The DOI matches its DataCite record; 23 of the 23 references it cites were found in a registry or an index; 3 of the 4 sources its claims cite were checked for support.
DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter
Open accessAbstract
As Transfer Learning from large-scale pre-trained models becomes more prevalent in Natural Language Processing (NLP), operating these large models in on-the-edge and/or under constrained computational training or inference budgets remains challenging. In this work, we propose a method to pre-train a smaller general-purpose language representation model, called DistilBERT, which can then be fine-tuned with good performances on a wide range of tasks like its larger counterparts. While most prior work investigated the use of distillation for building task-specific models, we leverage knowledge distillation during the pre-training phase and show that it is possible to reduce the size of a BERT model by 40%, while retaining 97% of its language understanding capabilities and being 60% faster. To leverage the inductive biases learned by larger models during pre-training, we introduce a triple loss combining language modeling, distillation and cosine-distance losses. Our smaller, faster and lighter model is cheaper to pre-train and we demonstrate its capabilities for on-device computations in a proof-of-concept experiment and a comparative on-device study.
Overview
AI snapshot
Written from the paper’s full text.
TL;DR
DistilBERT is a distilled version of BERT pre-trained with knowledge distillation, reducing parameters by 40% while retaining 97% of BERT’s GLUE performance and running 60% faster at inference. It is presented as a general-purpose, lighter model suitable for constrained and on-device NLP applications.
Key findings
- DistilBERT has 66M parameters vs. BERT-base’s 110M and takes 410s vs. 668s for a CPU inference pass on STS-B, i.e. 40% smaller and 60% faster.
- On GLUE dev sets, DistilBERT scores 77.0 macro-score vs. BERT-base’s 79.5, retaining 97% of performance; it is always on par or improving over the ELMo baseline.
- On IMDb test accuracy, DistilBERT reaches 92.82 vs. BERT-base’s 93.46; on SQuAD v1.1, DistilBERT reaches 77.7/85.8 EM/F1 vs. BERT-base’s 81.2/88.5.
- Ablations show removing the cosine loss costs -2.96 GLUE macro-score, removing distillation loss costs -1.46, removing MLM loss costs -0.31, and random initialization costs -3.69.
Methods
- Knowledge distillation during pre-training: a student with BERT’s architecture but half the layers, no token-type embeddings and no pooler, initialized from every other teacher layer.
- Training objective is a triple loss combining distillation loss \(L_{ce}\), masked language modeling loss \(L_{mlm}\), and cosine embedding loss \(L_{cos}\), using softmax-temperature.
- Trained on English Wikipedia + Toronto Book Corpus with large batches up to 4K examples, dynamic masking, no next-sentence prediction, on 8 16GB V100 GPUs for approximately 90 hours; evaluated on GLUE, IMDb, SQuAD v1.1, CPU inference, and an iPhone 7 Plus on-device proof-of-concept.
Limitations
- The paper does not state an explicit limitations section.
- Evident from the text: evaluation is limited to English benchmarks and a small set of downstream tasks, and the on-device study is described as a proof-of-concept; the model could be further reduced with quantization.
- Evident from the text: the SQuAD second-distillation result is reported inconsistently—the prose says 79.8 F1 and 70.4 EM, while Table 2 lists DistilBERT (D) as 79.1/86.9.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
15 Own result · 6 Own interpretation · 2 Own method · 2 No basis given · 1 Supported · 1 Partly supported · 1 Sources not found · accuracy not yet measured
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“we leverage knowledge distillation during the pre-training phase and show that it is possible to reduce the size of a BERT model by 40%, while retaining 97% of its language understanding capabilities and being 60% faster.”
- Also stated in: § Size and inference speed · p. 4 · plain
“DistilBERT has 40% fewer parameters than BERT and is 60% faster than BERT.”
- Also stated in: § 6 Conclusion and future work · p. 5 · plain
“We introduced DistilBERT, a general-purpose pre-trained version of BERT, 40% smaller, 60% faster, that retains 97% of the language understanding capabilities.”
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.
- Also stated in: § Size and inference speed · p. 4 · plain
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“we propose a method to pre-train a smaller generalpurpose language representation model, called DistilBERT, which can then be finetuned with good performances on a wide range of tasks like its larger counterparts.”
- Also stated in: § 1 Introduction · p. 2 · plain
“Our general-purpose pre-trained models can be fine-tuned with good performances on several downstream tasks, keeping the flexibility of larger models.”
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.
- Also stated in: § 1 Introduction · p. 2 · plain
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Further ablation studies indicate that all the components of the triple loss are important for best performances.”
hedgeddirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Excluding the tokenization step, DistilBERT is 71% faster than BERT, and the whole model weighs 207 MB (which could be further reduced with quantization).”
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.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“we leverage knowledge distillation during the pre-training phase and show that it is possible to reduce the size of a BERT model by 40%, while retaining 97% of its language understanding capabilities and being 60% faster.”
- Also stated in: § Size and inference speed · p. 4 · plain
“DistilBERT has 40% fewer parameters than BERT and is 60% faster than BERT.”
- Also stated in: § 6 Conclusion and future work · p. 5 · plain
“We introduced DistilBERT, a general-purpose pre-trained version of BERT, 40% smaller, 60% faster, that retains 97% of the language understanding capabilities.”
plaindirect- 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: Inference speed measured on a single GLUE task (STS-B) on CPU is representative of DistilBERT's speed advantage generally.
- Also stated in: § Size and inference speed · p. 4 · plain
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“we propose a method to pre-train a smaller generalpurpose language representation model, called DistilBERT”
- Also stated in: § 6 Conclusion and future work · p. 5 · plain
“We introduced DistilBERT, a general-purpose pre-trained version of BERT”
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.
- Also stated in: § 6 Conclusion and future work · p. 5 · plain
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We found it beneficial to add a cosine embedding loss ( L cos ) which will tend to align the directions of the student and teacher hidden states vectors.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“On SQuAD, DistilBERT is within 3.9 points of the full BERT.”
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.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“we leverage knowledge distillation during the pre-training phase and show that it is possible to reduce the size of a BERT model by 40%, while retaining 97% of its language understanding capabilities and being 60% faster.”
- Also stated in: § 4 Experiments · p. 3 · plain
“DistilBERT also compares surprisingly well to BERT, retaining 97% of the performance with 40% fewer parameters.”
- Also stated in: § 6 Conclusion and future work · p. 5 · plain
“We introduced DistilBERT, a general-purpose pre-trained version of BERT, 40% smaller, 60% faster, that retains 97% of the language understanding capabilities.”
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.
- Also stated in: § 4 Experiments · p. 3 · plain
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“removing the Masked Language Modeling loss has little impact while the two distillation losses account for a large portion of the performance.”
plaindirect- 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.
- Probably not a substantial claim. 7
- It may state two separate claims as one. 2
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“environmental cost of exponentially scaling these models' computational requirements as mentioned in Schwartz et al. [2019], Strubell et al. [2019].”
What the sources say
- Supportedcites Schwartz et al., 2019 source
The Abstract (p. 1) says the computations required for deep learning research have been doubling every few months and have a large carbon footprint. Section 2, “Red AI” (p. 4), likewise warns that the environmental cost of increasingly large models will not decrease soon. (LLM, conf=0.93)
- Partly supportedcites Strubell et al., 2019 sourceExperimental — accuracy not yet measured
The Introduction (page unavailable) says that computationally hungry NLP models require considerable energy and have associated environmental costs. It also says model development can multiply those costs thousands of times, but it does not establish that the models’ computational requirements are scaling exponentially. (LLM, conf=0.84)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“We applied best practices for training BERT model recently proposed in Liu et al. [2019].”
What the sources say
- Supportedcites Liu et al., 2019 source
The Introduction (p. 1) says Liu et al. propose an improved recipe for training BERT, and Section 4 (p. 4) evaluates which choices matter for BERT pretraining. The citing authors’ statement that they applied those practices to DistilBERT describes their own work, not a finding attributed to Liu et al. (LLM, conf=0.92)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
- Supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Among the 9 tasks, DistilBERT is always on par or improving over the ELMo baseline (up to 19 points of accuracy on STS-B).”
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.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In this case, we were able to reach interesting performances given the size of the model: 79.8 F1 and 70.4 EM, i.e. within 3 points of the full model.”
plaindirect- 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.
It may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We further demonstrated that DistilBERT is a compelling option for edge applications.”
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: fails — It may be too vague to check against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. It may be too vague to check against evidence.
What the paper did — it cites no source for this, so there is none to check. The quote shows it is in the paper.
In the paper
“BERT and DistilBERT results are the medians of 5 runs with different seeds.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
In the paper
“current research 1 on pre-trained models indicates that training even larger models still leads to better performances on downstream tasks.”
What the sources say
- Not matchedcites MegatronLM
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
hedgeddirect- 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.
Probably not a substantial claim.
- Not matched
What the paper did — it cites no source for this, so there is none to check. The quote shows it is in the paper.
