ir-query-understanding

name	ir-query-understanding
description	Query expansion, spell correction, semantic search, query classification, entity recognition, and autocomplete

Information Retrieval: Query Understanding

Scope: Query preprocessing including expansion, spell correction, semantic understanding, classification, entity recognition, and suggestions Lines: ~300 Last Updated: 2025-10-25 Format Version: 1.0 (Atomic)

---

When to Use This Skill

Activate this skill when:

• Improving search recall through query expansion
• Handling misspelled queries with spell correction
• Understanding query intent (informational, navigational, transactional)
• Extracting entities from queries (people, places, products)
• Implementing autocomplete and query suggestions
• Bridging vocabulary gap between users and documents
• Supporting multilingual search queries
• Personalizing query interpretation based on user context

Core Concepts

Concept 1: Query Expansion

Techniques:

Synonym Expansion: Add related terms

• "laptop" → "laptop", "notebook", "computer"

Pseudo-Relevance Feedback: Use top results to expand

• Search → Get top 5 → Extract terms → Re-search

Word Embeddings: Add semantically similar terms

• Use word2vec, GloVe, or transformer embeddings

from gensim.models import Word2Vec
import numpy as np

# Load pre-trained word embeddings
# For demo, train small model (use pre-trained in production)
sentences = [
    ["machine", "learning", "algorithms"],
    ["deep", "learning", "neural", "networks"],
    ["supervised", "learning", "classification"],
]
model = Word2Vec(sentences, vector_size=100, window=5, min_count=1)

def expand_query_embeddings(query, model, top_k=3):
    """Expand query with semantically similar terms"""
    terms = query.lower().split()
    expanded = set(terms)

    for term in terms:
        if term in model.wv:
            # Get similar words
            similar = model.wv.most_similar(term, topn=top_k)
            for word, score in similar:
                if score > 0.7:  # Threshold for similarity
                    expanded.add(word)

    return " ".join(expanded)

# Pseudo-relevance feedback
def pseudo_relevance_feedback(query, search_fn, top_k=5, expansion_terms=3):
    """Expand query using top search results"""
    # Initial search
    results = search_fn(query, limit=top_k)

    # Extract terms from top results
    from collections import Counter
    term_counts = Counter()

    for doc in results:
        # Simple tokenization (use proper NLP in production)
        terms = doc['content'].lower().split()
        term_counts.update(terms)

    # Add most common terms not in original query
    query_terms = set(query.lower().split())
    expansion = []

    for term, count in term_counts.most_common():
        if term not in query_terms and len(expansion) < expansion_terms:
            expansion.append(term)

    expanded_query = query + " " + " ".join(expansion)
    return expanded_query

# Example
original_query = "machine learning"
expanded = expand_query_embeddings(original_query, model, top_k=2)
print(f"Expanded query: {expanded}")

Concept 2: Spell Correction

Approaches:

Edit Distance: Levenshtein distance to find close matches Context-Aware: Use language model to pick best correction Search-Based: Correct based on index terms

import re
from collections import Counter

def edit_distance(s1, s2):
    """Compute Levenshtein distance"""
    if len(s1) < len(s2):
        return edit_distance(s2, s1)

    if len(s2) == 0:
        return len(s1)

    previous_row = range(len(s2) + 1)
    for i, c1 in enumerate(s1):
        current_row = [i + 1]
        for j, c2 in enumerate(s2):
            insertions = previous_row[j + 1] + 1
            deletions = current_row[j] + 1
            substitutions = previous_row[j] + (c1 != c2)
            current_row.append(min(insertions, deletions, substitutions))
        previous_row = current_row

    return previous_row[-1]

class SpellChecker:
    def __init__(self, corpus):
        """Build dictionary from corpus"""
        self.word_counts = Counter()
        for doc in corpus:
            words = re.findall(r'\w+', doc.lower())
            self.word_counts.update(words)

    def correction(self, word):
        
         word  .word_counts:
             word

        
        candidates = ._candidates(word)

        
         (candidates, key= w: .word_counts.get(w, ))

     ():
        
        
         (
            ._known([word]) 
            ._known(._edits1(word)) 
            ._known(._edits2(word)) 
            [word]
        )

     ():
        
         (w  w  words  w  .word_counts)

