Learning Helpful Inductive Biases from Self-Supervised Pretraining

By Alex Warstadt, Yian Zhang, Haau-Sing Li, Haokun Liu, Samuel R. Bowman. 2020

28 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Representation & Self-Supervised Learning, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Representation & Self-Supervised Learning

In Optimization & Training

Citations in the corpus

When Do You Need Billions of Words of Pretraining Data? (2020, methodology): “Third, we probe the models’ knowledge of various syntactic phenomena using unsupervised acceptability judgments on the BLiMP suite (Warstadt et al., 2020a).”
Masked Language Modeling and the Distributional Hypothesis: Order Word Matters Pre-training for Little (2021, methodology): “…to parametric probes, stimulibased probing (Linzen et al., 2016; Marvin and Linzen, 2018; Gulordava et al., 2018; Goldberg, 2019; Wolf, 2019; Warstadt et al., 2019a, 2020a,b; Ettinger, 2020) has been used to show that even without a learned probing function, BERT can predict syntactic…”