Learning Which Features Matter: RoBERTa Acquires a Preference for Linguistic Generalizations (Eventually)

One reason pretraining on self-supervised linguistic tasks is effective is\nthat it teaches models features that are helpful for language understanding.\nHowever, we want pretrained models to learn not only to represent linguistic\nfeatures, but also to use those features preferentially during fine-turning.\nWith this goal in mind, we introduce a new English-language diagnostic set\ncalled MSGS (the Mixed Signals Generalization Set), which consists of 20\nambiguous binary classification tasks that we use to test whether a pretrained\nmodel prefers linguistic or surface generalizations during fine-tuning. We\npretrain RoBERTa models from scratch on quantities of data ranging from 1M to\n1B words and compare their performance on MSGS to the publicly available\nRoBERTa-base. We find that models can learn to represent linguistic features\nwith little pretraining data, but require far more data to learn to prefer\nlinguistic generalizations over surface ones. Eventually, with about 30B words\nof pretraining data, RoBERTa-base does demonstrate a linguistic bias with some\nregularity. We conclude that while self-supervised pretraining is an effective\nway to learn helpful inductive biases, there is likely room to improve the rate\nat which models learn which features matter.\n