Title: SEA-Embedding: An Open and Reproducible Framework for Text Embeddings in Southeast Asia

Text embeddings serve as a critical foundation for numerous downstream applications, rendering their robustness essential for practical natural language processing (NLP). Yet, current state-of-the-art embedding models often lack reproducibility due to their dependence on closed or undisclosed training datasets. Furthermore, these models frequently demonstrate inadequate resilience when applied to Southeast Asian languages. To address these challenges, we introduce SEA-Embedding, a fully transparent and reproducible text-embedding pipeline designed specifically for Southeast Asian languages. This system is trained exclusively on publicly accessible data. We utilize this framework to investigate three fundamental components of robust embedding design: base encoder initialization, training objectives, and data composition. SEA-Embedding not only delivers state-of-the-art performance on the SEA-BED benchmark but also facilitates systematic and reproducible analysis of robust text embeddings tailored to the region.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC