Task-Conditional Flow Matching Enhances Multilingual Text Embedding Adaptation
A new framework named Task-Conditional Flow Matching (TCFM) has been developed by researchers to enhance multilingual text embedding models. In contrast to conventional approaches that utilize a uniform training objective for various tasks, TCFM strategically applies Flow Matching specifically to translation tasks, while opting for more appropriate objectives for retrieval, classification, and pair-classification tasks. Additionally, it incorporates teacher-guided representation preservation and a three-stage curriculum to facilitate stable adaptation. When tested on the Indic Massive Text Embedding Benchmark, TCFM demonstrated state-of-the-art results, significantly enhancing embedding quality across multiple multilingual tasks and showing versatility among different embedding model families. The codebase and datasets will be made publicly available following paper acceptance. The paper can be found on arXiv with the identifier 2608.05785.
Key facts
- TCFM is a new framework for multilingual text embedding adaptation.
- It selectively applies Flow Matching to translation tasks.
- Retrieval, classification, and pair-classification tasks use objectives aligned with their learning dynamics.
- TCFM combines teacher-guided representation preservation with a three-stage curriculum.
- Evaluation on the Indic Massive Text Embedding Benchmark shows state-of-the-art results.
- TCFM improves embedding quality across diverse multilingual tasks.
- The framework generalizes across embedding model families.
- Codebase and datasets will be released upon acceptance of the paper.
Entities
Institutions
- arXiv