SciBERT: A Pretrained Language Model for Scientific Text
Open accessThe DOI matches its Crossref record; 37 of the 37 references it cites were found in a registry or an index; 9 of the 33 sources its claims cite were checked for support.
Abstract
Iz Beltagy, Kyle Lo, Arman Cohan. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.
Overview
CoChat Summary
TL;DR
SciBERT is a BERT-based pretrained language model trained on a 3.17B-token corpus of 1.14M scientific papers, with an in-domain WordPiece vocabulary. It improves over BERT-Base across scientific NLP tasks and achieves new state-of-the-art results on several datasets.
Key findings
- SciBERT outperforms BERT-Base on scientific tasks by +2.11 F1 with finetuning and +2.43 F1 without finetuning.
- Domain breakdowns: biomedical +1.92 F1 with finetuning and +3.59 F1 without; computer science +3.55 F1 with finetuning and +1.13 F1 without; multidomain +0.49 F1 with finetuning and +0.93 F1 without.
- Finetuning beats frozen BERT embeddings by +3.25 F1 with SciBERT and +3.58 F1 with BERT-Base on average.
- Using SCIVOCAB instead of BASEVOCAB gives +0.60 F1 averaged across datasets: +0.76 F1 biomedical, +0.61 F1 computer science, +0.11 F1 multidomain. SciBERT achieves new SOTA on BC5CDR, ChemProt, EBM-NLP, ACL-ARC, the NER part of SciERC, and SciCite.
Methods
- SciBERT follows the BERT-Base architecture and is pretrained on a random sample of 1.14M Semantic Scholar papers: 18% computer science and 82% broad biomedical, using full text, averaging 154 sentences (2,769 tokens) per paper, for 3.17B tokens.
- The authors build SCIVOCAB, a 30K WordPiece vocabulary with 42% token overlap with BASEVOCAB, and train four variants: cased/uncased and BASEVOCAB/SCIVOCAB.
- Evaluation covers NER, PICO extraction, text classification, relation classification, and dependency parsing, using datasets including EBM-NLP, SciERC, ACL-ARC, SciCite, and Paper Field; results are macro F1 for most tasks, micro F1 for ChemProt, and LAS/UAS for DEP, averaged over multiple random seeds.
Limitations
- The paper states future work will release a BERT-Large analogue and experiment with different proportions of papers from each domain, noting these models are costly to train and aiming for a single resource useful across multiple domains.
- The paper notes SciBERT performs slightly worse than SOTA on JNLPBA, NCBI-disease, and GENIA, attributing these to comparison models: a BiLSTM-CRF ensemble trained on multiple NER datasets, BIOBERT trained on 18B biomedical tokens, and a parser using POS features not used here.
- Evident from the text: the pretraining corpus is limited to computer science and biomedical papers, so coverage of other scientific domains is not established; frozen-embedding experiments did not use extensive hyperparameter search; and SciERC relation results are not comparable to prior work because this paper uses gold entities while prior work performs joint entity and relation extraction.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
9 Valid · 0 Not valid · 16 Not checked · 29 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Yet while both BERT and ELMo have released pretrained models, they are still trained on general domain corpora such as news articles and Wikipedia.”
§ SCIBERT: A Pretrained Language Model for Scientific Text > 1 Introduction
What the sources say
- Not checked — no full text of the sourcecites Devlin et al., 2019 source
- Partly supportedcites Peters et al., 2018 sourceExperimental — accuracy not yet measured
Section 3.4 says the ELMo biLM was pretrained on the 1B Word Benchmark, and the Introduction says the trained models were made publicly available. The paper does not discuss BERT or Wikipedia, or identify the benchmark text as news articles, so it does not support those parts of the sentence. (LLM, conf=0.86)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“As shown through ELMo (Peters et al., 2018), GPT (Radford et al., 2018) and BERT (Devlin et al., 2019), unsupervised pretraining of language models on large corpora significantly improves performance on many NLP tasks.”
§ SCIBERT: A Pretrained Language Model for Scientific Text > 1 Introduction
What the sources say
- Supportedcites Peters et al., 2018 source
Section 2 describes pretraining ELMo’s bidirectional language model on unlabeled data, and Section 3.4 identifies the large 1B Word Benchmark training corpus. Section 4 reports that adding ELMo improved results across six NLP benchmark tasks, with relative error reductions of 6–20%. This supports the claim as it relates to ELMo; the other cited models are not attributed to this source. (LLM, conf=0.92)
- Not matchedcites Radford et al., 2018
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not checked — no full text of the sourcecites Devlin et al., 2019 source
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“SCIBERT significantly outperformed BERT-Base and achieves new SOTA results on several of these tasks, even compared to some reported BIOBERT (Lee et al., 2019) results on biomedical tasks.”
