Latin American Spanish training data for a European e-commerce group

The challenge

The group had been treating Latin American Spanish as one thing with accents, training on data translated into regional variants that was grammatically regional but pragmatically Castilian, which improved benchmark scores while leaving real-world performance flat. Its intent, search and moderation models misclassified across variants.

Approach and methodology

• Organised 58 annotators into five variant desks: Caribbean, Mexican, Andean, Rioplatense and Castilian
• Extended the label schema to capture pragmatic features, not just semantics
• Drew on Punta Cana's genuinely multi-variant, pan-Latin-American labour pool
• Built a divergence taxonomy across the Spanish variants

Results

Real-world model performance improved where benchmark-driven work had not, because the annotation captured the pragmatic divergence, directness, politeness register and urgency signalling, that was the actual source of misclassification. Full metrics are in the downloadable portfolio.

The complete metrics table and charts for this engagement are in the downloadable portfolio.

Enduring value

The variant divergence taxonomy is a client-owned asset now shaping the group's Latin American product strategy, and the five-variant annotation model is a Corpshore Dominicana standard.