# ACL Anthology: BERT

Dense two-column academic paper with many tables, figures, and links. Categories: complex-columns, text-heavy, tables, figures. Canonical source: https://aclanthology.org/N19-1423/

- Source PDF: [source.pdf](source.pdf)
- Continuous semantic HTML: [semantic-continuous/index.html](semantic-continuous/index.html)
- Side-by-side comparison: [compare.html](compare.html)
- Quality probe: [quality/quality-report.md](quality/quality-report.md)
- Pages: 16
- Text runs: 7431
- Text lines: 1236
- Image placements: 38
- Vector paths: 983
- Exported assets: 15
- Links: 285
- Diagnostics: 22
- Quality status: needs-review
- Quality checks needing review: 7

## Quality Artifacts

- [diff p1](quality/page-1-diff.png)
- [visual report p1](quality/page-1-visual-report.html)
- [diff p2](quality/page-2-diff.png)
- [color heatmap p2](quality/page-2-color-heatmap.png)
- [visual report p2](quality/page-2-visual-report.html)

## Text Preview

BERT:Pre-trainingofDeepBidirectionalTransformersfor LanguageUnderstanding JacobDevlinMing-WeiChangKentonLeeKristinaToutanova GoogleAILanguage {jacobdevlin,mingweichang,kentonl,kristout}@google.com AbstractTherearetwoexistingstrategiesforapply- ingpre-trainedlanguagerepresentationstodown- Weintroduceanewlanguagerepresenta- streamtasks:feature-basedandfine-tuning.The tionmodelcalledBERT,whichstandsfor feature-basedapproach,suchasELMo(Peters BidirectionalEncoderRepresentationsfrom etal.,2018a),usestask-specificarchitecturesthat Transformers.Unlikerecentlanguagerepre- sentationmodels(Petersetal.,2018a;Rad-includethepre-trainedrepresentationsasaddi- fordetal.,2018),BERTisdesignedtopre-tionalfeature...