     ():
        
        letters = 
        splits = [(word[:i], word[i:])  i  ((word) + )]
        deletes = [L + R[:]  L, R  splits  R]
        transposes = [L + R[] + R[] + R[:]  L, R  splits  (R) > ]
        replaces = [L + c + R[:]  L, R  splits  R  c  letters]
        inserts = [L + c + R  L, R  splits  c  letters]
         (deletes + transposes + replaces + inserts)

     ():
        
         (e2  e1  ._edits1(word)  e2  ._edits1(e1))


corpus = [, , ]
speller = SpellChecker(corpus)

misspelled = 
corrected = .join([speller.correction(word)  word  misspelled.split()])
()  

Concept 3: Query Classification

Intent Types:

• Informational: Seeking knowledge ("what is machine learning")
• Navigational: Finding specific site/page ("github login")
• Transactional: Ready to act ("buy macbook pro")

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression

# Training data: (query, intent)
training_data = [
    ("what is machine learning", "informational"),
    ("how does neural network work", "informational"),
    ("github login", "navigational"),
    ("facebook home page", "navigational"),
    ("buy macbook pro", "transactional"),
    ("book flight to paris", "transactional"),
    ("machine learning tutorial", "informational"),
    ("amazon prime", "navigational"),
    ("order pizza online", "transactional"),
]

queries, intents = zip(*training_data)

# Train classifier
vectorizer = TfidfVectorizer()
X = vectorizer.fit_transform(queries)

classifier = LogisticRegression()
classifier.fit(X, intents)

def classify_query(query, vectorizer, classifier):
    """Classify query intent"""
    X_query = vectorizer.transform([query])
    intent = classifier.predict(X_query)[0]
    probabilities = classifier.predict_proba(X_query)[0]

    return {
        'intent': intent,
        'confidence': max(probabilities)
    }

# Example
test_query = "how to learn python"
result = classify_query(test_query, vectorizer, classifier)
print(f"Intent: {result['intent']} (confidence: )")

Concept 4: Entity Recognition

Extract: People, places, products, dates, prices from queries

import re

class QueryEntityExtractor:
    def __init__(self):
        # Patterns for common entities
        self.patterns = {
            'price': r'\$?\d+(?:\.\d{2})?',
            'date': r'\d{4}-\d{2}-\d{2}|\d{1,2}/\d{1,2}/\d{4}',
            'email': r'[\w\.-]+@[\w\.-]+\.\w+',
            'phone': r'\(?\d{3}\)?[-.\s]?\d{3}[-.\s]?\d{4}',
        }

        # Product catalog (in production, load from database)
        self.products = {'macbook', 'iphone', 'ipad', 'airpods'}
        self.brands = {'apple', 'google', 'microsoft', 'amazon'}

    def extract(self, query):
        """Extract entities from query"""
        entities = {}

        # Regex-based extraction
        for entity_type, pattern in self.patterns.items():
            matches = re.findall(pattern, query, re.IGNORECASE)
            if matches:
                entities[entity_type] = matches

        # Dictionary-based extraction
        query_lower = query.lower()
        words = set(query_lower.split())

        entities['products'] = [p for p in self.products if p  query_lower]
        entities[] = [b  b  .brands  b  words]

         entities


extractor = QueryEntityExtractor()
query = 
entities = extractor.extract(query)
()

---

Patterns

Pattern 1: Autocomplete with Prefix Matching

When to use: Real-time query suggestions as user types

class TrieNode:
    def __init__(self):
        self.children = {}
        self.is_end = False
        self.frequency = 0

class Autocomplete:
    def __init__(self):
        self.root = TrieNode()

    def insert(self, query, frequency=1):
        """Insert query into trie"""
        node = self.root
        for char in query.lower():
            if char not in node.children:
                node.children[char] = TrieNode()
            node = node.children[char]

        node.is_end = True
        node.frequency += frequency

    def search(self, prefix, max_results=5):
        """Get top suggestions for prefix"""
        # Navigate to prefix
        node = self.root
        for char in prefix.lower():
            if char not in node.children:
                return []
            node = node.children[char]

        # Collect all completions
        results = []
        self._collect_completions(node, prefix, results)

        # Sort by frequency and return top-k
        results.sort(key= x: x[], reverse=)
         [query  query, freq  results[:max_results]]

