An Ensemble Distillation Framework for Sentence Embeddings with Multilingual Round-Trip Translation

In this work, we propose a novel unsupervised contrastive learning framework to improve state-of-the-art sentence embeddings. First, we train a set of contrastive submodels which take multilingual round-trip translation(RTT) as data augmentation. The RTT naturally changes the length of the same sentence and replaces Synonyms simultaneously. Then we incorporate them into a single model through knowledge distillation. Specifically, it takes an input sentence and predicts the ensemble output of all submodels via a contrastive objective. Thus we preserve nearly the same semantic expressiveness as the ensemble model without increasing the test cost. We evaluate our framework on standard semantic textual similarity (STS) tasks. Experimental results show the advantage of our framework that we achieve an average of 79.27% Spearman's correlation, a 3.02% improvement compared to the previous best results using BERT-base.