While large language models dominate many NLP benchmarks, their computational cost and limited coverage for low-resource languages remain significant challenges. At NAACL 2025, CERTAIN researcher Simon Ostermann co-authored a paper introducing GrEmLIn, a centralized repository of green, static baseline embeddings for 87 mid- and low-resource languages.
GrEmLIn offers an efficient and inclusive alternative since static embeddings require no parameters at inference time, making them far more energy efficient. Despite their simplicity, GrEmLIn embeddings outperform certain LLM-based embeddings (like E5) on lexical similarity tasks and remain competitive in sentiment analysis and natural language inference. By integrating multilingual graph knowledge into enhanced GloVe embeddings, GrEmLIn shows that sustainable and inclusive AI is not only possible, but practical.