existing search stacks; evaluate integration overhead. 3. For product teams: focus on user-perceivable gains (answer credibility, multi-turn consistency) rather than offline nDCG alone.

> Note: For exact quantitative results, consult the original paper at https://arxiv.org/abs/2305.15053.