In the paper
“We report scores on the development sets for each task by fine-tuning DistilBERT without the use of ensembling or multi-tasking scheme for fine-tuning (which are mostly orthogonal to the present work).”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“While most prior work investigated the use of distillation for building task-specific models, we leverage knowledge distillation during the pre-training phase”
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: fails — The paper may give no reason why it holds.
Probably not a substantial claim. The paper may give no reason why it holds.
The paper relies on this without support.
Operating large pre-trained models on-device in real-time is hampered by their growing computational and memory requirements.
“the growing computational and memory requirements of these models may hamper wide adoption.”
The paper relies on this without support.
Near-zero probabilities in a trained model's output distribution reflect its generalization capabilities.
“some of these”
The paper relies on this without support.
Inference speed measured on a single GLUE task (STS-B) on CPU is representative of DistilBERT's speed advantage generally.
“Inference time of a full pass of GLUE task STS-B (sentiment analysis) on CPU with a batch size of 1.”
“DistilBERT has 40% fewer parameters than BERT and is 60% faster than BERT.”
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
“This idea has been successfully applied to LLMs, where distilled students approximate the generation patterns of larger teachers while retaining much of their performance (Sanh et al., 2019).”
From Vision-Guided Iterative Refinement for Frontend Code Generation (2026) · cites this paper as Sanh et al., 2019
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
“even benchmarking is not a reliable estimator of parameter count [12], [15], [28], [29]”
From Hollow-LLM Attack: Computationally Trivial Weights in Zero-Knowledge Verification of LLM Inference (2026) · cites this paper as [29]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“It achieves similar performance to the original BERT model while using fewer resources. The approach leveraged knowledge distillation during the pretraining phase and reduced the size of general BERT by 40%, yet it retained a language understanding capability of 97%.”
From A Review on Deep-Learning-Based Cyberbullying Detection (2023) · cites this paper as [53]
The cited source says what the claim attributes to it.
In the citing paper
“Unlike in many similar circumstances, we do not employ knowledge distillation [56], because we discovered that for our sizes of the models, this technique does not provide any quality improvement.”
From Full Line Code Completion: Bringing AI to Desktop (2024) · cites this paper as [56]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ 4.1 Downstream task benchmark · p. 3
“In this case, we were able to reach interesting performances given the size of the model: 79.8 F1 and 70.4 EM, i.e. within 3 points of the full model.”
§ 3 DistilBERT: a distilled version of BERT · p. 3
“| DistilBERT (D) | - | 79.1/86.9 |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
“DistilBERT is always on par or improving over the ELMo baseline (up to 19 points of accuracy on STS-B).”
Reported 19 · recomputed 16.5
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
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
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
Distilling the Knowledge in a Neural Network
Hinton, G., Vinyals, O., & Dean, J. (2015) · arXiv · doi:10.48550/arxiv.1503.02531 · Cited by 14,204
Deep 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
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
Self-citationSQuAD: 100,000+ Questions for Machine Comprehension of Text
Rajpurkar, P., Zhang, J., Lopyrev, K., & Liang, P. (2016) · Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d16-1264 · Cited by 6,208
GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding
Wang, A., Singh, A., Michael, J., Hill, F., Levy, O., & Bowman, S. (2018) · Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP · doi:10.18653/v1/w18-5446 · Cited by 4,222
Learning Word Vectors for Sentiment Analysis
Maas, A. L., Daly, R. E., Pham, P. T., Huang, D., Ng, A. Y., & Potts, C. (2011) · Cited by 3,710
Energy and Policy Considerations for Deep Learning in NLP
Strubell, E., Ganesh, A., & McCallum, A. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1355 · Cited by 3,684
Research works printed but not linked: 1
- Cristian Bucila, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In KDD , 2006.
