Batch-Softmax Contrastive Loss for Pairwise Sentence Scoring Tasks
This work addresses the need for improved sentence embeddings in NLP tasks like classification and ranking, though it is incremental as it adapts existing contrastive loss methods from computer vision to NLP.
The paper tackled the problem of learning better task-specific sentence embeddings for pairwise sentence scoring tasks by fine-tuning pre-trained transformer models with a batch-softmax contrastive loss, resulting in sizable improvements on various datasets and tasks including classification, ranking, and regression.
The use of contrastive loss for representation learning has become prominent in computer vision, and it is now getting attention in Natural Language Processing (NLP). Here, we explore the idea of using a batch-softmax contrastive loss when fine-tuning large-scale pre-trained transformer models to learn better task-specific sentence embeddings for pairwise sentence scoring tasks. We introduce and study a number of variations in the calculation of the loss as well as in the overall training procedure; in particular, we find that data shuffling can be quite important. Our experimental results show sizable improvements on a number of datasets and pairwise sentence scoring tasks including classification, ranking, and regression. Finally, we offer detailed analysis and discussion, which should be useful for researchers aiming to explore the utility of contrastive loss in NLP.