Abstract
In the domain of natural language augmentation, the eligibility of generated samples remains not well understood. To gather insights around this eligibility issue, we apply a transformer-based similarity calculation within the BET framework based on backtranslation, in the context of automated paraphrase detection. While providing a rigorous statistical foundation to BET, we push their results by analyzing statistically the impacts of the level of qualification, and several sample sizes. We conducted a vast amount of experiments on the MRPC corpus using six pre-trained models: BERT, XLNet, Albert, RoBERTa, Electra, and DeBerta. We show that our method improves significantly these “base” models while using only a fraction of the corpus. Our results suggest that using some of those smaller pre-trained models, namely RoBERTa base and Electra base, helps us reach F1 scores very close to their large counterparts, as reported on the GLUE benchmark. On top of acting as a regularizer, the proposed method is efficient in dealing with data scarcity with improvements of around 3% in F1 score for most pre-trained models, and more than 7.5% in the case of Electra.