Cited by
Sorted by citations
Items 2,901–3,000 of 4,692
Fain, M., Twomey, N., & Bollegala, D. (2021) · International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) · doi:10.1145/3404835.3463027 · Cited by 3
Leveraging knowledge sources for detecting self-reports of particular health issues on social media
Bagherzadeh, P., & Bergler, S. (2021) · Cited by 3
Kotyushev, M., Glazkova, A. V., & Morozov, D. A. (2021) · International Workshop on Semantic Evaluation (SemEval) · doi:10.18653/v1/2021.semeval-1.124 · Cited by 3
The Effect of Using Masked Language Models in Random Textual Data Augmentation
Rashid, M. A., & Amirkhani, H. (2021) · International Computer Conference, Computer Society of Iran (CSICC) · doi:10.1109/csicc52343.2021.9420616 · Cited by 2
WikiMorph: Learning to Decompose Words into Morphological Structures
Yarbro, J. T., & Olney, A. M. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-78270-2_72 · Cited by 3
Pimpalkhute, V., Nakhate, P., & Diwan, T. (2021) · Social Media Mining for Health Research and Applications Workshop and Shared Tasks (SMM4H) · doi:10.18653/v1/2021.smm4h-1.24 · Cited by 4
SeqDialN: Sequential Visual Dialog Network in Joint Visual-Linguistic Representation Space
Yang, L., Meng, F., Liu, X., Wu, M.-K. D., Ying, V., & Xu, J. Y. (2021) · doi:10.18653/v1/2021.dialdoc-1.2 · Cited by 3
QA4GIS: A novel approach learning to answer GIS developer questions with API documentation
Wang, W., Li, Y., Wang, S., & Ye, X. (2021) · Transactions in GIS · doi:10.1111/tgis.12798 · Cited by 4
Gombert, S., & Bartsch, S. (2021) · International Workshop on Semantic Evaluation (SemEval) · doi:10.18653/v1/2021.semeval-1.12 · Cited by 3
Efficient Deep Feature Calibration for Cross-Modal Joint Embedding Learning
Xie, Z., Liu, L., Li, L., & Zhong, L. (2021) · INTERNATIONAL CONFERENCE ON MULTIMODAL INTERACTION · doi:10.1145/3462244.3479892 · Cited by 3
Veri Artırımı için Yarı-Denetimli Bağlamsal Anlam Belirsizliği Giderme
Torunoğlu-Selamet, D., & Eryiğit, G. (2021) · Türkiye Bilişim Vakfı Bilgisayar Bilimleri ve Mühendisliği Dergisi · doi:10.54525/tbbmd.835744 · Cited by 3
A Compression-Compilation Framework for On-mobile Real-time BERT Applications
Niu, W., Kong, Z., Yuan, G., Jiang, W., Guan, J., Ding, C., Zhao, P., Liu, S., Ren, B., & Wang, Y. (2021) · International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2021/712 · Cited by 7
Li, X., Rong, Y., Zhao, M., & Fan, J. (2021) · Communications in computer and information science · doi:10.1007/978-981-16-5188-5_41 · Cited by 3
Knowledge Distillation with Noisy Labels for Natural Language Understanding
Bhardwaj, S., Ghaddar, A., Rashid, A., Bibi, K., Li, C., Ghodsi, A., Langlais, P., & Rezagholizadeh, M. (2021) · Workshop on Noisy User-generated Text (W-NUT) · doi:10.18653/v1/2021.wnut-1.33 · Cited by 3
Adversarial Variational Knowledge Distillation
Tang, X. L., & Lin, T. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-86365-4_45 · Cited by 3
Compressibility of Distributed Document Representations
Škrlj, B., & Petković, M. (2021) · Proceedings · doi:10.1109/icdm51629.2021.00166 · Cited by 3
Predicting Doctor Ratings from User Reviews using Deep Learning
Jhaveri, Y., Gandhi, T., Naik, T., Nisar, S., & Sonawane, P. (2022) · International Conference on Applied Artificial Intelligence and Computing (ICAAIC) · doi:10.1109/icaaic53929.2022.9793206 · Cited by 3
Assessment of Massively Multilingual Sentiment Classifiers
Rajda, K., Augustyniak, L., Gramacki, P., Gruza, M., Woźniak, S., & Kajdanowicz, T. J. (2022) · Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis (WASSA) · doi:10.18653/v1/2022.wassa-1.13 · Cited by 5
Dawkins, H., & Nejadgholi, I. (2022) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/2022.acl-short.59 · Cited by 3
Embarrassingly Simple Performance Prediction for Abductive Natural Language Inference
Kadiķis, E., Srivastav, V., & Klinger, R. (2022) · Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT) · doi:10.18653/v1/2022.naacl-main.441 · Cited by 4
BE3R: BERT based Early-Exit Using Expert Routing
Mangrulkar, S., Ankith, M. S., & Sembium, V. (2022) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3534678.3539132 · Cited by 3
