Muppet: Massive Multi-task Representations with Pre-Finetuning

Armen Aghajanyan, Anchit Gupta, Akshat Shrivastava, Xilun Chen, Luke Zettlemoyer, Sonal Gupta


Abstract
We propose pre-finetuning, an additional large-scale learning stage between language model pre-training and fine-tuning. Pre-finetuning is massively multi-task learning (around 50 datasets, over 4.8 million total labeled examples), and is designed to encourage learning of representations that generalize better to many different tasks. We show that pre-finetuning consistently improves performance for pretrained discriminators (e.g. RoBERTa) and generation models (e.g. BART) on a wide range of tasks (sentence prediction, commonsense reasoning, MRC, etc.), while also significantly improving sample efficiency during fine-tuning. We also show that large-scale multi-tasking is crucial; pre-finetuning can hurt performance when few tasks are used up until a critical point (usually above 15) after which performance improves linearly in the number of tasks.
Anthology ID:
2021.emnlp-main.468
Volume:
Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing
Month:
November
Year:
2021
Address:
Online and Punta Cana, Dominican Republic
Venue:
EMNLP
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
5799–5811
Language:
URL:
https://aclanthology.org/2021.emnlp-main.468
DOI:
10.18653/v1/2021.emnlp-main.468
Bibkey:
Cite (ACL):
Armen Aghajanyan, Anchit Gupta, Akshat Shrivastava, Xilun Chen, Luke Zettlemoyer, and Sonal Gupta. 2021. Muppet: Massive Multi-task Representations with Pre-Finetuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5799–5811, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.
Cite (Informal):
Muppet: Massive Multi-task Representations with Pre-Finetuning (Aghajanyan et al., EMNLP 2021)
Copy Citation:
PDF:
https://aclanthology.org/2021.emnlp-main.468.pdf
Video:
 https://aclanthology.org/2021.emnlp-main.468.mp4
Code
 facebook/muppet-roberta-base +  additional community code
Data
ANLIBoolQCNN/Daily MailCoLACommonsenseQAGLUEHellaSwagMultiNLIQNLIRACERTERedditReddit TIFUSQuADSSTSWAGSuperGLUE