§ SCIBERT: A Pretrained Language Model for Scientific Text > 7 Conclusion and Future Work
What the sources say
- Supportedcites Lee et al., 2019 source
The citation identifies the source of the reported BioBERT results used as a comparator, rather than attributing the SciBERT comparison to Lee et al. The Experimental results (§4.3) report BioBERT performance on biomedical named-entity recognition, relation extraction, and question-answering tasks. (LLM, conf=0.90)
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“In Table 2, we compare SCIBERT results with reported BIOBERT results on the subset of datasets included in (Lee et al., 2019) BC5CDR and ChemProt, and performs similarly on JNLPBA despite being trained on a substantially smaller biomedical corpus.”
§ SCIBERT: A Pretrained Language Model for Scientific Text > 4.1 Biomedical Domain
What the sources say
- Supportedcites Lee et al., 2019 source
The results tables in our copy report BioBERT scores for BC5CDR, ChemProt, and JNLPBA (section 6; page unavailable). The citation identifies the source of those reported BioBERT results; the comparison with SciBERT is the citing authors’ own conclusion. (LLM, conf=0.90)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“In addition, SCIBERT outperforms the SOTA on SciCite (Co-han et al., 2019).”
§ SCIBERT: A Pretrained Language Model for Scientific Text > 4.3 Multiple Domains
What the sources say
- Partly supportedcites Cohan et al., 2019 sourceExperimental — accuracy not yet measured
Table 4 (section 4.3 Results; page unavailable) shows that the paper’s BiLSTM-Attn model with ELMo and both scaffolds outperformed the listed previous SOTA on SciCite (84.0 versus 79.6 macro F1). It does not report a SciBERT result, so it does not support attributing that improvement to SciBERT. (LLM, conf=0.85)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
- Partly supported
17 more claims
Create a free account to see all 54 claims and to verify the 13 references still open.
- Probably not a substantial claim. 17
- It may state two separate claims as one. 17
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“With their word-piece vocabulary, contextual embeddings, and self-supervised pre-training on large unlabelled corpora these models have essentially removed the need for most pre-processing beyond automatically-applied lowercasing. 14,31”
From Data extraction methods for systematic review (semi)automation: Update of a living systematic review (2025) · cites this paper as [31]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“SciBERT, introduced by Beltagy et al. [16], was the first language model tailored for scientific texts.”
From Evaluating Large Language Models for Structured Science Summarization in the Open Research Knowledge Graph (2024) · cites this paper as [16]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“Hence, we base our experiments on well-established large language models (LLM) from the general and scientific domain -RoBERTa (Liu et al., 2019), BioBERT (Lee et al., 2020) and SciBERT ( Beltagy et al., 2019) -with the recommended default hyperparameter settings.”
From NLPeer: A Unified Resource for the Computational Study of Peer Review (2023) · cites this paper as Beltagy et al., 2019
The cited source says what the claim attributes to it.
In the citing paper
“Although the first experiments with BERT ( 39 ) were disappointing, fine-tuning SciBERT ( 40 ) led to the best results and is the current production architecture of the platform.”
From scite: a smart citation index that displays the context of citations and classifies their intent using deep learning (2021) · cites this paper as [40]
The cited source says what the claim attributes to it.
In the citing paper
“Similar observations have been made in prior work, e.g. Beltagy et al. (2019) found that pre-training BERT on text from research papers improved its performance on scientific tasks.”
From Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (2019) · cites this paper as Beltagy et al. (2019)
7 more claims about this paper
Create a free account to see every claim other papers make about this one.
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
Adam: A Method for Stochastic Optimization
Kingma, D. P., & Ba, J. (2014) · arXiv · doi:10.48550/arxiv.1412.6980 · Cited by 98,946
Exploiting Generative AI to Scale up Intelligent Tutoring Systems
Jakubův, J., Chvalovský, K., Goertzel, Z., Kaliszyk, C., Olšák, M., Piotrowski, B., Schulz, S., Suda, M., & Urban, J. (2023) · DROPS (Schloss Dagstuhl – Leibniz Center for Informatics) · doi:10.4230/lipics.itp.2023.19 · Cited by 76,658
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325 · cited 7×
“3617 Casing We follow Devlin et al. (2019) in using the cased models for NER and the uncased models for all other tasks.”