Wu, F., He, S., Guo, S., Qu, Z., Wang, H., Zhuang, W., & Zhang, J. (2022) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3489517.3530417 · Cited by 4
Bhagya, S. S., & Nadera, B. S. (2022) · International Conference on Computing, Communication, Security and Intelligent Systems (IC3SIS) · doi:10.1109/ic3sis54991.2022.9885285 · Cited by 4
Yi̇rmi̇beşoğlu, Z., & Güngör, T. (2022) · ACM Transactions on Asian and Low-Resource Language Information Processing · doi:10.1145/3571073 · Cited by 4
Improving Healthcare Question Answering System by Identifying Suitable Answers
Das, B., & Nirmala, S. J. (2022) · IEEE Mysore Sub Section International Conference (MysuruCon) · doi:10.1109/mysurucon55714.2022.9972435 · Cited by 3
An Implemented review for Intent Creation using different Clustering Techniques
Magoo, C., & Singh, M. (2022) · International Conference on Computational Intelligence and Communication Technologies (CCICT) · doi:10.1109/ccict56684.2022.00027 · Cited by 3
Parameter-efficient feature-based transfer for paraphrase identification
Liu, X., Rzepka, R., & Araki, K. (2022) · Natural Language Engineering · doi:10.1017/s135132492200050x · Cited by 3
Béchara, H., Zhang, R., Yuan, S., & Mikhaylov, S. (2022) · IEEE International Conference on Big Data (BigData) · doi:10.1109/bigdata55660.2022.10020787 · Cited by 6
Caron, M. (2022) · IEEE International Conference on Big Data (BigData) · doi:10.1109/bigdata55660.2022.10020933 · Cited by 3
Improvement of Semantic-BERT Model with Bidirectional LSTM for Arabic Language
Mahdi, A. F., Farhan, R. N., & Al-Rawi, S. S. (2022) · International Conference on Contemporary Information Technology and Mathematics (ICCITM) · doi:10.1109/iccitm56309.2022.10031836 · Cited by 3
CpG Island Detection Using Transformer Model with Conditional Random Field
Hossain, M. J., Bhuiyan, M. I. H., & Abdullah, Z. R. (2022) · IEEE Bombay Section Signature Conference (IBSSC) · doi:10.1109/ibssc56953.2022.10037492 · Cited by 3
Natural language processing: fast forwarding research to the “good stuff”
Lindquist, J. M. (2023) · American Journal of Clinical Nutrition · doi:10.1016/j.ajcnut.2022.12.012 · Cited by 3
Tweet Summarization Using Clustering Mechanisms
Ampili, K., & Kanakala, S. (2022) · International Conference on Advances in Computing, Communication Control and Networking (ICAC3N) · doi:10.1109/icac3n56670.2022.10074514 · Cited by 3
Light RAT-SQL: A RAT-SQL with More Abstraction and Less Embedding of Pre-existing Relations
Ndongala, N. M. (2023) · Texila international journal of academic research · doi:10.21522/tijar.2014.10.02.art001 · Cited by 4
Kharazmi, P., Zhao, Z., Chung, C., & Choudhary, S. (2023) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp49357.2023.10095536 · Cited by 3
Fast and highly coupled model for time series forecasting
Guo, T., Wang, H., Zhang, M., Liu, Y., & Zhang, F. (2023) · Multimedia Tools and Applications · doi:10.1007/s11042-023-15787-y · Cited by 3
Predicting Depression Risk in Patients with Cancer Using Multimodal Data
de Hond, A., van Buchem, M. M., Fanconi, C., Roy, M., Blayney, D. W., Kant, I. M. J., Steyerberg, E. W., & Hernandez‐Boussard, T. (2023) · Studies in health technology and informatics · doi:10.3233/shti230274 · Cited by 3
An NLP-Empowered Virtual Course Assistant for Online Teaching and Learning
Liu, S., Man, S., & Song, L. (2022) · IEEE International Conference on Teaching, Assessment, and Learning for Engineering · doi:10.1109/tale54877.2022.00068 · Cited by 3
Fake News Detection in Tweets: Challenges and Adaptations imposed by the COVID-19
Santana, C., Claro, D. B., & Souza, M. (2022) · iSys - Brazilian Journal of Information Systems · doi:10.5753/isys.2022.2286 · Cited by 3
On the Expressive Flexibility of Self-Attention Matrices
Likhosherstov, V., Choromański, K., & Weller, A. (2023) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v37i7.26055 · Cited by 5
EMS®: A Massive Computational Experiment Management System towards Data-driven Robotics
Lin, Q., Guo, Y., & Liu, H. (2023) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48891.2023.10160948 · Cited by 3