     ():
        
         node.is_end:
            results.append((current, node.frequency))

         char, child  node.children.items():
            ._collect_completions(child, current + char, results)


ac = Autocomplete()
query_log = [
    (, ),
    (, ),
    (, ),
    (, ),
]

 query, freq  query_log:
    ac.insert(query, freq)


suggestions = ac.search(, max_results=)
()

Pattern 2: Contextual Query Rewriting

Use case: Rewrite query based on user context or conversation

def contextual_query_rewrite(query, user_context, conversation_history):
    """Rewrite query with context"""
    rewritten = query

    # Resolve pronouns using conversation history
    if any(pronoun in query.lower() for pronoun in ['it', 'this', 'that', 'he', 'she']):
        if conversation_history:
            last_entity = extract_main_entity(conversation_history[-1])
            rewritten = rewritten.replace('it', last_entity)
            rewritten = rewritten.replace('this', last_entity)

    # Add implicit location
    if user_context.get('location') and 'near' not in query.lower():
        if any(term in query.lower() for term in ['restaurant', 'hotel', 'store']):
            rewritten += f" near {user_context['location']}"

    # Add temporal context
    if user_context.get('time_period'):
        if any(term in query.lower() for term in ['news', 'events', 'weather']):
            rewritten += 

     rewritten


context = {: , : }
history = []

query = 
rewritten = contextual_query_rewrite(query, context, history)
()

Pattern 3: Multi-Language Query Understanding

When to use: Support queries in multiple languages

from langdetect import detect

def multilingual_query_processing(query):
    """Detect language and process accordingly"""
    try:
        lang = detect(query)
    except:
        lang = 'en'

    # Language-specific processing
    if lang == 'en':
        # English stopwords
        stopwords = {'the', 'a', 'an', 'is', 'are'}
    elif lang == 'es':
        # Spanish stopwords
        stopwords = {'el', 'la', 'los', 'las', 'es', 'son'}
    elif lang == 'zh-cn':
        # Chinese - different tokenization
        # Use jieba or similar
        stopwords = set()
    else:
        stopwords = set()

    # Process with language-specific rules
    terms = query.lower().split()
    filtered_terms = [t for t in terms if t not in stopwords]

    return {
        'language': lang,
        'processed_query': ' '.join(filtered_terms),
        'original_query': query
    }

# Example
query_en = "what is machine learning"
query_es = 

result_en = multilingual_query_processing(query_en)
result_es = multilingual_query_processing(query_es)

Pattern 4: Query Segmentation

Use case: Break complex queries into components

import re

def segment_query(query):
    """Segment query into structured components"""
    segments = {
        'what': None,    # What user is looking for
        'where': None,   # Location constraint
        'when': None,    # Time constraint
        'how': None,     # Method/manner
        'attributes': [] # Additional attributes
    }

    query_lower = query.lower()

    # Extract location
    location_pattern = r'(?:in|at|near)\s+([a-z\s]+?)(?:\s|$)'
    location_match = re.search(location_pattern, query_lower)
    if location_match:
        segments['where'] = location_match.group(1).strip()

    # Extract time
    time_keywords = ['today', 'tomorrow', 'this week', 'next month']
    for keyword in time_keywords:
        if keyword in query_lower:
            segments['when'] = keyword
            break

    # Extract main intent
    if query_lower.startswith('how'):
        segments['how'] = query
    else:
        # Remove location and time, rest is "what"
        what = query_lower
        if segments['where']:
            what = what.replace(f"in ", )
            what = what.replace(, )
         segments[]:
            what = what.replace(segments[], )
        segments[] = what.strip()

     segments


query = 
segments = segment_query(query)
()

Pattern 5: Query Relaxation

When to use: No results for strict query, relax constraints

def relaxed_search(query, search_fn, min_results=3):
    """Progressive query relaxation if no results"""
    # Try exact query first
    results = search_fn(query)

    if len(results) >= min_results:
        return results, "exact"

