ML-Embed: Efficient Multilingual Text Embeddings

A new research paper introduces ML-Embed, a suite of multilingual text embedding models designed to address three barriers: high computational costs, narrow linguistic focus, and lack of transparency. The models are built on a 3-Dimensional Matryoshka Learning (3D-ML) framework, which includes Matryoshka Representation Learning (MRL) for storage efficiency, Matryoshka Layer Learning (MLL) for flexible inference, and a new Matryoshka Embedding Learning (MEL) for parameter efficiency. The authors curated a massively multilingual dataset to train the models, aiming to make embeddings more inclusive and efficient for a wide range of languages.