Tahsin, M. U., Shanto, M. S. H., & Rahman, R. M. (2023) · SN Computer Science · doi:10.1007/s42979-023-01989-6 · Cited by 3
A Robust Log Classification Approach Based on Natural Language Processing
Li, D., Zhang, J., Yang, J., Zhang, X., Lin, F., Wang, C., & Chang, L. (2023) · International Conference on Computer, Control and Robotics (ICCCR) · doi:10.1109/icccr56747.2023.10194171 · Cited by 3
A Full-fledged Commit Message Quality Checker Based on Machine Learning
Faragó, D., Färber, M., & Petrov, C. (2023) · IEEE Annual International Computer Software and Applications Conference (COMPSAC) · doi:10.1109/compsac57700.2023.00108 · Cited by 4
Privacy-Preserving Knowledge Transfer through Partial Parameter Sharing
Youssef, P., Schlötterer, J., & Seifert, C. (2023) · Clinical Natural Language Processing Workshop (ClinicalNLP) · doi:10.18653/v1/2023.clinicalnlp-1.3 · Cited by 3
When Gradient Descent Meets Derivative-Free Optimization: A Match Made in Black-Box Scenario
Han, C., Cui, L., Zhu, R., Wang, J., Chen, N., Sun, Q., Li, X., & Gao, M. (2023) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2023.findings-acl.55 · Cited by 5
Shmalts, M. (2023) · International Workshop on Semantic Evaluation (SemEval) · doi:10.18653/v1/2023.semeval-1.289 · Cited by 3
Guerin, N., & Chemla, E. (2023) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2023.findings-acl.560 · Cited by 3
Language Detoxification with Attribute-Discriminative Latent Space
Kwak, J. M., Kim, M., & Hwang, S. J. (2023) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/2023.acl-long.565 · Cited by 9
Dynamic and Efficient Inference for Text Generation via BERT Family
Liang, X., Li, J., Wu, L., Cao, Z., & Zhang, M. (2023) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/2023.acl-long.162 · Cited by 3
Campos, D., Magnani, A., & Zhai, C. (2023) · Workshop on Simple and Efficient Natural Language Processing (SustaiNLP) · doi:10.18653/v1/2023.sustainlp-1.4 · Cited by 8
Automatic Named Entity Obfuscation in Speech
Preiss, J. (2023) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2023.findings-acl.39 · Cited by 3
Matching Pairs: Attributing Fine-Tuned Models to their Pre-Trained Large Language Models
Foley, M., Rawat, A., Lee, T., Hou, Y., Picco, G., & Zizzo, G. (2023) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/2023.acl-long.410 · Cited by 8
Effective and Efficient Query-aware Snippet Extraction for Web Search
Yi, J., Wu, F., Wu, C., Huang, X., Jiao, B., Sun, G., & Xie, X. (2022) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2022.emnlp-main.197 · Cited by 4
Shen, B., Lin, Z., Liu, Y., Liu, Z., Wang, L., & Wang, W. (2022) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2022.emnlp-main.112 · Cited by 4
Knowledge Distillation Transfer Sets and their Impact on Downstream NLU Tasks
Peris, C., Tan, L., Gueudré, T., Gojayev, T., Wei, P., & Oz, G. (2022) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2022.emnlp-industry.12 · Cited by 4
Evidence > Intuition: Transferability Estimation for Encoder Selection
Bassignana, E., Müller-Eberstein, M., Zhang, M., & Plank, B. (2022) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2022.emnlp-main.283 · Cited by 7
WordTies: Measuring Word Associations in Language Models via Constrained Sampling
Yao, P., Renwick, T., & Barbosa, D. (2022) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2022.findings-emnlp.440 · Cited by 3
DomainNet: Homograph Detection and Understanding in Data Lake Disambiguation
Leventidis, A. S., Rocco, L. D., Gatterbauer, W., Miller, R. J., & Riedewald, M. (2023) · ACM Transactions on Database Systems · doi:10.1145/3612919 · Cited by 17
Hasan, M. R., Hossain, M. Z., Gedeon, T., Soon, S., & Rahman, S. (2023) · Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis (WASSA) · doi:10.18653/v1/2023.wassa-1.47 · Cited by 6
Barrière, V., Sedoc, J., Tafreshi, S., & Giorgi, S. (2023) · Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis (WASSA) · doi:10.18653/v1/2023.wassa-1.44 · Cited by 5
Lorandi, M., & Belz, A. (2023) · Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis (WASSA) · doi:10.18653/v1/2023.wassa-1.30 · Cited by 3
