Query Rewriting

Automatically augments a query with terms that share identical or highly similar meanings to the original keywords. This technique directly increases recall by ensuring documents using different but equivalent vocabulary are matched.

• Mechanism: Utilizes a controlled thesaurus, WordNet, or static embedding models like Word2Vec.
• Example: A query for 'attorney' is expanded to search for 'lawyer' and 'counsel' simultaneously.
• Risk: Unchecked expansion can introduce ambiguity and reduce precision if synonyms have multiple senses.

Detects and rectifies typographical errors before the query hits the index, preventing zero-result searches. This relies on edit distance metrics to find the closest valid term.

• Levenshtein Distance: Calculates the minimum number of single-character edits (insertions, deletions, substitutions) required to transform the misspelled word into a dictionary term.
• Fuzzy Matching: Tolerates minor string variations, ensuring 'iphoen' correctly retrieves 'iPhone'.
• Implementation: Often uses finite-state automata or Damerau-Levenshtein distance for high-speed, high-volume query logs.

Modifies query specificity by traversing the semantic hierarchy of concepts. This structural approach relies on lexical databases or taxonomies rather than statistical similarity.

• Hypernym Expansion (Broadening): Adds more general terms to increase recall. A search for 'sedan' is expanded to include 'car' or 'vehicle'.
• Hyponym Expansion (Narrowing): Adds specific instances to refine the search. A query for 'car' is expanded to include 'sedan', 'SUV', and 'coupe'.
• Application: Critical for e-commerce platforms where users search with varying levels of product category granularity.

Reduces inflected word forms to a common base to normalize the query and document vocabulary. This ensures that 'running', 'ran', and 'runs' are treated as the same root concept.

• Stemming: A crude, rule-based heuristic that chops off affixes (e.g., Porter Stemmer reduces 'fishing' to 'fish'). It is fast but often produces non-words.
• Lemmatization: A morphological analysis that considers the part-of-speech to return the canonical dictionary form (e.g., 'better' becomes 'good'). It requires linguistic context but is more accurate.
• Trade-off: Stemming is computationally cheaper, while lemmatization preserves semantic integrity for downstream NLP tasks.

Bridges the gap between shorthand abbreviations and their full textual forms, as well as variations in pronunciation-based spelling.

• Acronym Expansion: Replaces 'WHO' with 'World Health Organization' (or vice versa) to ensure documents using either form are retrieved. This is vital for biomedical and legal domains.
• Phonetic Expansion: Uses algorithms like Soundex or Metaphone to index words by their sound. A query for 'fone' will match 'phone'.
• Use Case: Essential for voice search interfaces and databases containing proper names with multiple spelling variants.

An automatic, iterative expansion technique that assumes the top-k results from an initial retrieval are relevant. The system extracts dominant terms from these pseudo-relevant documents and adds them to the original query.

• Process: Execute initial query -> Analyze top 10 documents -> Extract high-frequency, discriminative terms -> Re-weight and re-execute the query.
• Rocchio Algorithm: A classic vector-space implementation that reformulates the query vector by adding the centroid of relevant document vectors.
• Risk: Can cause query drift if the initial top-k results are actually non-relevant, poisoning the feedback loop.

Automatically appending terms with identical or highly similar meanings to the original query to bridge the vocabulary gap between user intent and document indexing.

• Mechanism: Utilizes controlled vocabularies, WordNet, or static embeddings.
• Example: Rewriting 'laptop' to 'laptop OR notebook'.
• Risk: Can introduce noise if polysemy is not handled, reducing precision.

The process of detecting and rectifying typographical errors before the query hits the index, preventing zero-result searches.

• Mechanism: Often relies on Levenshtein Distance or probabilistic noisy channel models.
• Example: Correcting 'iphone chargr' to 'iphone charger'.
• Implementation: Requires a high-confidence threshold to avoid altering valid rare terms like proper nouns.

An automatic expansion technique that assumes the top-k initially retrieved documents are relevant, extracting salient terms from them to augment the original query.

• Process: Execute initial search → Analyze top results → Add frequent terms to the query.
• Advantage: Adapts dynamically to the corpus without manual thesauri.
• Pitfall: Topic drift can occur if the initial results are poor, poisoning the expanded query.

Leveraging Large Language Models (LLMs) like GPT or T5 to generate hypothetical answers, synonyms, or alternative phrasings for a query.

• Technique: Uses prompt engineering to instruct the model to act as a search assistant.
• Example: Expanding 'best cooling sheets' to include 'breathable bedding' and 'temperature regulating linens'.
• Strength: Captures long-tail semantic relationships missed by static dictionaries.

Reducing words to a common base form to match morphological variants.

• Stemming: A crude, rule-based chopping of affixes (e.g., 'running' → 'run'). Often uses the Porter Stemmer.
• Lemmatization: Morphological analysis using a vocabulary to return the dictionary form (e.g., 'better' → 'good').
• Trade-off: Stemming is faster but less accurate; lemmatization requires part-of-speech tagging.

Using deep transformer models like BERT to generate expansion terms that are semantically appropriate for the specific context of the query, not just the word in isolation.

• Differentiator: Unlike static Word2Vec, it distinguishes between 'bank' (river) and 'bank' (financial).
• Method: Masks terms in the query and predicts replacements, or uses encoder representations to find nearest neighbors in a dense vector space.