Learning Which Features Matter: RoBERTa Acquires a Preference for Linguistic Generalizations (Eventually)
Alex Warstadt, Yian Zhang, Xiaocheng Li, Haokun Liu, Samuel R. Bowman
摘要
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 finetuning. 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. Feature type Feature description Positive example Negative example Surface Absolute position Is the first token of S "the"? The cat chased a mouse. A cat chased a mouse. Length Is S longer than n (e.g., 3) words? The cat chased a mouse. The cat meowed. Lexical content Does S contain "the"? That cat chased the mouse. That cat chased a mouse. Relative position Does "the" precede "a"? The cat chased a mouse. A cat chased the mouse. Orthography Does S appear in title case? The Cat Chased a Mouse. The cat chased a mouse. Linguistic Morphology Does S have an irregular past verb? The cats slept. The cats meow. Syn. category Does S have an adjective? Lincoln was tall. Lincoln was president. Syn. construction Is S the control construction? Sue is eager to sleep. Sue is likely to sleep. Syn. position Is the main verb in "ing" form? Cats who eat mice are purring. Cats who are eating mice purr.
问问这篇 Paper
智能体会读完全文。
Lune 把这篇 Paper 索引到了每一个公式,引用它的顶会 Paper 也一样。你提问,回答直接引用原文。
引用它的顶会 Paper23
- LIFT: Language-Interfaced Fine-Tuning for Non-language Machine Learning TasksTuan Dinh, Yuchen Zeng, Ruisu Zhang, Ziqian Lin 等NeurIPS 2022 · 被引用 222 次
- Masked Language Modeling and the Distributional Hypothesis: Order Word Matters Pre-training for LittleKoustuv Sinha, Robin Jia, Dieuwke Hupkes, Joelle Pineau 等EMNLP 2021 · 被引用 177 次
- Sudden Drops in the Loss: Syntax Acquisition, Phase Transitions, and Simplicity Bias in MLMsAngelica Chen, Ravid Shwartz-Ziv, Kyunghyun Cho, Matthew L. Leavitt 等ICLR 2024 · 被引用 119 次
- Probing for the Usage of Grammatical NumberKarim Lasri, Tiago Pimentel, Alessandro Lenci, Thierry Poibeau 等ACL 2022 · 被引用 72 次
- LLM Circuit Analyses Are Consistent Across Training and ScaleCurt Tigges, Michael Hanna, Qinan Yu, Stella BidermanNeurIPS 2024 · 被引用 71 次
它引用的顶会 Paper4
- Language Models are Few-Shot LearnersTom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah 等NeurIPS 2020 · 被引用 64,255 次
- A Systematic Assessment of Syntactic Generalization in Neural Language ModelsJennifer Hu, Jon Gauthier, Peng Qian, Ethan Wilcox 等ACL 2020 · 被引用 124 次
- Information-Theoretic Probing with Minimum Description LengthElena Voita, Ivan TitovEMNLP 2020 · 被引用 34 次
- Information-Theoretic Probing for Linguistic StructureTiago Pimentel, Josef Valvoda, Rowan Hall Maudslay, Ran Zmigrod 等ACL 2020 · 被引用 21 次
相关 Paper
- When Do You Need Billions of Words of Pretraining Data?Yian Zhang, Alex Warstadt, Xiaocheng Li, Samuel R. BowmanACL 2021
- Muppet: Massive Multi-task Representations with Pre-FinetuningArmen Aghajanyan, Anchit Gupta, Akshat Shrivastava, Xilun Chen 等EMNLP 2021 · 被引用 176 次
- Fast, Effective, and Self-Supervised: Transforming Masked Language Models into Universal Lexical and Sentence EncodersFangyu Liu, Ivan Vulic, Anna Korhonen, Nigel CollierEMNLP 2021 · 被引用 85 次
- Do Transformers Parse while Predicting the Masked Word?Haoyu Zhao, Abhishek Panigrahi, Rong Ge, Sanjeev AroraEMNLP 2023 · 被引用 5 次
- How much pretraining data do language models need to learn syntax?Laura Pérez-Mayos, Miguel Ballesteros, Leo WannerEMNLP 2021 · 被引用 31 次
