What BERT Is Not: Lessons from a New Suite of Psycholinguistic Diagnostics for Language Models
Open accessThe DOI matches its Crossref record; 36 of the 38 references it cites were found in a registry or an index.
Abstract
Pre-training by language modeling has become a popular and successful approach to NLP tasks, but we have yet to understand exactly what linguistic capacities these pre-training processes confer upon models. In this paper we introduce a suite of diagnostics drawn from human language experiments, which allow us to ask targeted questions about information used by language models for generating predictions in context. As a case study, we apply these diagnostics to the popular BERT model, finding that it can generally distinguish good from bad completions involving shared category or role reversal, albeit with less sensitivity than humans, and it robustly retrieves noun hypernyms, but it struggles with challenging inference and role-based event prediction— and, in particular, it shows clear insensitivity to the contextual impacts of negation.
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 source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“The second trend is that neural networks routinely fail on exactly the phenomena where such symbolic models succeed, for example, logical inference [6,9], compositionality [7,23] and out-of-distribution generalization [24–26].”
From Symbols and grounding in large language models (2023) · cites this paper as [6]
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
“Besides factual knowledge, large-scale pre-training also allows LMs to handle linguistic phenomena such as analogies [9], negations [25], semantic role sensitivity [25], semantic similarity [131], cant understanding [131], and rare word understanding [116].”
From Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing (2023) · cites this paper as [25]
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
“For these kinds of tasks, we can also score the generation probability of each candidate and choose the one with the highest probability [25].”
From Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing (2023) · cites this paper as [25]
Experimental — accuracy not yet measured
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 2×
“BERT is a deep bidirectional transformer network (Vaswani et al., 2017) pre-trained on tasks of masked language modeling (predicting masked words given bidirectional context) and next-sentence prediction (binary classification of whether two sentences are a sequence).”
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 2×
“As a case study, we use these three diagnostics to examine the predictive capacities of the pre-trained BERT model (Devlin et al., 2019), which has been the basis of impressive performance across a wide range of tasks.”
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 · cited 3×
“, 2019) and ELMo (Peters et al., 2018a) have advanced the state of the art in a wide variety of tasks, suggesting that these models acquire valuable, generalizable linguistic competence during the pre-training process.”
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 · cited 2×
“Much work has attempted to evaluate systems’ overall level of ‘‘understanding’’, often with tasks such as semantic similarity and entailment (Wang et al., 2018; Bowman et al., 2015; Agirre et al., 2012; Dagan et al., 2005; Bentivogli et al., 2016), and additional work has been done to design curated versions of these…”
A large annotated corpus for learning natural language inference
Bowman, S. R., Angeli, G., Potts, C., & Manning, C. D. (2015) · Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d15-1075 · Cited by 3,377
Brain potentials during reading reflect word expectancy and semantic association
Kutas, M., & Hillyard, S. A. (1984) · Nature · doi:10.1038/307161a0 · Cited by 2,218 · cited 1×
“Like cloze, the N400 can be used to gauge how expected a word w is in a context c —the amplitude of the N400 response appears to be sensitive to fit of a word in context, and has been shown to correlate with cloze in many cases (Kutas and Hillyard, 1984).”
