Query Expansion

Pseudo-relevance feedback is an automatic technique that assumes the top-k initially retrieved documents are relevant. It extracts significant terms from these documents to expand the original query.

• Process: Execute the original query, analyze the top results, identify frequent or discriminative terms (using metrics like TF-IDF), and add them to the query.
• Key Algorithm: The Rocchio algorithm is a classic PRF method that creates a new query vector as a weighted combination of the original query vector and the centroid vectors of the relevant (top-retrieved) documents.
• Use Case: Foundational method in traditional search engines and a baseline for modern dense retrieval systems. It is computationally efficient but risks query drift if initial results are poor.

Thesaurus-based expansion uses a predefined lexical database of synonyms, hyponyms, and related concepts to add terms to a query.

• Resources: Can utilize general linguistic resources like WordNet or domain-specific ontologies and knowledge graphs.
• Process: For each query term, lookup synonyms and semantically related words, then append them to the query, often with a lower boost factor.
• Advantage: Provides controlled, explainable expansion. Highly effective in specialized domains with stable vocabularies (e.g., medicine, law).
• Limitation: Static nature can miss emerging terminology and contextual nuances captured by distributional methods.

This technique uses distributional semantic models like Word2Vec, GloVe, or contextual embeddings from models like BERT to find semantically related terms for expansion.

• Process: Encode query terms into a vector space, then retrieve the nearest neighbor vectors as expansion candidates based on cosine similarity.
• Contextual Expansion: Modern approaches use masked language models (e.g., BERT) to predict likely words given the full query context, generating more coherent expansions than static embeddings.
• Advantage: Captures complex semantic and syntactic relationships not found in a thesaurus. Adapts to the specific corpus used for training the embeddings.

Generative expansion employs a large language model to directly generate a rewritten or expanded version of the query.

• Process: The original query is fed to an LLM with an instruction (e.g., 'Generate a more comprehensive search query for: [original query]'). The model outputs a natural language query with added context, synonyms, and clarifying phrases.
• Fine-Tuning: Models can be fine-tuned on query-reformulation pairs from search logs for higher quality.
• Advantage: Produces fluent, multi-term expansions that read like natural human queries. Can perform query clarification and disambiguation implicitly.
• Consideration: Introduces latency and cost due to LLM inference. Requires careful prompting to avoid hallucination of irrelevant terms.

Explicit relevance feedback is a supervised technique where a user identifies relevant and non-relevant documents from initial results, and the system uses this signal to reformulate the query.

• Process: User marks documents as 'relevant' or 'not relevant'. The system computes a new query vector that moves toward the centroid of relevant documents and away from the centroid of non-relevant documents (using the Rocchio algorithm).
• Key Feature: Provides high-precision signal, often leading to very effective query refinement.
• Use Case: Common in academic literature search systems and enterprise settings where user interaction is expected. Less common in web search due to the requirement for explicit feedback.

This advanced approach uses a cross-encoder model not for initial retrieval but for scoring candidate expansion terms. It's often part of a multi-stage retrieval pipeline.

• Process: Generate a list of potential expansion terms using another method (e.g., embeddings). A cross-encoder model, which jointly processes the query and a candidate term, scores how useful each term would be for retrieval. Top-scoring terms are selected.
• Advantage: Cross-encoders provide highly accurate pairwise relevance judgments, leading to precise term selection. This mitigates the noise introduced by simpler expansion methods.
• System Role: Typically used as a reranking or filtering stage after a broader expansion technique to ensure quality, balancing recall and precision.

A retrieval strategy that combines sparse retrieval (e.g., keyword-based BM25) and dense retrieval (e.g., vector similarity search) to leverage both exact lexical matches and semantic understanding. It typically uses a weighted score fusion (like reciprocal rank fusion) to rank final results.

• Purpose: Mitigates the weaknesses of either approach alone—sparse search can miss semantic matches, while dense search can be overly broad.
• Implementation: A query is executed against both a traditional inverted index and a vector index; results are merged and re-ranked.

A probabilistic ranking function used by search engines to estimate the relevance of documents to a given search query. It is the foundational algorithm for sparse retrieval and keyword search.

• Mechanism: Scores documents based on term frequency (TF) and inverse document frequency (IDF), with built-in normalization for document length.
• Role in Expansion: Often used to retrieve an initial set of documents from which terms for query expansion are extracted (pseudo-relevance feedback).

A neural retrieval architecture that uses two separate BERT-based encoders—a question encoder and a passage encoder—to map queries and documents into a shared dense vector space. Relevance is measured by vector similarity (e.g., dot product).

• Training: Trained end-to-end on (question, positive passage, negative passage) triplets.
• Contrast with Expansion: DPR learns semantic matching directly; query expansion can be used as a preprocessing step to enrich the input to the question encoder.

A re-ranking algorithm designed to reduce redundancy in search result sets. It selects items that are both relevant to the query and sufficiently diverse from items already selected.

• Formula: Balances query similarity and inter-document similarity.
• Application: Often used after query expansion to diversify the final results, preventing a list of near-identical documents that all match the expanded terms.

A specific query expansion technique where the top k documents from an initial search are assumed to be relevant. Terms are then extracted from these documents to augment the original query for a second, more comprehensive search.

• Process: 1. Initial retrieval. 2. Term extraction from top docs. 3. Query reformulation. 4. Final retrieval.
• Key Assumption: The top-ranked documents are mostly relevant, which can fail if the initial search is poor.

The core data structure enabling fast keyword search. It maps each unique term (or token) to a postings list—a record of all documents (and often positions) where the term appears.

• Function: Enables efficient Boolean queries and term-frequency calculations.
• Relationship: Query expansion often relies on statistics (like IDF) derived from the global inverted index to select valuable new terms.