TurkEmbed: Turkish Embedding Model on NLI & STS Tasks

TurkEmbed, a novel Turkish language embedding model, has been introduced to address the limitations of existing models that often rely on machine-translated datasets, which can hinder accuracy and semantic understanding. By employing diverse datasets and advanced training techniques such as matryoshka representation learning, TurkEmbed achieves significant improvements in Natural Language Inference (NLI) and Semantic Textual Similarity (STS) tasks. Evaluations on the Turkish STS-b-TR dataset reveal that TurkEmbed surpasses the current state-of-the-art model, Emrecan, with an improvement of 1-4%. This advancement is crucial for enhancing the Turkish NLP ecosystem, as it provides a more nuanced understanding of the language and facilitates progress in downstream applications, ultimately contributing to more robust and accurate language processing capabilities.