    # Relaxation strategies
    strategies = [
        ('remove_quotes', lambda q: q.replace('"', '')),
        ('remove_stopwords', lambda q: remove_stopwords(q)),
        ('expand_synonyms', lambda q: expand_with_synonyms(q)),
        ('stem_terms', lambda q: stem_query(q)),
    ]

    for strategy_name, strategy_fn in strategies:
        relaxed_query = strategy_fn(query)
        results = search_fn(relaxed_query)

        if len(results) >= min_results:
            return results, strategy_name

    # Ultimate fallback: search individual terms (OR query)
    terms = query.split()
    results = search_fn(" OR ".join(terms))

    return results, "individual_terms"

def remove_stopwords(query):
    stopwords = {'the', 'a', 'an', 'is', 'are', 'in', 'on', 'at'}
    return ' '.join([w for w in query.split()  w.lower()   stopwords])

---

Quick Reference

Query Expansion Techniques

Technique                | Use Case                   | Pros              | Cons
-------------------------|----------------------------|-------------------|-------------
Synonyms                 | Vocabulary mismatch        | Simple, fast      | Limited coverage
Pseudo-relevance feedback| High-quality corpus        | Domain-adaptive   | Query drift
Word embeddings          | Semantic similarity        | Broader coverage  | May add noise
User click data          | Large query logs           | User-validated    | Needs logs

Intent Classification

Intent Type      | Indicators                       | Search Strategy
-----------------|----------------------------------|------------------
Informational    | what, how, why, definition       | Return documents, snippets
Navigational     | brand name, login, homepage      | Return direct link
Transactional    | buy, order, download, book       | Return products, CTAs

Key Guidelines

✅ DO: Spell-check queries before search
✅ DO: Expand queries to bridge vocabulary gap
✅ DO: Classify intent to customize results
✅ DO: Extract entities for structured search
✅ DO: Suggest queries as user types (autocomplete)
✅ DO: Handle multilingual queries

❌ DON'T: Over-expand queries (query drift)
❌ DON'T: Ignore user's original terms
❌ DON'T: Auto-correct without showing original
❌ DON'T: Suggest irrelevant queries
❌ DON'T: Apply same processing to all languages

---

Anti-Patterns

Critical Violations

# ❌ NEVER: Auto-correct without showing original query
def bad_search(query):
    corrected = spell_correct(query)
    return search(corrected)
    # User doesn't know query was changed

# ✅ CORRECT: Show correction, let user confirm
def good_search(query):
    corrected = spell_correct(query)
    if corrected != query:
        # Return both original and corrected results
        return {
            'original_results': search(query),
            'did_you_mean': corrected,
            'corrected_results': search(corrected)
        }
    return {'results': search(query)}

❌ Silent auto-correction: User confusion, wrong results ✅ Correct approach: Show "Did you mean?" with original results

Common Mistakes

# ❌ Don't: Expand query too aggressively
expanded = original_query + " " + " ".join(all_synonyms)
# Query drift: "python programming" → "python programming snake coding development"

# ✅ Correct: Selective expansion with scoring
def selective_expansion(query, synonyms, threshold=0.8):
    expanded = query
    for term in query.split():
        if term in synonyms:
            # Add only high-confidence synonyms
            for syn, score in synonyms[term]:
                if score > threshold:
                    expanded += f" {syn}"
    return expanded

❌ Over-expansion: Query drift, irrelevant results ✅ Better: Use confidence thresholds, limit expansion

# ❌ Don't: Ignore query structure
query = "machine learning NOT deep learning"
processed = " ".join(query.split())  # Loses NOT operator

# ✅ Correct: Preserve operators
def parse_query(query):
    # Preserve boolean operators, quotes, etc.
    tokens = []
    operators = {'AND', 'OR', 'NOT'}

    for token in query.split():
        if token.upper() in operators:
            tokens.append(token.upper())
        else:
            tokens.append(token)

    return tokens

❌ Destroying query structure: Loses user intent ✅ Better: Respect operators, quotes, special syntax

---

Related Skills

• ir-search-fundamentals.md - Use expanded queries with BM25/Elasticsearch
• ir-vector-search.md - Semantic query understanding with embeddings
• ir-ranking-reranking.md - Rank results after query understanding
• ir-recommendation-systems.md - Query-based recommendations
• frontend-forms-validation.md - Client-side query validation

---

Last Updated: 2025-10-25 Format Version: 1.0 (Atomic)