MethodWhat Does BERT Look at? An Analysis of BERT’s Attention
Clark, K., Khandelwal, U., Levy, O., & Manning, C. D. (2019) · Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP · doi:10.18653/v1/w19-4828 · Cited by 1,373
BERT Rediscovers the Classical NLP Pipeline
Tenney, I., Das, D., & Pavlick, E. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1452 · Cited by 1,307
SemEval-2012 Task 6: A Pilot on Semantic Textual Similarity
Agirre, E., Cer, D. M., Diab, M., & González-Agirre, A. (2012) · Joint Conference on Lexical and Computational Semantics · Cited by 1,063
Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference
McCoy, R. T., Pavlick, E., & Linzen, T. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1334 · Cited by 944
Research works printed but not linked: 19
- Adi Yossi, 2016, International Conference on Learning Representations
- Agirre Eneko, 2012, Proceedings of the First Joint Conference on Lexical and Computational Semantics-Volume 1: Proceedings of the main conference and the shared task, and Volume 2: Proceedings of the Sixth International Workshop on Semantic Evaluation, 385
- Chowdhury Shammur Absar, 2018, Proceedings of the 27th International Conference on Computational Linguistics, 133
- Clark Kevin, 2019, arXiv preprint arXiv:1906.04341 · arXiv
- Conneau Alexis, 2018, ACL 2018-56th Annual Meeting of the Association for Computational Linguistics, 2126
- Dagan Ido, 2005, Machine Learning Challenges Workshop, 177
- Dasgupta Ishita, 2018, Proceedings of the 40th Annual Meeting of the Cognitive Science Society
- Devlin Jacob, 2019, Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies
- Ettinger Allyson, 2018, Proceedings of the 27th International Conference on Computational Linguistics, 1790
- Frank Stefan L., 2013, ACL 2013-51st Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference, 2, 878
- Futrell Richard, 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), 32
- Giulianelli Mario, 2018, Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, 240, 10.18653/v1/W18-5426 · DOI
- Gulordava Kristina, 2018, Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), 1, 1195
- Kim Najoung, 2019, Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM 2019), 235
- Peters Matthew, 2018, Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), 1, 2227
- Tenney Ian, 2019, arXiv preprint arXiv:1905.05950 · arXiv
- Tenney Ian, 2019, Proceedings of the International Conference on Learning Representations
- Vaswani Ashish, 2017, Advances in Neural Information Processing Systems, 5998
- Wang Alex, 2018, Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, 353
Cited by
Items 201–300 of 553
Semantics of Multiword Expressions in Transformer-Based Models: A Survey
Miletić, F., & Walde, S. S. I. (2024) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00657 · Cited by 31
“Detailed evidence of this trend comes from evaluations of phrase similarity by Yu and Ettinger (2020).”
BackgroundMethodInfluentialThe Non-native Speaker Aspect: Indian English in Social Media
Sarkar, R., Mahinder, S., & KhudaBukhsh, A. R. (2020) · Workshop on Noisy User-generated Text (W-NUT) · doi:10.18653/v1/2020.wnut-1.9 · Cited by 14
“Word prediction as a test of LM’s language understanding has been explored in Paperno et al. (2016); Ettinger (2020) and recent studies lever-aged it in novel applications such as relation extraction (Petroni et al., 2019) and political insight mining (Palakodety et al., 2020b).”
BackgroundZeng, C., & Li, S. (2021) · Wireless Communications and Mobile Computing · doi:10.1155/2021/5375334 · Cited by 15
“With the MLM probing study, Ettinger applied a set of diagnostic methods derived from human language experiments to the BERT model and found that BERT has a certain understanding of semantic roles [19, 22].”
MethodRetractedThe Role of Complex NLP in Transformers for Text Ranking
Rau, D., & Kamps, J. (2022) · ACM SIGIR International Conference on Theory of Information Retrieval (ICTIR) · doi:10.1145/3539813.3545144 · Cited by 14
“In previous work, Ettinger [7] showed that BERT relies on the word order during pre-training for the MLM task, finding evidence of the model encoding syntactic information in its representations.”
MethodAccurate, yet inconsistent? Consistency Analysis on Language Understanding ModelsPreprint
Jang, M., Kwon, D. S., & Lukasiewicz, T. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2108.06665 · Cited by 19
“…adversarial attacks exhibit a critical difference to our work in that they do not guarantee semantic identicalness due to the unnatural looking, grammatical errors, and inse-curity of PLM’s MLM performance (Ravichander et al., 2020; Ettinger, 2020; Kassner and Schütze, 2020; Elazar et al., 2021).”
BackgroundThe pragmatic profile of ChatGPT: assessing the communicative skills of a conversational agentPreprint
Pietro, C. B. D. S., Frau, F., Mangiaterra, V., & Bambini, V. (2023) · PsyArXiv (OSF Preprints) · doi:10.31234/osf.io/ckghw · Cited by 15
NOPE: A Corpus of Naturally-Occurring Presuppositions in English
Parrish, A., Schuster, S., Warstadt, A., Agha, O., Lee, S.-H., Zhao, Z., Bowman, S. R., & Linzen, T. (2021) · Conference on Computational Natural Language Learning · doi:10.18653/v1/2021.conll-1.28 · Cited by 37
Probing Simile Knowledge from Pre-trained Language Models
Chen, W., Chang, Y., Zhang, R., Pu, J., Chen, G., Zhang, L., Xi, Y., Chen, Y., & Su, C. (2022) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/2022.acl-long.404 · Cited by 21
Madrid, J., Diehl, P., Selig, M., Rolauffs, B., Hans, F. P., Busch, H.-J., Scheef, T., & Benning, L. (2024) · JMIR Medical Education · doi:10.2196/58375 · Cited by 18
“Since LLMs use a statistical approach to reconstruct human-written text, we anticipated difficulties with logic and mathematics, which require formal symbolic representation [4-8].”
Probing for Multilingual Numerical Understanding in Transformer-Based Language Models
Johnson, D., Mak, D., Barker, A., & Loessberg-Zahl, L. (2020) · BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP) · doi:10.18653/v1/2020.blackboxnlp-1.18 · Cited by 13
Collateral facilitation in humans and language models
Michaelov, J. A., & Bergen, B. K. (2022) · Conference on Computational Natural Language Learning · doi:10.18653/v1/2022.conll-1.2 · Cited by 21
On the Interplay Between Fine-tuning and Composition in Transformers
Yu, L., & Ettinger, A. (2021) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2021.findings-acl.201 · Cited by 14
Self-citationAutomatic Construction of Evaluation Suites for Natural Language Generation DatasetsPreprint
Mille, S., Dhole, K., Mahamood, S., Perez-Beltrachini, L., Gangal, V., Kale, M., van Miltenburg, E., & Gehrmann, S. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2106.09069 · Cited by 31
Probing LLMs for Joint Encoding of Linguistic Categories
Starace, G., Papakostas, K., Choenni, R., Panagiotopoulos, A., Rosati, M., Leidinger, A., & Shutova, E. (2023) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2023.findings-emnlp.476 · Cited by 30
When Language Models Fall in Love: Animacy Processing in Transformer Language Models
Hanna, M., Belinkov, Y., & Pezzelle, S. (2023) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2023.emnlp-main.744 · Cited by 22
Linguistically Differentiating Acts and Recalls of Racial Microaggressions on Social Media
Gunturi, U. S., Kumar, A., Ding, X., & Rho, E. H. R. (2024) · Proceedings of the ACM on Human-Computer Interaction · doi:10.1145/3637366 · Cited by 14
Mukherjee, A., Hassija, V., Chamola, V., & Gupta, K. K. (2025) · IEEE Open Journal of the Computer Society · doi:10.1109/ojcs.2025.3560333 · Cited by 19
“Additionally, LLMs like B ERT have known limitations, such as insensitivity to negation [24], word order issues [25], imprecise handling of numbers [26], and difficulty with named entities [27].”
How Relevant Are Selectional Preferences for Transformer-based Language Models?
Metheniti, E., de Cruys, T. V., & Hathout, N. (2020) · International Conference on Computational Linguistics · doi:10.18653/v1/2020.coling-main.109 · Cited by 20
“Ettinger (2020) has also pointed out that B ERT performance on predictions dropped in cases of true/false statements and negations.”
BackgroundOn Positivity Bias in Negative Reviews
Aithal, M., & Tan, C. (2021) · doi:10.18653/v1/2021.acl-short.39 · Cited by 12
“The prevalence of indirect expressions in negative reviews underscores the importance of modeling and understanding negation in sentiment analysis and sentiment transfer (Ettinger, 2020).”
BackgroundA Multilingual Benchmark for Probing Negation-Awareness with Minimal Pairs
Hartmann, M., de Lhoneux, M., Hershcovich, D., Kementchedjhieva, Y., Nielsen, L., Qiu, C., & Søgaard, A. (2021) · Conference on Computational Natural Language Learning · doi:10.18653/v1/2021.conll-1.19 · Cited by 27
“…recently been designed to explicitly investigate the negation processing capabilities of pretrained language models, either directly on the encoder (Ettinger, 2020; Kassner and Schütze, 2020) or after fine-tuning the model for downstream tasks such as natural language inference (NLI) (Naik et al.,…”
BackgroundChauhan, S., Daniel, P., Saxena, S., & Sharma, A. (2022) · Applied Artificial Intelligence · doi:10.1080/08839514.2022.2031817 · Cited by 17
“Many models on various leader boards use models from the BERT family, but BERT fails (Ettinger 2020; Kodge and Roy 2021).”
MethodNegatER: Unsupervised Discovery of Negatives in Commonsense Knowledge Bases
Safavi, T., Zhu, J. H., & Koutra, D. (2021) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2021.emnlp-main.456 · Cited by 17
“Other studies have demonstrated that pretrained LMs struggle to distinguish affirmative sentences from their negated counterparts (Kassner and Schütze, 2020; Ettinger, 2020), although this can again be addressed by fine-tuning (Kassner and Schütze, 2020; Jiang et al.”
BackgroundTextual form features for text readability assessment
Pan, W., Li, X., Chen, X., & Xu, R. (2024) · Natural language processing. · doi:10.1017/nlp.2024.50 · Cited by 14
“It possesses the inherent capability to encode linguistic knowledge, such as hierarchical parse trees (Hewitt and Manning 2019), syntactic chunks (Liu et al. 2019), and semantic roles (Ettinger 2020).”
This is a BERT. Now there are several of them. Can they generalize to novel words?
Haley, C. (2020) · BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP) · doi:10.18653/v1/2020.blackboxnlp-1.31 · Cited by 26
“Ettinger (2020) differentiates her work from these works by noting their primarily syntactic focus, and promises to test more diverse linguistic capacities, but focuses on semantic and pragmatic capacities, showing among other things that BERT fails to fully model the meaning of negation.”
BackgroundExploring Strategies for Generalizable Commonsense Reasoning with Pre-trained Models
Ma, K., Ilievski, F., Francis, J., Ozaki, S., Nyberg, E., & Oltramari, A. (2021) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2021.emnlp-main.445 · Cited by 23
“Ettinger (2020) diagnoses the BERT model, finding that it struggles with complex inference, role-based event prediction, and grasping the contextual impacts of negation.”
BackgroundAutomatic Expansion of Domain-Specific Affective Models for Web Intelligence Applications
Weichselbraun, A., Steixner, J., Brașoveanu, A., Scharl, A., Göbel, M. C., & Nixon, L. J. B. (2021) · Cognitive Computation · doi:10.1007/s12559-021-09839-4 · Cited by 12
Vahtola, T., Creutz, M., & Tiedemann, J. (2022) · BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP) · doi:10.18653/v1/2022.blackboxnlp-1.20 · Cited by 27
“BERT does not seem to be highly robust for the transformation, and it does not seem that negation alone suffices to prime the model for systematically predicting the opposite of the prediction in the affirmative sentence (Ettinger, 2020).”
BackgroundFact-Driven Abstractive Summarization by Utilizing Multi-Granular Multi-Relational Knowledge
Mao, Q., Li, J., Peng, H. M., He, S., Wang, L., Yu, P. S., & Wang, Z. (2022) · IEEE/ACM Transactions on Audio Speech and Language Processing · doi:10.1109/taslp.2022.3161157 · Cited by 16
Competence-Based Analysis of Language ModelsPreprint
Davies, A., Jiang, J., & Zhai, C. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2303.00333 · Cited by 12
Natural Language Processing RELIES on Linguistics
Opitz, J., Wein, S., & Schneider, N. (2025) · Computational Linguistics · doi:10.1162/coli_a_00560 · Cited by 10
“In the sections that follow, we tackle each of the six facets of NLP research which make up the mnemonic “RELIES,” for which we argue that linguistics has enduring relevance.”
Lorge, I., Joyce, D. W., Taylor, N., Nevado‐Holgado, A. J., Cipriani, A., & Kormilitzin, A. (2025) · Computers in Biology and Medicine · doi:10.1016/j.compbiomed.2025.110246 · Cited by 12
Bauer, L., Deng, L., & Bansal, M. (2021) · Workshop on Knowledge Extraction and Integration for Deep Learning Architectures: Deep Learning Inside Out (DeeLIO) · doi:10.18653/v1/2021.deelio-1.7 · Cited by 10
“Ettinger (2020) constructed a series of tests for language models and applied these to BERT to study strengths and weakness.”
MethodProbing BERT’s priors with serial reproduction chains
Yamakoshi, T., Griffiths, T. L., & Hawkins, R. (2022) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2022.findings-acl.314 · Cited by 18
Enhancing Class Understanding Via Prompt-Tuning For Zero-Shot Text Classification
Dan, Y., Zhou, J., Chen, Q., Bai, Q., & He, L. (2022) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp43922.2022.9746200 · Cited by 15
Multi-task Learning of Negation and Speculation for Targeted Sentiment Classification
Moore, A., & Barnes, J. (2021) · Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT) · doi:10.18653/v1/2021.naacl-main.227 · Cited by 15
A Model of Online Temporal-Spatial Integration for Immediacy and Overrule in Discourse Comprehension
Uchida, T., Lair, N., Ishiguro, H., & Dominey, P. F. (2020) · Neurobiology of Language · doi:10.1162/nol_a_00026 · Cited by 13
“Fine characterization of language processing via psycholinguistics can continue to be used to better understand what computational language models are capable of (Ettinger, 2020), and computational models can continue to be used to characterize what the brain is doing (Dehghani et al., 2017; Huth, DeHeer, et al., 2016…”
BackgroundMethodResultInfluentialKementchedjhieva, Y., Anderson, M., & Søgaard, A. (2021) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2021.findings-acl.429 · Cited by 10
Explainability of Text Processing and Retrieval Methods: A Survey
Saha, S., Majumdar, D., & Mitra, M. (2026) · ACM Computing Surveys · doi:10.1145/3801957 · Cited by 10
CrossOIE: Cross-Lingual Classifier for Open Information Extraction
Cabral, B. S., Glauber, R., Souza, M., & Claro, D. B. (2020) · Lecture notes in computer science · doi:10.1007/978-3-030-41505-1_35 · Cited by 10
Investigating Negation in Pre-trained Vision-and-language Models
Dobreva, R., & Keller, F. (2021) · BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP) · doi:10.18653/v1/2021.blackboxnlp-1.27 · Cited by 10
“Pretrained language models fail to make correct predictions in the presence of negation or even to distinguish between positive and negative sentences (Ettinger, 2020; Kassner and Schütze, 2020).”
Back to Square One: Bias Detection, Training and Commonsense Disentanglement in the Winograd SchemaPreprint
Elazar, Y., Zhang, H., Goldberg, Y., & Roth, D. (2021) · arXiv (Cornell University) · Cited by 10
Negation Neglect: When models fail to learn negations in trainingPreprint
Mayne, H., McKinney, L., Dubiński, J., Karvonen, A., Chua, J., & Evans, O. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.13829 · Cited by 10
Knowledgeable or Educated Guess? Revisiting Language Models as Knowledge Bases
Cao, B., Lin, H., Han, X., Sun, L., Yan, L., Liao, M., Xue, T., & Xu, J. (2021) · arXiv · doi:10.48550/arxiv.2106.09231 · Cited by 10
“…et al., 2019; Hewitt and Man-ning, 2019; Goldberg, 2019; Warstadt et al., 2019), semantic knowledge (Tenney et al., 2019; Wal-lace et al., 2019; Ettinger, 2020) and world knowledge (Davison et al., 2019; Bouraoui et al., 2020; Forbes et al., 2019; Zhou et al., 2019; Roberts et al., 2020; Lin…”
BackgroundSometimes We Want Ungrammatical Translations
Parthasarathi, P., Sinha, K., Pineau, J., & Williams, A. (2021) · Findings of the Association for Computational Linguistics: EMNLP 2021 · doi:10.18653/v1/2021.findings-emnlp.275 · Cited by 9
“been predominantly fairly simple, including reverse and word shuffle, and usually target only single sentences (Ettinger, 2020; Li et al., 2020; Sinha et al., 2020).”
BackgroundDahl, S. C., Bøgsted, M., Sagi, T., & Vesteghem, C. (2025) · JMIR Medical Informatics · doi:10.2196/68707 · Cited by 13
“Another study stated that BERT fails completely to show a generalizable understanding of negation, raising questions about the aptitude of language models to learn this type of meaning [50].”
Generalized Quantifiers as a Source of Error in Multilingual NLU Benchmarks
Cui, R., Hershcovich, D., & Søgaard, A. (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.359 · Cited by 18
LLM Cognitive Judgements Differ from Human
Lamprinidis, S. (2024) · Frontiers of artificial intelligence, ethics and multidisciplinary applications. · doi:10.1007/978-981-99-9836-4_2 · Cited by 16
Loconte, R., Orrù, G., Tribastone, M., Pietrini, P., & Sartori, G. (2024) · Heliyon · doi:10.1016/j.heliyon.2024.e38911 · Cited by 17
Circling the void: Using Heidegger and Lacan to think about large language models
Heimann, M., & Hübener, A.-F. (2025) · Cognitive Systems Research · doi:10.1016/j.cogsys.2025.101349 · Cited by 9
“…they are able to mark determinate absences, and this positivization of lack is central in language models, since negations are only links between positive tokens (Gubelmann & Handschuh, 2022), even if LLMs had some difficulties with this at the beginning (Ettinger, 2020; Kassner & Schütze, 2020).”
On the evolution of syntactic information encoded by BERT’s contextualized representations
Pérez-Mayos, L., Carlini, R., Ballesteros, M., & Wanner, L. (2021) · Conference of the European Chapter of the Association for Computational Linguistics · doi:10.18653/v1/2021.eacl-main.191 · Cited by 14
Morph Call: Probing Morphosyntactic Content of Multilingual Transformers
Mikhailov, V., Serikov, O., & Artemova, E. (2021) · doi:10.18653/v1/2021.sigtyp-1.10 · Cited by 13
Semantic Similarity as a Window into Vector- and Graph-Based Metrics
Leung, W. C., Wein, S., & Schneider, N. (2022) · Workshop on Generation, Evaluation and Metrics (GEM) · doi:10.18653/v1/2022.gem-1.8 · Cited by 9
“Ettinger (2019) finds that pre-trained BERT is unable to capture the effect on negation on meaning.”
BackgroundEnabling High-Level Machine Reasoning with Cognitive Neuro-Symbolic Systems
Oltramari, A. (2024) · Proceedings of the AAAI Symposium Series · doi:10.1609/aaaiss.v2i1.27701 · Cited by 9
Azzajjad, M. F., Ahmar, D. S., & Kilo, A. K. (2024) · Jambura Journal of Educational Chemistry · doi:10.37905/jjec.v6i1.22386 · Cited by 9
“Meskipun sebagian besar mahasiswa memahami materi dengan baik, masih dapat meningkatkan pemahaman mereka dengan menambahkan materi yang lebih mendalam atau studi kasus yang lebih kompleks (Ettinger, 2020).”
BackgroundDagoBERT: Generating Derivational Morphology with a Pretrained Language Model
Hofmann, V., Pierrehumbert, J. B., & Schütze, H. (2020) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2020.emnlp-main.316 · Cited by 39
“This can be related to recent work showing that BERT has difficulties with negated expressions (Ettinger, 2020; Kassner and Schütze, 2020).”
BackgroundSemEval-2022 Task 3: PreTENS-Evaluating Neural Networks on Presuppositional Semantic Knowledge
Zamparelli, R., Chowdhury, S. A., Brunato⋄, D., Chesi, C., Dell’Orletta⋄, F., Hasan, M. A., & Venturi⋄, G. (2022) · International Workshop on Semantic Evaluation (SemEval) · doi:10.18653/v1/2022.semeval-1.29 · Cited by 8
“…and Zamparelli (2018), clause embedding, Futrell et al. 2019, etc.) or at the syntax/semantics interface (negative-polarity items, Jumelet and Hupkes 2018; argument structure, quantifier restrictions, Warstadt et al. 2019), mostly using LSTM architectures (but see Tran et al. 2018; Ettinger 2020).”
MethodModeling the Influence of Verb Aspect on the Activation of Typical Event Locations with BERT
Cho, W. I., Chersoni, E., Hsu, Y.-Y., & Huang, C. (2021) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2021.findings-acl.258 · Cited by 14
“Interestingly, Ferretti et al. (2007) found that atypical locations elicited significantly larger N400 amplitudes in the imperfective condition, but there were no significant differences in the perfective one, suggesting that the differences in location typicality become salient only when the event…”
BackgroundMethodResultInfluentialShaking Syntactic Trees on the Sesame Street: Multilingual Probing with Controllable Perturbations
Taktasheva, E., Mikhailov, V., & Artemova, E. (2021) · Workshop on Multilingual Representation Learning (MRL) · doi:10.18653/v1/2021.mrl-1.17 · Cited by 16
“edge of LMs and their sensitivity to word order employing text perturbations (Futrell et al., 2018, 2019; Ettinger, 2020).”