Efficient NLP Model Finetuning via Multistage Data Filtering
Ouyang, X., Ansari, S. M. A., Lin, F. X., & Ji, Y. (2023) · International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2023/455 · Cited by 4
AdaDS: Adaptive data selection for accelerating pre-trained language model knowledge distillation
Zhou, Q., Li, P., Liu, Y., Guan, Y., Xing, Q., Chen, M., Sun, M., & Liu, Y. (2023) · AI Open · doi:10.1016/j.aiopen.2023.08.005 · Cited by 3
Applying the Diamond Model of Intrusion Analysis with Generative Pre-trained Transformer 3
Chen, S.-S., Pai, T., & Sun, C. (2023) · IEEE International Conference on Consumer Electronics - Taiwan (ICCE-Taiwan) · doi:10.1109/icce-taiwan58799.2023.10226923 · Cited by 4
The Functional Relevance of Probed Information: A Case Study
Hanna, M., Zamparelli, R., & Mareček, D. (2023) · Conference of the European Chapter of the Association for Computational Linguistics · doi:10.18653/v1/2023.eacl-main.58 · Cited by 3
Dataset Construction and Opinion Holder Detection Using Pre-trained Models
Mahmud, A., & Shimada, K. (2023) · International Journal of Service and Knowledge Management · doi:10.52731/ijskm.v7.i2.779 · Cited by 4
Similarity versus Supervision: Best Approaches for HS Code Prediction
Stassin, S., Amel, O., Mahmoudi, S. A., & Siebert, X. (2023) · European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN) · doi:10.14428/esann/2023.es2023-163 · Cited by 4
Text classification by CEFR levels using machine learning methods and BERT language model
Lagutina, N. S., Lagutina, K. V., Brederman, A. M., & Николаевна, К. Н. (2023) · Modeling and Analysis of Information Systems · doi:10.18255/1818-1015-2023-3-202-213 · Cited by 4
Text embeddings and clustering for characterizing online communities on Reddit
Sawicki, J. (2023) · Annals of Computer Science and Information Systems · doi:10.15439/2023f6275 · Cited by 4
Efficient Detection of Noise Reviews Over a Large Number of Places
Kim, H. G., & Park, Y. H. (2023) · IEEE Access · doi:10.1109/access.2023.3324654 · Cited by 3
Erdoğanyılmaz, C., Mengünoğul, B., & Balci, M. (2023) · International Conference on Computer Science and Engineering (UBMK) · doi:10.1109/ubmk59864.2023.10286653 · Cited by 6
Mohammadi, M., Eftekhari, M., & Hassani, A. (2023) · IEEE Access · doi:10.1109/access.2023.3327339 · Cited by 5
Izquierdo-Domenech, J., Linares-Pellicer, J., & Ferri-Molla, I. (2023) · Multimedia Tools and Applications · doi:10.1007/s11042-023-17516-x · Cited by 4
MOOCMaven: Bridging M-Learning and Open Data for Enhanced Unified MOOC Exploration
Tatang, A. F., & Algarni, A. (2023) · International Journal of Interactive Mobile Technologies (iJIM) · doi:10.3991/ijim.v17i20.42445 · Cited by 3
CSformer: Enhancing deep learning efficiency for intelligent IoT
Jia, X., Wu, H., Zhang, R., & Peng, M. (2023) · Computer Communications · doi:10.1016/j.comcom.2023.11.007 · Cited by 3
Singh, A., & Kumar, S., V (2023) · International Conference on Computing, Communication and Networking Technologies (ICCCNT) · doi:10.1109/icccnt56998.2023.10306507 · Cited by 3
DisCo: Distilled Student Models Co-training for Semi-supervised Text Mining
Jiang, W., Mao, Q., Lin, C., Li, J., Deng, T., Yang, W., & Wang, Z. (2023) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2023.emnlp-main.244 · Cited by 4
Linking Surface Facts to Large-Scale Knowledge Graphs
Radevski, G., Gashteovski, K., Hung, C.-C., Lawrence, C., & Glavaš, G. (2023) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2023.emnlp-main.445 · Cited by 4
Confidence-based Ensembling of Perspective-aware Models
Casola, S., Lo, S. M., Basile, V., Frenda, S., Cignarella, A. T., Patti, V., & Bosco, C. (2023) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2023.emnlp-main.212 · Cited by 4
DUnE: Dataset for Unified Editing
Akyürek, A. F., Pan, E., Kuwanto, G., & Wijaya, D. T. (2023) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2023.emnlp-main.114 · Cited by 14
INGENIOUS: Using Informative Data Subsets for Efficient Pre-Training of Language Models
Renduchintala, H. S. V. N. S. K., Killamsetty, K., Bhatia, S., Aggarwal, M., Ramakrishnan, G., Iyer, R., & Krishnamurthy, B. (2023) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2023.findings-emnlp.445 · Cited by 4
