On Losses for Modern Language Models

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 4970–4981, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 4970 On Losses for Modern Language Models St´ephane Aroca-Ouellette1,2 and Frank Rudzicz1,2,3 1University of Toronto 2Vector Institute for Artificial Intelligence 3Li Ka Shing Knowledge Institute, St Michael’s Hospital {stephaneao, frank}@cs.toronto.edu Abstract BERT set many state-of-the-art results over varied NLU benchmarks by pre-training over two tasks: masked language modelling (MLM) and next sentence prediction (NSP), the latter of which has been highly criticized. In this paper, we 1) clarify NSP’s effect on BERT pre-training, 2) explore fourteen possi- ble auxiliary pre-training tasks, of which seven are novel to modern language models, and 3) investigate different ways to include mul- tiple tasks into pre-training. We show that NSP is detrimental to training due to its con- text splitting and shallow semantic signal. We also identify six auxiliary pre-training tasks – sentence ordering, adjacent sentence predic- tion, TF prediction, TF-IDF prediction, a Fast- Sent variant, and a Quick Thoughts variant – that outperform a pure MLM baseline. Fi- nally, we demonstrate that using multiple tasks in a multi-task pre-training framework pro- vides better results than using any single aux- iliary task. Using these methods, we outper- form BERTBase on the GLUE benchmark using fewer than a quarter of the training tokens. 1 Introduction When Devlin et al. (2018) released BERT, a trans- former network (Vaswani et al., 2017) trained using a ‘masked language model’ (MLM) task and a ‘next sentence prediction’ (NSP), it redefined the NLP landscape, establishing itself as the state-of-the-art (SoTA) on many natural language understanding (NLU) benchmarks including the GLUE (Wang et al., 2018), SQuAD (Rajpurkar et al., 2016), and SWAG (Zellers et al., 2018) benchmarks. Many models inspired by BERT have since sur- passed its performance. However, in contrast to the original BERT paper, many obtained better results by excluding the NSP task. Some, such as XLNET (Yang et al., 2019) and RoBERTa (Liu et al., 2019), rely solely on a MLM variant, while others (Wang et al., 2020; Joshi et al., 2019; Cheng et al., 2019; Sun et al., 2019b) incorporate one or more different auxiliary loss functions. To our knowledge, there is no published work comparing or fully exploring auxiliary tasks for modern language models. With multi-task learning’s long history in trans- fer learning (Caruana, 1997; Parisotto et al., 2015; Ren et al., 2018), its use in language understand- ing models deserves further exploration. In this paper, we study existing and novel auxiliary tasks in a BERT paradigm to guide future research in an informed manner. Specifically, we test and provide insight on: 1) NSP’s effect on BERT pre- training; 2) the result of 14 other auxiliary tasks on BERT pre-training; 3) how to combine multiple tasks in BERT pre-training; and 4) the advantages of multi-task learning in BERT pre-training. Al- though all experiments in this paper are conducted using BERT, we believe the results are applicable to BERT’s successors (e.g. XLNET, RoBERTa, ERNIE...) and …

Content truncated to 500 words