Training Universal Text Encoders with Pair Relevance Classification Loss
Keywords: sentence embeddings, text encoders, contrastive learning
Abstract: Finetuning large language models (LLMs) using contrastive learning objectives has become the dominant approach for representation learning in general-purpose text embedding tasks. Our work seeks to enable going beyond strictly positive (or negative) pairs of text, to more fine-grained annotations that can capture the nuances of complex language tasks. We propose training text encoders with a simple pair classification loss that utilizes binary cross-entropy on relevance labels. When compared to the standard softmax-based loss for multi-class classification against multiple text alternatives, we find that training with our proposed loss improves the average score across 56 English language tasks of the Massive Text Embedding Benchmark (MTEB), while finetuning the same Meta-Llama-3-8B-Instruct model on the same mix of open datasets. Furthermore, our models excel in the Pair Classification and the Semantic Textual Similarity benchmarks, outperforming many models that are trained on more extensive data. Finally, thorough experiments using graded relevance data from TREC-DL 2023 during training demonstrate that binary cross-entropy provides generalization improvements that the softmax-based loss fails to achieve.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12915
Loading