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

arXiv: https://arxiv.org/abs/2010.05358 · HTML: https://mlsanity.com/paper/2010.05358/

## Key facts

- Cited by 12 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 11 Oct 2020.

## Built on by (most cited first, 12 of 12)

- [Measuring and Improving Consistency in Pretrained Language Models](https://mlsanity.com/paper/2102.01017/) (2021)
- [Masked Language Modeling and the Distributional Hypothesis: Order Word Matters Pre-training for Little](https://mlsanity.com/paper/2104.06644/) (2021)
- [Transformer Grammars: Augmenting Transformer Language Models with Syntactic Inductive Biases at Scale](https://mlsanity.com/paper/2203.00633/) (2022)
- [Neural reality of argument structure constructions](https://mlsanity.com/paper/2202.12246/) (2022)
- [Unit Testing for Concepts in Neural Networks](https://mlsanity.com/paper/2208.10244/) (2022)
- [Language model acceptability judgements are not always robust to context](https://mlsanity.com/paper/2212.08979/) (2022)
- [How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech](https://mlsanity.com/paper/2301.11462/) (2023)
- [Distilling Relation Embeddings from Pre-trained Language Models](https://mlsanity.com/paper/2110.15705/) (2021)
- [Uncovering Constraint-Based Behavior in Neural Models via Targeted Fine-Tuning](https://mlsanity.com/paper/2106.01207/) (2021)
- [How often are errors in natural language reasoning due to paraphrastic variability?](https://mlsanity.com/paper/2404.11717/) (2024)
- [Analysis and Prediction of NLP Models Via Task Embeddings](https://mlsanity.com/paper/2112.05647/) (2021)
- [RelBERT: Embedding Relations with Language Models](https://mlsanity.com/paper/2310.00299/) (2023)

## Abstract (arXiv)

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.