Sentence Embeddings using Supervised Contrastive Learning

Sentence embeddings encode sentences in fixed dense vectors and have played an important role in various NLP tasks and systems. Methods for building sentence embeddings include unsupervised learning such as Quick-Thoughts(Logeswaran and Lee, 2018) and supervised learning such as InferSent(Conneau et al., 2017). With the success of pretrained NLP models, recent research(Reimers and Gurevych, 2019) shows that fine-tuning pretrained BERT (Devlin et al., 2019) on SNLI(Bowman et al., 2015) and Multi-NLI(Williams et al., 2018) data creates state-of-the-art sentence embeddings, outperforming previous sentence embeddings methods on various evaluation benchmarks. In this paper, we propose a new method to build sentence embeddings by doing supervised contrastive learning. Specifically our method fine-tunes pretrained BERT on SNLI data, incorporating both supervised crossentropy loss and supervised contrastive loss. Compared with baseline where fine-tuning is only done with supervised cross-entropy loss similar to current state-of-the-art method SBERT(Reimers and Gurevych, 2019), our supervised contrastive method improves 2.8% in average on Semantic Textual Similarity (STS) benchmarks(Cer et al., 2017) and 1.05% in average on various sentence transfer tasks.