Pseudo-Relevance Feedback (PRF)

The foundational premise of PRF is the Blind Feedback Assumption: the top k documents retrieved by an initial search (e.g., using BM25) are treated as pseudo-relevant, even without human verification. This assumption is necessary for automation but introduces risk. The technique extracts the most informative terms from these documents to augment the original query. The value of k is a critical hyperparameter; a small k may miss relevant terms, while a large k increases noise from irrelevant documents that slipped into the initial results.

PRF is mathematically formalized by algorithms like Rocchio and the more modern Relevance Model 3 (RM3).

• Rocchio Algorithm: Creates an updated query vector by moving it toward the centroid of the pseudo-relevant documents and away from non-relevant ones (if identified). The formula is: Q_new = α * Q_original + β * (1/|D_r|) * Σ(D_r) - γ * (1/|D_nr|) * Σ(D_nr).
• Relevance Model 3 (RM3): A probabilistic approach that builds a unigram language model from the pseudo-relevant documents and interpolates it with the original query model. It selects the highest-probability terms from this combined model for expansion, weighted by their estimated importance.

Not all terms from the top documents are useful for expansion. PRF systems employ scoring functions to select the most discriminative terms. Common criteria include:

• High frequency in pseudo-relevant docs: Terms that appear often in the top results.
• Low frequency in the general corpus: Measured by Inverse Document Frequency (IDF), favoring rare, specific terms.
• Proximity to original query terms: Terms that co-occur with the original query words in the documents.

Selected terms are assigned a weight, often proportional to their score, and added to the query. The original query terms are typically re-weighted (boosted) to maintain the core intent.

The principal goal of PRF is to improve recall—the system's ability to retrieve all relevant documents. The original query may be short, ambiguous, or use different vocabulary than the relevant documents (vocabulary mismatch problem). By adding synonymous or related terms from the top documents, PRF creates a richer query representation that matches a wider set of relevant documents. This is particularly effective for short-tail queries (1-2 words) and in domains with specialized jargon where users and documents may use different terms for the same concept.

The major failure mode of PRF is query drift, where the expanded query moves away from the user's original intent. This occurs when the top k documents from the initial search are not truly relevant. The system then extracts and amplifies terms from irrelevant documents, propagating and compounding the error in the second retrieval stage. Query drift is more likely with:

• Poor initial retrieval performance.
• Ambiguous or underspecified original queries.
• An incorrectly large k value. Robust implementations use negative feedback (identifying non-relevant terms) or dynamic k adjustment to mitigate this risk.

While rooted in classical information retrieval, PRF principles are adapted in modern systems:

• Dense Retrieval (DPR): The initial retrieval uses a bi-encoder to get top passages. Expansion can involve generating a new query embedding based on the embeddings of the top passages, or using a cross-encoder to score and select expansion terms.
• LLM-Powered PRF: A Large Language Model can be prompted to analyze the top-retrieved passages and generate a refined or expanded query, moving beyond simple term extraction to true query reformulation. This is sometimes called LLM-as-PRF.
• Hybrid Retrieval: PRF can be applied to both the sparse (keyword/BM25) and dense (vector) retrieval paths in a hybrid system, with the results fused for the final ranking.

The Rocchio Algorithm is a classic vector space model technique for query expansion and relevance feedback. It creates a new query vector by moving the original query vector towards the centroid of the pseudo-relevant documents and away from the centroid of non-relevant documents (though the latter is often omitted in PRF).

• Formula: Q_new = α * Q_original + β * (1/|D_rel|) * Σ(D_rel)
• Parameters: α controls the weight of the original query, β controls the influence of the relevant documents.
• Use Case: Foundational method in traditional information retrieval systems, often used as a baseline for modern neural approaches.

Relevance Model 3 (RM3) is a widely adopted, state-of-the-art probabilistic PRF method. It constructs a unigram language model from the top-ranked documents and interpolates it with the original query model.

• Process: Estimates a probability distribution over terms from the pseudo-relevant set, then mixes this distribution with the original query terms.
• Key Feature: Uses Dirichlet smoothing to account for term frequencies within the feedback documents and the entire collection.
• Advantage: Proven effective for both traditional bag-of-words retrievers (like BM25) and as a component in dense retrieval pipelines.

This approach frames PRF as an information retrieval optimization problem. The goal is to find an expanded query model that minimizes the Kullback-Leibler (KL) divergence between the query language model and the language model of the pseudo-relevant documents.

• Objective: Minimize D_KL(θ_F || θ_Q), where θ_F is the feedback document model and θ_Q is the query model.
• Outcome: Selects terms that best approximate the language used in the relevant documents.
• Application: Commonly used in language modeling frameworks for IR, providing a principled probabilistic foundation for term selection.

