Subspace Chronicles: How Linguistic Information Emerges, Shifts and Interacts during Language Model Training
- ,
- ,
- ,
- Ivan Titov
- ,
- ,
- University of Amsterdam,
- University of Edinburgh
Research Output:
Conference Article in Proceeding or Book/Report chapter
Article in proceedings
Peer-reviewOpen access
Publication Information
Output type
Research Output:
Conference Article in Proceeding or Book/Report chapter
Article in proceedings
Peer-reviewOriginal language
EnglishPages from-to (Number of pages)
Pages 13190-13208Publication milestones
- Published - 06/12/2023
Publication status
Published - 06/12/2023
Place of publication
SingaporePublisher
Association for Computational Linguistics, United StatesPublication IDs
- Scopus: 85183298663
Host publication title
Findings of the Association for Computational Linguistics: EMNLP 2023Abstract
Representational spaces learned via language modeling are fundamental to Natural Language Processing (NLP), however there has been limited understanding regarding how and when during training various types of linguistic information emerge and interact. Leveraging a novel information theoretic probing suite, which enables direct comparisons of not just task performance, but their representational subspaces, we analyze nine tasks covering syntax, semantics and reasoning, across 2M pre-training steps and five seeds. We identify critical learning phases across tasks and time, during which subspaces emerge, share information, and later disentangle to specialize. Across these phases, syntactic knowledge is acquired rapidly after 0.5% of full training. Continued performance improvements primarily stem from the acquisition of open-domain knowledge, while semantics and reasoning tasks benefit from later boosts to long-range contextualization and higher specialization. Measuring cross-task similarity further reveals that linguistically related tasks share information throughout training, and do so more during the critical phase of learning than before or after. Our findings have implications for model interpretability, multi-task learning, and learning from limited data.
Publication metrics
PlumX
Citations
8
Captures
18
Access to documents
Final published version
Final published version
Related Event
Title
Conference on Empirical Methods in Natural Language Processing
Event type
ConferenceDegree of recognition
International eventDate
06/12/2023 - 10/12/2023Location
Resorts World Convention CentreSingapore
