Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-Task Contrastive Learning for 8192-Token Bilingual Text Embeddings

Isabelle Mohr, Markus Krimmel, Saba Sturua, Mohammad Kalim Akram, Andreas Koukounas, Michael Günther, Georgios Mastrapas, Vinit Ravishankar, Joan Fontanals Martínez, Feng Wang, Qi Liu, Ziniu Yu, Jie Fu, Saahil Ognawala, Susana Guzman, Bo Wang, Maximilian Werk, Nan Wang, Han Xiao

TL;DR

The paper tackles the inefficiency and limited performance of multilingual text embeddings by proposing bilingual long-context models that pair English with a target language and support up to 8192 tokens. It combines a three-stage training pipeline (pre-training a JinaBERT-based backbone with language-pair vocabularies, followed by text-pair embedding fine-tuning and a multi-task stage for STS and retrieval) with a novel hard parameter-sharing strategy and a bi-directional InfoNCE-based loss. Empirical results show bilingual models outperform multilingual baselines on GLUE, XTREME, and MTEB benchmarks, with significant gains in cross-lingual retrieval and STS when using the multi-task objective; Spanish and German variants also enrich MTEB with new German/Spanish tasks. The work demonstrates that targeted bilingual backbones, longer context windows, and task-specific fine-tuning yield efficient, high-performing embeddings suitable for retrieval, clustering, and semantic similarity across languages, while reducing parameter counts through language-focused vocabularies.

Abstract

We introduce a novel suite of state-of-the-art bilingual text embedding models that are designed to support English and another target language. These models are capable of processing lengthy text inputs with up to 8192 tokens, making them highly versatile for a range of natural language processing tasks such as text retrieval, clustering, and semantic textual similarity (STS) calculations. By focusing on bilingual models and introducing a unique multi-task learning objective, we have significantly improved the model performance on STS tasks, which outperforms the capabilities of existing multilingual models in both target language understanding and cross-lingual evaluation tasks. Moreover, our bilingual models are more efficient, requiring fewer parameters and less memory due to their smaller vocabulary needs. Furthermore, we have expanded the Massive Text Embedding Benchmark (MTEB) to include benchmarks for German and Spanish embedding models. This integration aims to stimulate further research and advancement in text embedding technologies for these languages.

Multi-Task Contrastive Learning for 8192-Token Bilingual Text Embeddings

TL;DR

The paper tackles the inefficiency and limited performance of multilingual text embeddings by proposing bilingual long-context models that pair English with a target language and support up to 8192 tokens. It combines a three-stage training pipeline (pre-training a JinaBERT-based backbone with language-pair vocabularies, followed by text-pair embedding fine-tuning and a multi-task stage for STS and retrieval) with a novel hard parameter-sharing strategy and a bi-directional InfoNCE-based loss. Empirical results show bilingual models outperform multilingual baselines on GLUE, XTREME, and MTEB benchmarks, with significant gains in cross-lingual retrieval and STS when using the multi-task objective; Spanish and German variants also enrich MTEB with new German/Spanish tasks. The work demonstrates that targeted bilingual backbones, longer context windows, and task-specific fine-tuning yield efficient, high-performing embeddings suitable for retrieval, clustering, and semantic similarity across languages, while reducing parameter counts through language-focused vocabularies.

Abstract

We introduce a novel suite of state-of-the-art bilingual text embedding models that are designed to support English and another target language. These models are capable of processing lengthy text inputs with up to 8192 tokens, making them highly versatile for a range of natural language processing tasks such as text retrieval, clustering, and semantic textual similarity (STS) calculations. By focusing on bilingual models and introducing a unique multi-task learning objective, we have significantly improved the model performance on STS tasks, which outperforms the capabilities of existing multilingual models in both target language understanding and cross-lingual evaluation tasks. Moreover, our bilingual models are more efficient, requiring fewer parameters and less memory due to their smaller vocabulary needs. Furthermore, we have expanded the Massive Text Embedding Benchmark (MTEB) to include benchmarks for German and Spanish embedding models. This integration aims to stimulate further research and advancement in text embedding technologies for these languages.
Paper Structure (23 sections, 4 equations, 11 tables)

This paper contains 23 sections, 4 equations, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multilingual Language Models
  4. Multi-Task Learning
  5. Training Paradigm Overview
  6. Pre-training a Modified BERT
  7. Model Architecture
  8. Training Data
  9. Training Algorithm
  10. Fine-Tuning for the Embedding Task
  11. Fine-tuning with Text Pairs
  12. Training Data
  13. Loss Function
  14. Fine-tuning with Multi-Task Objective
  15. Training Data
  16. ...and 8 more sections