Modern neural PRF methods use transformer-based models like BERT or RoBERTa to score and select expansion terms in a context-aware manner, moving beyond simple term frequency.

• Mechanism: The model encodes the query and candidate passages, then scores terms based on their contextual importance to the query, not just their occurrence.
• Example: The BERT-QE model treats term selection as a classification problem, using the [CLS] token representation to predict useful expansion terms.
• Benefit: Captures semantic relatedness and can identify useful terms that are not lexically present in the top documents but are contextually implied.

A simple yet effective PRF strategy involves re-weighting terms from the original query and new expansion terms based on their Inverse Document Frequency (IDF) within the feedback set and the entire corpus.

• Process: Extract all terms from the top k documents. Score each term by tf_feedback * idf_collection. The highest-scoring terms are added to the query.
• Rationale: Favors terms that are frequent in the pseudo-relevant set but rare in the overall collection, maximizing discriminative power.
• Implementation: Often a core subroutine within more complex algorithms like RM3, and is straightforward to implement in systems like Elasticsearch or custom retrievers.

In dense retrieval systems like ANCE or DPR, PRF is adapted to work in the embedding space. The expanded query is often represented as a new dense vector.

• Method 1: Vector Averaging. The embeddings of the top pseudo-relevant passages are averaged, and this centroid is used as (or blended with) the new query embedding.
• Method 2: Learned Fusion. A neural aggregator (e.g., a transformer) learns to combine multiple passage embeddings into a single, improved query representation.
• Challenge: Requires the retriever to be robust to the shifted query vector. Systems like ANCE-PRF train the retriever end-to-end with simulated PRF data.

Query expansion is a general retrieval technique that augments an original user query with additional relevant terms or phrases to improve recall by matching a broader set of documents. Unlike PRF, which expands queries automatically using initial search results, expansion can also be driven by:

• Static thesauri or domain-specific knowledge graphs.
• Global analysis of term co-occurrence across an entire corpus.
• User interaction data from query logs and click-through rates. The core challenge is adding terms that increase recall without introducing semantic drift that harms precision.

Dense retrieval is a neural search paradigm where queries and documents are encoded into dense vector embeddings using models like BERT or Sentence Transformers. Relevance is computed via similarity (e.g., cosine) in this high-dimensional space. PRF can be adapted for dense retrieval through techniques like:

• Dense Passage Retrieval (DPR) with iterative refinement, where an initial retrieval is used to generate a new, improved query embedding.
• Learned query expansion models that predict expansion term vectors directly. This contrasts with traditional PRF applied to sparse, term-based retrieval models like BM25.

Cross-encoder reranking is a precision-focused technique that follows an initial broad retrieval (which may use PRF). A computationally intensive transformer model jointly processes a query-candidate document pair to produce a more accurate relevance score. Key aspects:

• It acts as a precision filter, reordering the top results from a fast, recall-oriented retriever.
• While PRF aims to improve the initial recall of the retriever, cross-encoders are used to improve the final ranking of already-retrieved documents.
• They are often too slow for first-stage retrieval but provide high accuracy for re-ranking the top 100-1000 candidates.

Relevance Feedback is the interactive, supervised counterpart to PRF. In this process:

• A user issues an initial query and receives results.
• The user explicitly marks documents as relevant or non-relevant.
• The system uses this labeled data to update the query representation (e.g., boosting terms from relevant docs, down-weighting terms from non-relevant docs) and performs a new search. PRF is an automatic approximation of this ideal process, assuming the top k documents are relevant without user input, making it scalable but potentially noisy.

Query reformulation is the process of altering a user's original query to better align with their underlying information need. It encompasses a wider range of operations than PRF's term expansion, including:

• Spelling correction and grammatical normalization.
• Semantic rewriting using seq2seq models to generate a wholly new, clearer query.
• Query simplification or decomposition for complex questions. While PRF adds terms, reformulation may delete, reorder, or replace terms. Modern systems often use large language models for zero-shot query reformulation.

BM25 is a probabilistic ranking function and the dominant algorithm for sparse, term-based retrieval. It estimates relevance based on term frequency (TF) and inverse document frequency (IDF) with built-in saturation. PRF is classically and most frequently applied on top of BM25. The standard Rocchio algorithm for PRF uses BM25 scores to weight terms extracted from the top-ranked documents. Understanding BM25 is essential because:

• It defines the initial ranking from which PRF assumes pseudo-relevance.
• The mathematical formulation of term weighting in BM25 influences how expansion terms are selected and weighted during feedback.