Abstract
Point of Interest (POI) retrieval is a critical component of location-based services, yet it faces challenges from typos, incomplete queries, and complex input methods. This paper introduces the Deep POI Semantic Model (DPSM), a convolutional neural network architecture developed by researchers at DiDi Chuxing. DPSM leverages a dual-tower structure to model both character-level and word-level features across POI names and addresses. Experimental results on massive real-world datasets demonstrate that DPSM significantly outperforms traditional semantic matching methods like DSSM, achieving superior relevance and accuracy in production environments.
Introduction
When a user enters a query like "Starbucks" into a navigation application, the system must match this input against a vast database of locations. This task is complicated by several real-world factors: users frequently make typos, provide incomplete queries, or use aliases. Furthermore, in regions like China, input methods often involve a mix of Chinese characters and Pinyin (Romanized Chinese), adding another layer of complexity. Traditional lexical matching, which relies on exact string overlaps, often fails in these scenarios. While existing semantic models like the Deep Structured Semantic Model (DSSM) offer improvements, they often overlook the specific structure of POI data, which typically consists of both a name and a detailed address.
Methodology: The DPSM Architecture
To address these challenges, the researchers proposed the Deep POI Semantic Model (DPSM). The model employs a "dual-tower" architecture, comprising a Query Model and a POI Model.
Multi-Granularity Representation
DPSM avoids the trade-off between character-level and word-level features by utilizing both. Character-level features are effective for capturing incomplete queries and Pinyin inputs, while word-level features, segmented using a Hidden Markov Model (HMM), capture deeper semantic meanings.
Deep Convolutional Structure
Both towers in the DPSM architecture utilize a deep convolutional structure. A word embedding layer is placed before the convolution to reduce kernel size, followed by a max-pooling layer. This configuration allows the model to extract the most salient "n-gram" features regardless of the input query's length, enhancing its robustness.
Multi-Field POI Modeling
A key innovation of DPSM is its treatment of POI data. Unlike previous models that focus solely on the POI name, DPSM processes the POI Name and POI Address as separate fields. These are modeled independently and then concatenated into a final POI vector, allowing the model to resolve ambiguities that name-only models might miss.
Training with NCE Sampling
To train the model effectively on potentially biased click data, the researchers employed Noise Contrastive Estimation (NCE). This technique helps generate high-quality negative samples based on POI popularity, ensuring the model learns to distinguish relevant results from popular but irrelevant ones.
Experimental Evaluation
The model was trained and tested using massive real-world datasets from DiDi’s logs, including a training set of 10 million clicked POIs, a click test set of 1 million samples, and a human-annotated relevance test set of 60,000 pairs.
Performance Results
DPSM consistently outperformed state-of-the-art baselines, including DSSM and the Convolutional Latent Semantic Model (CDESM).
On the click dataset, DPSM achieved an NDCG@10 of 0.8812, surpassing the baseline DSSM score of 0.8726. On the human-labeled relevance dataset, it achieved an NDCG@10 of 0.6333, compared to 0.6099 for DSSM. Beyond accuracy, the model is highly efficient, maintaining a response time of approximately 15ms under heavy load (1,000 queries per second), leading to its successful deployment in DiDi’s production environment.
Key Insights and Discussion
- The Importance of Address: Explicitly modeling the address field is crucial for semantic matching, as it provides vital context that names alone may lack.
- Robustness through Hybrid Granularity: Combining character and word embeddings makes the model resilient to the "Out-of-Vocabulary" (OOV) problem common in real-world search.
- Neural Traceability: Analysis of neuron activations revealed that high-similarity matches share specific "high activation" neurons in the max-pooling layer, suggesting the model develops a structured understanding of text field relationships.
Conclusion and Future Work
DPSM represents a significant advancement in POI retrieval by effectively leveraging the multi-field nature of POI data and a robust convolutional architecture. Future work may involve incorporating additional contextual features such as user location, time of day, or historical preferences to further personalize search results. Additionally, exploring the model's effectiveness in other languages with different input complexities remains an area for further research.
Summary updated