How Gaode Map Boosts POI Search Accuracy with Advanced Query Analysis and AI

Gaode Map’s search functionality can be summarized as three core parts—location, search and navigation—each addressing the questions of where, what and how to get there.

Overall Technical Architecture

The architecture mirrors a generic retrieval system and consists of three main stages: query analysis, recall and ranking. Query analysis not only uses general NLP methods such as tokenization, synonym expansion and spelling correction, but also incorporates map‑specific intents like city analysis, where‑what parsing and route‑planning understanding.

Query Analysis Techniques

Map query analysis includes both generic components (segmentation, part‑of‑speech, synonym, error correction) and specialized components (city analysis, where‑what analysis, path‑planning analysis). The goal is to precisely identify the user’s intended POI and improve satisfaction, which is the most critical metric for search effectiveness.

Spelling Correction

Approximately 6%–10% of map search queries contain errors. The correction pipeline combines pinyin‑based recall, character‑shape recall and translation‑model‑based replacement, addressing low‑frequency cases that traditional rule‑based methods miss.

Rewrite (Query Rewriting)

To handle low‑frequency queries that cannot be corrected directly, a rewrite module transforms them into semantically similar high‑frequency queries using sentence embeddings (SIF) and Faiss‑based vector retrieval, followed by a ranking model (GBRank) and filtering with alignment models (fastalign).

Omission Detection

Many map queries contain stop words that hinder effective recall. An omission module identifies core terms and performs a lightweight recall when primary strategies fail, improving robustness for queries like "厦门市搜湖里区县后高新技术园新捷创运营中心11楼1101室".

City Analysis

Accurately recognizing the target city from user queries is essential. The upgraded pipeline treats city analysis as a two‑stage task—light recall followed by a GBDT‑based binary classifier—using query‑level and phrase‑level features, as well as engineered combination features.

Where‑What Analysis

Where‑what parsing separates spatial descriptors (where) from target entities (what). The system uses a CRF model for sequence labeling and a GBDT model for post‑hoc intent selection, addressing challenges such as low‑frequency queries, inverse order expressions, and ambiguous terms.

Path Planning Intent

Path‑planning queries (e.g., "从回龙观到来广营") are identified via keyword matching, then processed by a CRF model to extract start and end points. Training data are generated automatically by enriching template patterns and substituting real POIs, reducing labeling cost while maintaining high precision and recall.

Future Outlook

With machine‑learning now pervasive, further gains will focus on low‑frequency and long‑tail queries using deep seq2seq models, better knowledge integration, and robustness improvements to handle atypical expressions and adversarial cases.