RoBERTa Model for Merchant Categorization at Square

This post describes Square's development of a RoBERTa-based ML model for accurately categorizing merchants into business types.

• •Square uses merchant categorization for product personalization, business strategy, growth targeting, product eligibility, and accurate Merchant Category Codes (MCCs) for payment processing
• •Historical self-selection during onboarding is highly inaccurate due to merchants rushing through signup, overly granular categories, and unclear category definitions
• •Previous ML approaches relied on outdated methods and over-indexed on self-selected labels as ground truth
• •The new RoBERTa model is trained on 20,000+ manually reviewed sellers with high-quality ground truth labels
• •Data preprocessing includes removing auto-created services, ranking catalog items by purchase frequency, and formatting inputs as structured prompt strings