Refining Neural Query Representations with Pseudo Relevance Feedback

205 concepts163 papers6 proposals

When a user types a short query into a search engine, they rarely provide enough context to capture their true information need. A two-word query like "apple health" could refer to corporate wellness programs, agricultural safety, or an iOS application.

Historically, information retrieval systems solved this ambiguity using Pseudo-Relevance Feedback (PRF). The core intuition is simple: perform an initial search, assume the top $k$ retrieved documents are relevant (hence "pseudo"-relevance), and extract prominent terms from those documents to expand the original query.

While this keyword-extraction approach worked well for traditional lexical search engines like BM25, the rise of neural retrieval architectures—such as dense vector embeddings and large language models (LLMs)—requires a fundamental rethink of how feedback is processed. Modern research focuses on how to translate the intuition of PRF into continuous vector spaces and generative prompts, allowing systems to dynamically reshape search intent without losing semantic precision.


Vector Space PRF: Shifting the Embedding Centroid

In dense retrieval, queries and documents are mapped to continuous, high-dimensional vector spaces where semantic similarity is measured by a dot product or cosine distance. If an initial query vector is ambiguous, it may land on the periphery of the target document cluster.

Traditional Lexical PRF:  [Query] -> Add Keywords -> [Expanded Query]
Dense Vector PRF:         [Query Vector] -> Shift toward Centroid of Top Docs -> [Refined Vector]

To perform PRF in this paradigm, HongChien Yu et al. introduced ANCE-PRF in Improving Query Representations for Dense Retrieval with Pseudo Relevance Feedback. Instead of modifying text, the system performs vector arithmetic. By applying the classic Rocchio algorithm directly to dense embeddings, the query vector is mathematically shifted toward the semantic centroid of the top-ranked document embeddings.

As investigated by Hang Li et al. in To Interpolate or not to Interpolate: PRF, Dense and Sparse Retrievers, this vector shift acts as a semantic anchor. It updates the query representation to reflect the latent concepts shared by the initial results, allowing a second-stage retrieval pass to surface highly relevant documents that might not share a single explicit keyword with the original query.


Concatenation and Cross-Attention Feedback

While shifting a vector toward a centroid is computationally efficient, it compresses all the rich, multi-faceted information of the feedback documents into a single point. An alternative approach feeds the raw text of the top results back into the neural encoder.

In a reproducibility study of dense PRF methods, Improving Query Representations for Dense Retrieval with Pseudo Relevance Feedback: A Reproducibility Study, Hang Li et al. detailed concatenation-based PRF. In this setup, the text of the top-ranked documents is appended directly to the original query string, and the entire sequence is re-encoded by a deep Transformer model.

This works because it leverages the encoder's internal cross-attention mechanisms. The Transformer layers can model token-level interactions between the query and the feedback text, resolving ambiguities and generating a highly contextualized representation. However, passing multiple long documents through a Transformer introduces a heavy computational burden.

To solve this, HongChien Yu et al. proposed PGT: Pseudo Relevance Feedback Using a Graph-Based Transformer. PGT uses a graph-structured attention mechanism to sparsify interactions across multiple feedback documents, capturing deep contextual relationships without the quadratic cost of full attention.


Expanding to Sparse Neural and Generative Models

The evolution of neural PRF extends beyond dense embeddings into sparse neural representations and generative language models:

  • Sparse Neural Models: Models like SPLADE map text to high-dimensional, sparse vectors aligned with a vocabulary. Carlos Eduardo Rosar Kos Lassance et al. demonstrated in Naver Labs Europe (SPLADE) @ TREC NeuCLIR 2022 that applying the Rocchio algorithm to these sparse neural vectors effectively expands and re-weights explicit terms, combining the interpretability of keyword expansion with neural semantic awareness.
  • Generative LLM Rerankers: In multi-stage ranking pipelines, LLMs are often used to re-score documents. Xiaoyang Chen et al. showed in CIP at TREC Deep Learning Track 2023 that injecting top-ranked documents directly into an LLM's prompt acts as in-context pseudo-relevance feedback. The text provides a dynamic reference point that helps the model better differentiate relevant nuances from background noise.

Mitigating Query Drift

The fatal flaw of any PRF system is query drift: if the initial search returns irrelevant results, feeding them back into the model will amplify the errors, pulling the subsequent search further away from the user's actual intent.

High Coherence:  [Doc 1] [Doc 2] [Doc 3] are tightly clustered -> Trust Feedback
Low Coherence:   [Doc 1] ... [Doc 2] ... [Doc 3] are scattered      -> Reject Feedback

To prevent this, researchers have introduced quality-gating mechanisms. Jiyin He et al. explored this in On the Topical Structure of the Relevance Feedback Set, utilizing a structural coherence score to evaluate the topical compactness of the top-retrieved documents. If the initial results are scattered and lack consistency, the system refutes or minimizes the feedback step to avoid drift.

Similarly, Ben He et al. developed an adaptive optimization framework called ReAL in Not All Terms Matter: Recall-Oriented Adaptive Learning for PLM-aided Query Expansion in Open-Domain Question Answering. ReAL uses a classifier to dynamically label retrieved documents as pseudo-relevant or pseudo-irrelevant, iteratively adjusting query term weights to emphasize helpful context while suppressing noisy or irrelevant feedback signals.

By formalizing these feedback loops—a challenge also addressed mathematically by Yuanhua Lv in Revisiting the Divergence Minimization Feedback Model—modern information retrieval transforms PRF from a brittle keyword heuristic into a robust, multi-modal optimization process.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 160 papers
  2. Version 1 · 2026-10-04 · 123 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_pseudo_relevance_feedback_5dfd412e,
  title = {pseudo-relevance feedback},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/pseudo-relevance-feedback/rdr_5dfd412efa9740cf9384dff8c44e18ff},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_pseudo_relevance_feedback_5dfd412e_v2,
  title = {pseudo-relevance feedback},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/pseudo-relevance-feedback/rdr_5dfd412efa9740cf9384dff8c44e18ff/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake