Query Intent Recognition in Vertical Search: Challenges, Methods, and Case Studies

In modern search systems, simple keyword matching is insufficient; recognizing the user's intent behind a query is crucial for returning diverse, relevant results, especially in vertical domains where intent recognition often becomes category classification.

What is intent recognition? It maps a user’s query to a specific category (e.g., "Apple 6" → "Mobile‑Phones", "red apple" → "Fruits‑Apple"), influencing downstream ranking and recommendation modules.

Key difficulties include irregular input (varying phrasing), multi‑intent queries (e.g., "water" could mean mineral water or facial toner), temporal shifts in intent (seasonal or product‑life‑cycle changes), and cold‑start scenarios with limited user data.

Common practical solutions :

Lexicon enumeration: maintain a dictionary of entities (brands, locations, product types) and map segmented query terms to these types, then match predefined patterns. Example: - Query: 澳洲[addr]cemony[brand]水乳[product]面霜[sub_product] - Pattern: [brand]+[product];[addr]+[product]+[sub_product]

Statistical analysis: use offline user‑behavior logs (clicks, dwell time, purchases) to infer the most likely intent for a query, supporting features like query suggestion and related search.

Machine‑learning models: treat intent recognition as a multi‑class classification problem, using algorithms such as MaxEnt, FastText, TextCNN, Bi‑LSTM + attention; FastText is noted for rapid deployment.

Feature engineering covers part‑of‑speech tagging, named‑entity recognition, term weighting (core word emphasis), term‑user profiling, session features (query count, pagination, click behavior), query‑type attributes (presence of spaces, keyword hits, length), and side information (search tab category, time of day).

Case Study 1 – E‑commerce platform demonstrates entity recognition, query rewriting (error correction, expansion, stop‑word removal, synonym conversion), category mapping, and term weighting to improve recall and ranking.

Case Study 2 – WeChat Search shows similar pipelines: segmentation, weighting, rewriting, and intent classification using both query content and contextual signals (user location, history) before retrieving and ranking results.

General workflow for vertical search combines rule‑based coverage with model‑based prediction, often employing ensemble methods to fuse multiple offline models for the final intent classifier.

References to external articles and resources are provided at the end of the original document.