Ko, J., Park, S., Kim, Y., Ahn, S., Chang, D.-S., Ahn, E., & Yun, S.-Y. (2023) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2023.findings-emnlp.404 · Cited by 10
End-to-end Adversarial Sample Generation for Data Augmentation
Liu, T., & Sun, Y. (2023) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2023.findings-emnlp.760 · Cited by 4
SHARCS: Efficient Transformers Through Routing with Dynamic Width Sub-networks
Salehi, M., Mehta, S., Kusupati, A., Farhadi, A., & Hajishirzi, H. (2023) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2023.findings-emnlp.706 · Cited by 5
EDeR: Towards Understanding Dependency Relations Between Events
Li, R., Haslum, P., & Cui, L. (2023) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2023.emnlp-main.926 · Cited by 3
Learn From One Specialized Sub-Teacher: One-to-One Mapping for Feature-Based Knowledge Distillation
Saadi, K., Mitrović, J., & Granitzer, M. (2023) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2023.findings-emnlp.882 · Cited by 2
Modeling Easiness for Training Transformers with Curriculum Learning
Ranaldi, L., Pucci, G., & Zanzotto, F. M. (2023) · International conference Recent advances in natural language processing · doi:10.26615/978-954-452-092-2_101 · Cited by 3
Peruski, R., Aykac, D., Torkelson, L., et al. (2024) · Electronic Imaging · doi:10.2352/ei.2024.36.17.avm-112 · Cited by 3
Vairamani, A. D., & Nayyar, A. (2024) · Elsevier eBooks · doi:10.1016/b978-0-323-95315-3.00003-6 · Cited by 5
Identifying going concern issues in auditor opinions: link to bankruptcy events
Bougiatiotis, K., Zavitsanos, E., & Παλιούρας, Γ. (2023) · IEEE International Conference on Big Data (BigData) · doi:10.1109/bigdata59044.2023.10386452 · Cited by 3
Sanjana, S., Kuranagatti, S., Devisetti, J. G., Sharma, R., & Arya, A. (2023) · International Conference on Contemporary Computing and Informatics (IC3I) · doi:10.1109/ic3i59117.2023.10398034 · Cited by 4
Benaissa, A. R., Harbaoui, A., & Ghezala, H. H. B. (2024) · Behaviour and Information Technology · doi:10.1080/0144929x.2024.2313142 · Cited by 4
Deep Metric Learning: Loss Functions Comparison
Vasilev, R. L., & D’yakonov, A. G. (2023) · Doklady Mathematics · doi:10.1134/s1064562423701053 · Cited by 3
Quan, Z., Vong, C., Zeng, W., & Yang, W. (2024) · IEEE/ACM Transactions on Audio Speech and Language Processing · doi:10.1109/taslp.2024.3364610 · Cited by 3
Dartiko, F., Yusa, M., Erlansari, A., & Basha, S. A. (2024) · INTENSIF Jurnal Ilmiah Penelitian dan Penerapan Teknologi Sistem Informasi · doi:10.29407/intensif.v8i1.21971 · Cited by 4
Textual Pre-Trained Models for Age Screening Across Community Question-Answering
Figueroa, A., & Timilsina, M. (2024) · IEEE Access · doi:10.1109/access.2024.3368929 · Cited by 3
Ngamwal, S., Yogita, Y., & Pal, V. (2024) · Expert Systems with Applications · doi:10.1016/j.eswa.2024.123572 · Cited by 3
Hasan, M. M., & Nasrabadi, N. M. (2023) · IEEE International Joint Conference on Biometrics (IJCB) · doi:10.1109/ijcb57857.2023.10448749 · Cited by 4
Poncelet, J., & hamme, H. V. (2024) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp48485.2024.10446344 · Cited by 3
Metadata
- Authors
- Sanh, V., Debut, L., Chaumond, J., & Wolf, T.
- Venue
- arXiv
- Publisher
- arXiv
- Published
- 2019
- ISSN
- 2331-8422 (ISSN-L)
- arXiv
- 1910.01108
- OpenAlex
- W2978017171
- Cited by
- 4,653
- References
- 18
- Citations by year
- 2026: 67 · 2025: 443 · 2024: 801 · 2023: 1136 · 2022: 736 · 2021: 968 · 2020: 486 · 2019: 15
- Type
- posted content
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Transactions on Graph Data and Knowledge (cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Keywords
- DistilBERT, knowledge distillation, BERT, pre-training, model compression, on-device computation, language modeling, transfer learning
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.
A language model drew the claims from the paper on 2026-10-23, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-23 (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
