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

TL;DR

This paper investigates how pretraining data size influences RoBERTa's ability to prefer linguistically meaningful features over surface cues during fine-tuning, introducing a diagnostic set to measure this bias.

Contribution

It introduces MSGS, a diagnostic set for testing linguistic versus surface generalizations, and analyzes how different pretraining data sizes affect RoBERTa's feature preferences.

Findings

RoBERTa learns linguistic features with little data

Large pretraining data (~30B words) leads to a linguistic bias

Pretraining data size impacts the rate of acquiring feature preferences

Abstract

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