Bridging Semantic Gaps with Generative Query Expansion

252 concepts186 papers2 proposals

Traditional information retrieval relies on a fragile assumption: that the words a user chooses to describe their need will match the words an author used to describe the solution. This "vocabulary gap" is the primary bottleneck in search. A user might search for "past tense verbs" while the relevant document is a French exercise containing only the phrase "J'ai mangé."

Generative Query Expansion (GQE) solves this by using Large Language Models (LLMs) not as final answer-engines, but as semantic translators. Instead of searching with the user's raw, often fragmented input, these systems use generative models to "hallucinate" what a relevant answer or document might look like, then use that synthetic content to perform the actual search.

The Hypothetical Pivot

The most influential intuition in this field is the "Generate-then-Retrieve" (GtR) strategy. As Abbasiantaeb et al. demonstrate, generating a hypothetical answer before retrieval significantly outperforms standard query rewriting.

This works because of how dense retrieval models represent information. In a vector space, a short question and a long, factual document often sit far apart. However, a synthetic answer generated by an LLM—even if it contains factual errors—shares the same semantic structure and "tone" as the real documents in the database. Gao et al. formalize this as Hypothetical Document Embeddings (HyDE). By transforming a query into a fake document, the system creates a "pivot" that sits much closer to the target information in vector space, capturing the expected semantic patterns of a relevant response.

Translating Intent to Distribution

GQE also acts as a bridge between different data distributions. Users often search using "compact and fragmented queries," whereas vision-language models are typically trained on verbose, descriptive captions. Liu et al. show that using an LLM to expand these fragments into full sentences aligns the user's intent with the model's training distribution, improving retrieval without needing to retrain the underlying search model.

This mechanism is particularly powerful in specialized domains:

  • Education: Bicknell et al. use LLMs to transform abstract instructional goals (e.g., "practice verbs") into concrete hypothetical exercises that match the actual content of a learning database.
  • Database Querying: In Text-to-SQL tasks, Kothyari et al. propose "schema hallucination," where the model predicts an idealized database structure to bridge the gap between natural language and formal naming conventions.
  • Design: Son et al. allow designers to "concretize" vague aesthetic terms into multiple specific prompts, providing the rich lexical detail necessary for visual search.

Inverting the Search with Synthetic Questions

While most GQE happens at query time, a parallel strategy involves expanding the documents themselves during indexing. The authors of the Ski framework (Zhang et al.) argue that matching a user's question to a raw paragraph is inherently difficult. Instead, they use LLMs to transform every document in a knowledge base into a set of synthetic questions that the document could answer.

This shifts the retrieval task from "Question-to-Paragraph" matching to "Question-to-Question" alignment. Because both the user's query and the synthetic index are now in the same format, the semantic gap is drastically reduced. Luo et al. and Ren et al. similarly find that these "pseudo-queries" serve as highly effective document identifiers, especially in large-scale web searches where traditional titles are often uninformative.

Iterative Refinement and Memory

The most advanced applications of GQE treat retrieval as a conversation rather than a single step. Gong et al. introduce ITER-RETGEN, which uses a model's initial (potentially imperfect) response as a source of semantic cues to find even better documents in a second pass.

For extremely long contexts, Qian et al. propose using a "memory module" to generate intermediate clues. These clues act as internal signposts, helping the retriever navigate complex, multi-hop questions that lack direct keyword matches. By treating the LLM's internal knowledge as a bridge to external evidence, these systems move beyond simple keyword matching toward a more human-like understanding of information seeking.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 179 papers
  2. Version 1 · 2026-10-04 · 277 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_generative_query_expansion_6699be4d,
  title = {generative query expansion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generative-query-expansion/rdr_6699be4da7a6465e8a4695b462a3ac32},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 2, frozen text):

@misc{lacuna_generative_query_expansion_6699be4d_v2,
  title = {generative query expansion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generative-query-expansion/rdr_6699be4da7a6465e8a4695b462a3ac32/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake