A Primer in BERTology: What We Know About How BERT Works
Open accessThe DOI matches its Crossref record; 213 of the 213 references it cites were found in a registry or an index.
Abstract
Transformer-based models have pushed state of the art in many areas of NLP, but our understanding of what is behind their success is still limited. This paper is the first survey of over 150 studies of the popular BERT model. We review the current state of knowledge about how BERT works, what kind of information it learns and how it is represented, common modifications to its training objectives and architecture, the overparameterization issue, and approaches to compression. We then outline directions for future research.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
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 cited source says what the claim attributes to it.
In the citing paper
“It is advised to apply Bidirectional Encoder Representations from Transformers (BERT) [95,144] to cope with the transferability in cross-domain for cyber grooming detection as it has profoundly bidirectional contextualization and allows the model to gain information from various representations with different positions [146].”
From Online grooming detection: A comprehensive survey of child exploitation in chat logs (2023) · cites this paper as [144]
The cited source says what the claim attributes to it.
In the citing paper
“Contextual models such as BERT 9 can represent each term based on the chat context and lead to a more profound knowledge for semantic analysis [144].”
From Online grooming detection: A comprehensive survey of child exploitation in chat logs (2023) · cites this paper as [144]
The cited source says what the claim attributes to it.
In the citing paper
“Before the emergence of ChatGPT, BERT is a commonly used pre-trained machine learning model for NLP tasks (Wolf et al., 2020), and a ubiquitous baseline in text classification experiments of research in recent years (Rogers et al., 2021).”
From Towards reliable generative AI-driven scaffolding: Reducing hallucinations and enhancing quality in self-regulated learning support (2026) · cites this paper as Rogers et al., 2021
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
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 · cited 5×
“…search-language-understanding-bert 2 Overview of BERT architecture Fundamentally, BERT is a stack of Transformer encoder layers (Vaswani et al., 2017) which consist of multiple “heads”, i.e., fully-connected neural networks augmented with a self-attention mechanism.”
BackgroundInfluentialBERT: 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 · cited 5×
“The input representations are computed as fol-Figure 1: BERT fine-tuning (Devlin et al., 2019). lows: BERT first tokenizes the given word into wordpieces (Wu et al., 2016b), and then combines three embedding layers (token, position, and segment) to obtain a fixed-length vector.”
BackgroundMethodInfluentialBERT: 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 5×
“The input representations are computed as fol-Figure 1: BERT fine-tuning (Devlin et al., 2019). lows: BERT first tokenizes the given word into wordpieces (Wu et al., 2016b), and then combines three embedding layers (token, position, and segment) to obtain a fixed-length vector.”
BackgroundMethodInfluentialGlove: Global Vectors for Word Representation
Pennington, J., Socher, R., & Manning, C. (2014) · Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.3115/v1/d14-1162 · Cited by 34,071 · cited 1×
“Unlike the conventional static embeddings (Mikolov et al., 2013a; Pennington et al., 2014), BERT’s representations are contextualized, i.e., every input token is represented by a vector dependent on the particular context of occurrence.”
BackgroundDistributed Representations of Words and Phrases and their Compositionality
Mikolov, T., Sutskever, I., Chen, K., Corrado, G., & Dean, J. (2013) · arXiv · doi:10.48550/arxiv.1310.4546 · Cited by 24,533
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
RoBERTa: A Robustly Optimized BERT Pretraining Approach
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019) · arXiv · doi:10.48550/arxiv.1907.11692 · Cited by 16,515 · cited 11×
“• Removing NSP does not hurt or slightly improves task performance (Liu et al., 2019b; Joshi et al., 2020; Clinchant et al., 2019), especially in cross-lingual setting (Wang et al., 2019b).”
BackgroundMethodResultInfluentialDistilling the Knowledge in a Neural Network
Hinton, G., Vinyals, O., & Dean, J. (2015) · arXiv · doi:10.48550/arxiv.1503.02531 · Cited by 14,204 · cited 1×
“The studies in the knowledge distillation framework (Hinton et al., 2015) use a smaller student-network that is trained to mimic the behavior of a larger teacher-network (BERT-large or BERT-base).”
BackgroundExploring 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 · cited 1×
“(2020) show the same for open-domain QA using T5 model (Raffel et al., 2019).”
BackgroundLewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., & Zettlemoyer, L. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.703 · Cited by 8,821
Research works printed but not linked: 76
- Arase Yuki, 2019, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 5393
- Baevski Alexei, 2019, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 5360
- He Bai, 2020, arXiv:2004. 14996 [cs]
- Bouraoui Zied, 2019, arXiv:1911.12753 [cs] · arXiv
- Brunner Gino, 2020, International Conference on Learning Representations
- Clark Kevin, 2020, International Conference on Learning Representations
- Conneau Alexis, 2019, arXiv:1911.02116 [cs] · arXiv
- Cui Yiming, 2019, arXiv:1906.08101 [cs] · arXiv
- Da Jeff, 2019, Proceedings of the First Workshop on Commonsense Inference in Natural Language Processing, 1
- Devlin Jacob, 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), 4171
- Elazar Yanai, 2020, arXiv:2006. 00995 [cs]
- Ettinger Allyson, 2019, arXiv: 1907.13528 [cs] · arXiv
- Fan Angela, 2019, International Conference on Learning Representations
- Forbes Maxwell, 2019, Proceedings of the 41st Annual Conference of the Cognitive Science Society (CogSci 2019), 7
- Frankle Jonathan, 2019, International Conference on Learning Representations
- Ganesh Prakhar, 2020, arXiv preprint arXiv:2002.11985 · arXiv
- Garg Siddhant, 2020, AAAI
- Gong Linyuan, 2019, International Conference on Machine Learning, 2337
- Gordon Mitchell A., 2020, arXiv preprint arXiv:2002.08307 · arXiv
- Goyal Saurabh, 2020, arXiv preprint arXiv:2001. 08950
- Hao Yaru, 2019, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 4143
- Hinton Geoffrey, 2014, Deep Learning and Representation Learning Workshop: NIPS 2014
- Hoover Benjamin, 2019, arXiv:1910. 05276 [cs]
- Jain Sarthak, 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), 3543
- Jawahar Ganesh, 2019, 57th Annual Meeting of the Association for Computational Linguistics (ACL), Florence, Italy
- Jiang Haoming, 2019, arXiv preprint arXiv:1911.03437 · arXiv
- Jiang Zhengbao, 2019, arXiv:1911. 12543 [cs]
- Di Jin, 2020, AAAI 2020
- Kim Taeuk, 2020, ICLR 2020
- Kong Lingpeng, 2019, International Conference on Learning Representations
- Kovaleva Olga, 2019, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 4356
- Krishna Kalpesh, 2020, ICLR 2020
- Kuznetsov Ilia, 2020, arXiv:2004.14999 [cs] · arXiv
- Lan Zhenzhong, 2020, ICLR
- Li Changmao, 2020, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5709
- Lin Yongjie, 2019, Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, 241
- Liu Nelson F., 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), 1073
- Manning Christopher D., 2020, Proceedings of the National Academy of Sciences, 201907367
- May Chandler, 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), 622
- Thomas McCoy R., 2019, International Conference on Learning Representations
- Michel Paul, 2019, Advances in Neural Information Processing Systems 32 (NIPS 2019)
- Mikolov Tomas, 2013, Advances in Neural Information Processing Systems 26 (NIPS 2013), 3111
- Jiaqi Mu, 2018, International Conference on Learning Representations
- Pimentel Tiago, 2020, arXiv:2004. 03061 [cs]
- Poerner Nina, 2019, arXiv preprint arXiv: 1911.03681 · arXiv
- Raganato Alessandro, 2020, arXiv:2002.10260 [cs] · arXiv
- Richardson Kyle, 2020, AAAI 2020
- Richardson Kyle, 2019, arXiv:1912.13337 [cs] · arXiv
- Roberts Adam, 2020, arXiv preprint arXiv:2002.08910 · arXiv
- Rogers Anna, 2020, AAAI, 11
- Sanh Victor, 2019, 5th Workshop on Energy Efficient Machine Learning and Cognitive Computing - NeurIPS 2019
- Serrano Sofia, 2019, arXiv:1906.03731 [cs] · arXiv
- Shen Sheng, 2019, arXiv preprint arXiv:1909. 05840
- Stickland Asa Cooper, 2019, International Conference on Machine Learning, 5986
- Strubell Emma, 2019, ACL 2019
- Sugawara Saku, 2020, AAAI
- Sun Siqi, 2019, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 4314
- Talmor Alon, 2019, arXiv:1912.13283 [cs] · arXiv
- Tenney Ian, 2019, International Conference on Learning Representations
- Schijndel Marten van, 2019, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 5831
- Vaswani Ashish, 2017, Advances in Neural Information Processing Systems, 5998
- Vig Jesse, 2019, Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, 63
- Vilares David, 2020, Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI-20)
- Voita Elena, 2019, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 4387
- Voita Elena, 2020, arXiv:2003.12298 [cs] · arXiv
- Wallace Eric, 2019, arXiv preprint arXiv:1909. 07940
- Warstadt Alex, 2020, Proceedings of the 42nd Annual Virtual Meeting of the Cognitive Science Society
- Warstadt Alex, 2019, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 2870
- Wolf Thomas, 2020, arXiv:1910.03771 [cs] · arXiv
- Felix Wu, 2019, International Conference on Learning Representations
- Zhiyong Wu, 2020, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 4166
- Yin Pengcheng, 2020, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 8413
- Zadeh Ali Hadi, 2020, arXiv:2005.03842 [cs, stat] · arXiv
- Zhang Zhuosheng, 2020, AAAI 2020
- Zhao Sanqiang, 2019, arXiv preprint arXiv:1909.11687 · arXiv
- Zhou Xuhui, 2020, AAAI 2020
Cited by
Items 1,001–1,014 of 1,014
Testing the Effectiveness of the Diagnostic Probing Paradigm on Italian Treebanks
Miaschi, A., Alzetta, C., Brunato⋄, D., Dell’Orletta⋄, F., & Venturi⋄, G. (2023) · Information · doi:10.3390/info14030144 · Cited by 6
Fine-tuning language models to recognize semantic relations
Roussinov, D., Sharoff, S., & Puchnina, N. (2023) · Language Resources and Evaluation · doi:10.1007/s10579-023-09677-w · Cited by 2
“While it has been shown that their weights can be used to predict some linguistic properties, such as semantic roles (Rogers et al., 2020) or word analogy (Vulic et al., 2020), our study is the first to show that the pre-trained models still have limitations, which are not always revealed by the…”
BackgroundGale, R. C., Salem, A. C., Fergadiotis, G., & Bedrick, S. D. (2023) · Workshop on Representation Learning for NLP (RepL4NLP) · doi:10.18653/v1/2023.repl4nlp-1.18 · Cited by 11
“Recently, large language models (LLMs) have shown notable success in capturing information across several linguistic layers, developing rich representations of syntactic and semantic structures within their hidden layers (Rogers et al., 2020).”
BackgroundHierarchical Representations in Dense Passage Retrieval for Question-Answering
Ennen, P., Freddi, F., Lin, C.-J., Kung, P.-N., Wang, R., Yang, C., Shiu, D.-S., & Bernacchia, A. (2023) · Fact Extraction and VERification Workshop (FEVER) · doi:10.18653/v1/2023.fever-1.2 · Cited by 3
“This is inspired by the observation that different layers provide different level of abstraction (Rogers et al., 2020) that can be all used for downstream tasks (Evci et al., 2022).”
BackgroundFault classification method for on-board equipment of metro train control system based on BERT-CNN
Qian, X. U., Zhang, L., Ou, D., & Yunpeng, H. E. (2023) · JOURNAL OF SHENZHEN UNIVERSITY SCIENCE AND ENGINEERING · doi:10.3724/sp.j.1249.2023.05529 · Cited by 2
Korichi, Y., Desharnais, J., Gambs, S., & Tawbi, N. (2024) · Lecture notes in computer science · doi:10.1007/978-3-031-70903-6_14 · Cited by 2
Suryono, M. E. P., & Djunaidy, A. (2024) · International Conference on Electronic and Electrical Engineering and Intelligent System (ICE3IS) · doi:10.1109/ice3is62977.2024.10775785 · Cited by 2
“BERT is a transformer-based model that learns and represents information in NLP [13].”
BERT For Classifying Intent in Ulos Fabric and Batak Culture Learning Chatbot
Atika, S., Lydia, M. S., & Muchtar, M. A. (2025) · International Conference on Computer Sciences, Engineering, and Technology Innovation (ICoCSETI) · doi:10.1109/icocseti63724.2025.11020457 · Cited by 2
Building a Discourse‐Argument Hybrid System for Vietnamese Why‐Question Answering
Nguyen, C. T., & Nguyen, D. T. (2021) · Computational Intelligence and Neuroscience · doi:10.1155/2021/6550871 · Cited by 2
Li, B., Wisniewski, G., & Crabbé, B. (2022) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/2022.acl-short.54 · Cited by 6
“Many works (Rogers et al., 2020) strive to analyze the representations uncovered by transformers to find out whether they are consistent with models derived from linguistic theories.”
BackgroundA Survey of Transformer-Based Stance Detection
Küçük, D. (2022) · Advances in computational intelligence and robotics book series · doi:10.4018/978-1-6684-6001-6.ch004 · Cited by 2
SURİYELİ MÜLTECİLERE UYGULANAN AYRIMCI-DIŞLAYICI TWİTLERİN BERT MODELİ İLE SINIFLANDIRILMASI
KARAMAN, E. (2022) · Ortadoğu ve Göç · doi:10.31834/ortadoguvegoc.1148440 · Cited by 4
Anatomy of a large-scale real-time peer evaluation system
Nyomi, X., & Moccozet, L. (2022) · International Conference on Information Technology Based Higher Education and Training (ITHET) · doi:10.1109/ithet56107.2022.10032005 · Cited by 3
Remy, F., Delobelle, P., Berendt, B., Demuynck, K., & Demeester, T. (2023) · arXiv · doi:10.48550/arxiv.2310.03477
“Encoding words as substring tokens is undoubtedly a trade-off, whose challenges are by now well-known (Provilkov et al., 2020; Rogers et al., 2020).”
Background
Metadata
- Authors
- Rogers, A., Kovaleva, O., & Rumshisky, A.
- Affiliations
- University of Copenhagen (DK)
- University of Massachusetts Lowell (US)
- Venue
- Transactions of the Association for Computational Linguistics
- Publisher
- MIT Press
- Published
- 2020-12
- Volume
- 8
- Pages
- 842-866
- ISSN
- 2307-387X (ISSN-L)
- OpenAlex
- W3006881356
- Cited by
- 1,552
- References
- 182
- Citations by year
- 2026: 171 · 2025: 249 · 2024: 183 · 2023: 292 · 2022: 237 · 2021: 320 · 2020: 100
- FWCI
- 93.20
- Type
- journal article
- Language
- en
- Open access
- diamond
- doi.org
- Licence
- cc-by
- Locations
- Transactions of the Association for Computational Linguistics (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- IT University Of Copenhagen (IT University of Copenhagen) (submittedVersion)
- arXiv (Cornell University)
- Funders
- National Science Foundation (1844740)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Keywords
- BERT, transformer models, pre-trained language models, model interpretability, training objectives, model architecture modification, overparameterization, model compression, learned representations
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.
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
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.