BackgroundMethodInfluentialDeep Contextualized Word Representations
Peters, M., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., & Zettlemoyer, L. (2018) · Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) · doi:10.18653/v1/n18-1202 · Cited by 11,983
MIMIC-III, a freely accessible critical care database
Johnson, A. E. W., Pollard, T. J., Shen, L., Lehman, L.-W. H., Feng, M., Ghassemi, M., Moody, B., Szolovits, P., Anthony Celi, L., & Mark, R. G. (2016) · Scientific Data · doi:10.1038/sdata.2016.35 · Cited by 8,141
BioBERT: a pre-trained biomedical language representation model for biomedical text mining
Lee, J., Yoon, W., Kim, S., Kim, D., Kim, S., So, C. H., & Kang, J. (2020) · Bioinformatics · doi:10.1093/bioinformatics/btz682 · Cited by 7,218 · cited 5×
“Recent work on domain adaptation of BERT includes B IO B ERT (Lee et al., 2019) and C LINI - CAL B ERT (Alsentzer et al., 2019; Huang et al., 2019).”
MethodResultInfluentialGoogle's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation
Wu, Y., Schuster, M., Chen, Z., Le, Q. V., Norouzi, M., Macherey, W., Krikun, M., Cao, Y., Gao, Q., Macherey, K., et al. (2016) · arXiv · doi:10.48550/arxiv.1609.08144 · Cited by 5,319
Universal Language Model Fine-tuning for Text Classification
Howard, J., & Ruder, S. (2018) · Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/p18-1031 · Cited by 4,100 · cited 2×
“We finetune for 2 to 5 epochs using a batch size of 32 and a learning rate of 5e-6, 1e5, 2e-5, or 5e-5 with a slanted triangular schedule (Howard and Ruder, 2018) which is equivalent to the linear warmup followed by linear decay (Devlin et al.”
Method
Cited by
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929
Transformers: State-of-the-Art Natural Language Processing
Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al. (2020) · Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations · doi:10.18653/v1/2020.emnlp-demos.6 · Cited by 8,741
“We highlight specific systems developed by users with different goals following the architect, trainer, and end-user distinction of Strobelt et al. (2017): Case 1: Model Architects AllenAI, a major NLP research lab, developed a new pretrained model for improved extraction from biomedical texts called SciBERT (Beltagy…”
InfluentialOn the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021) · Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3442188.3445922 · Cited by 6,988
“However, in both application areas, the training data has been shown to have problematic characteristics [18, 38, 42, 47, 61] resulting in models that encode stereotypical and derogatory associations along gender, race, ethnicity, and disability status [11, 12, 69, 69, 132, 132, 157].”
BackgroundHan (2023) · doi:10.1109/tpami.2022.3152247 · Cited by 4,682
“BioBERT [69] uses Transformer architecture for biomedical text mining tasks; SciBERT [7] is developed by training Transformer on 114M scientific articles covering biomedical and computer science field, aiming to execute NLP tasks related to scientific domain more precisely; Huang et al. [55] proposes ClinicalBERT util…”
MethodHuggingFace's Transformers: State-of-the-art Natural Language Processing
Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al. (2019) · arXiv · doi:10.48550/arxiv.1910.03771 · Cited by 3,347
Metadata
- Authors
- Beltagy, I., Lo, K., & Cohan, A.
- Affiliations
- Allen Institute for Artificial Intelligence (US)
- ORCID
- Venue
- Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)
- Published
- 2019
- Pages
- 3613-3618
- OpenAlex
- W2970771982
- Semantic Scholar
- 156d217b0a911af97fa1b5a71dc909ccef7a8028
- Cited by
- 3,218
- References
- 33
- Citations by year
- 2026: 328 · 2025: 534 · 2024: 542 · 2023: 626 · 2022: 478 · 2021: 486 · 2020: 203 · 2019: 19
- FWCI
- 113.46
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- Locations
- Conference on Empirical Methods in Natural Language Processing (EMNLP) (publishedVersion, cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Keywords
- SciBERT, pre-trained language models, scientific texts, scientific publications, biomedical text, natural language processing
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-11-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-11-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: SciBERT: A Pretrained Language Model for Scientific Text. https://app.cochat.ai/papers/10.18653/v1/d19-1371
- The paper
- Beltagy, I., Lo, K., & Cohan, A. (2019). SciBERT: A Pretrained Language Model for Scientific Text. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). https://doi.org/10.18653/v1/d19-1371
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.