Large-scale Vietnamese point-of-interest classification using weak labeling.

Area of Science:

• Computer Science
• Geographic Information Systems
• Natural Language Processing

Background:

• Crowd-sourced Point-of-Interests (POI) data often suffers from low-quality category labels, impacting location-based applications.
• Existing Vietnamese POI datasets are limited in size and scope, hindering advancements in localized digital mapping.
• Accurate POI categorization is crucial for services ranging from navigation to local business discovery.

Purpose of the Study:

• To create the first large-scale, annotated dataset for categorical classification of Vietnamese Points of Interest (POIs).
• To develop and validate a weak labeling approach for efficient POI dataset creation.
• To establish a benchmark for POI categorical classification in the Vietnamese language domain.

Main Methods:

• Collected 750,000 POIs from a Vietnamese digital map (WeMap).
• Implemented a novel weak labeling strategy to overcome the limitations of manual annotation.
• Constructed a dataset comprising 275,000 weak-labeled POIs for training and 30,000 gold-standard POIs for testing.
• Utilized BERT-based fine-tuning as a strong baseline for empirical evaluation.

Main Results:

• The developed dataset is the largest of its kind for Vietnamese POIs, covering 15 categories.
• The weak labeling approach proved highly efficient for large-scale data annotation.
• The BERT-based baseline achieved a 90% F1 score on the gold-standard test set.
• Significant improvement in WeMap's POI data accuracy, increasing from 56% to 93%.

Conclusions:

• The proposed weak labeling method is effective for creating large-scale, high-quality POI datasets.
• The new dataset provides a valuable resource for advancing POI classification research in Vietnamese.
• The findings demonstrate the feasibility and scalability of automated POI categorization using deep learning models.