SLABERT Talk Pretty One Day: Modeling Second Language Acquisition with BERT
Yadavalli, A., Yadavalli, A., & Tobin, V. (2023) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/2023.acl-long.657 · Cited by 34
Generative Artificial Intelligence and Authorship Gaps
Nawar, T. (2024) · American Philosophical Quarterly · doi:10.5406/21521123.61.4.05 · Cited by 11
“…beyond an inability to parse particular “tricky” relations such as negation (which has long been problematic for various distinct models, see e.g., Ettinger 2020; cf. Floridi and Chiriatti 2020) and seems to extend to various relational, syntactic, and crucial compositional features more…”
Benchmarking Gaslighting Negation Attacks Against Multimodal Large Language Models
Zhu, B., Gui, Y., Qi, H., Chen, J., Ngo, C., & Lim, E. (2026) · International Conference on Multimedia Retrieval · doi:10.1145/3805622.3810624 · Cited by 11
“Early studies showed that models like BERT struggle with distinguishing affirmation and negation [5, 12], leading to methods such as unlikelihood training to improve performance [8].”
Are Rotten Apples Edible? Challenging Commonsense Inference Ability with Exceptions
Do, N., & Pavlick, E. (2021) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2021.findings-acl.181 · Cited by 15
“However, as demonstrated by Ettinger (2020) and Kassner and Sch¨utze (2019), BERT is insensitive to various overt contextual cues, notably negation.”
BackgroundMethodAutomated Scoring of Tablet-Administered Expressive Language Tests
Gale, R. C., Bird, J., Wang, Y., van Santen, J. P. H., Prud’hommeaux, E. T., Dolata, J. K., & Asgari, M. (2021) · Frontiers in Psychology · doi:10.3389/fpsyg.2021.668401 · Cited by 8
“…corpus of unlabeled sentences—3.3 billion words from the English Wikipedia combined with BooksCorpus (Zhu et al., 2015)—language representations produced by BERT have been shown to be effective in capturing both syntactic and semantic aspect of the language (Rogers et al., 2020; Ettinger, 2020).”
BackgroundResultInfluentialChapter 13. Generalizable Neuro-Symbolic Systems for Commonsense Question Answering
Oltramari, A., Francis, J., Ilievski, F., Ma, K., & Mirzaee, R. (2021) · Frontiers in artificial intelligence and applications · doi:10.3233/faia210360 · Cited by 10
“Along these lines, recent work [13] has shown that complex inferences, role-based event prediction, and understanding the conceptual impacts of negation, are some of the missing capabilities diagnosed when BERT [14], one of the most popular and versatile opensource language models available today, is applied to benchm…”
BackgroundDoes BERT Know that the IS-A Relation Is Transitive?
Lin, R., & Ng, H. T. (2022) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/2022.acl-short.11 · Cited by 21
“…predictions but for the wrong reasons, including being insensitive to negation and exploiting only surface features (Kassner and Schutze, 2020; Ettinger, 2020), lacking understanding of perceptual properties (Forbes et al., 2019; Weir et al., 2020), and surface form competition (Holtzman et…”
BackgroundLogiTorch: A PyTorch-based library for logical reasoning on natural language
Helwe, C., Clavel, C., & Suchanek, F. M. (2022) · Conference on Empirical Methods in Natural Language Processing: System Demonstrations (EMNLP) · doi:10.18653/v1/2022.emnlp-demos.25 · Cited by 16
“However, the models can be distracted easily by trap words, syntactic variations (Kass-ner and Schütze, 2020), or negation (Kassner and Schütze, 2020; Ettinger, 2020; Hossain et al., 2020, 2022; Helwe et al., 2021).”
BackgroundA Semantic Approach to Negation Detection and Word Disambiguation with Natural Language Processing
Okpala, I., Rodriguez, G. R., Tapia, A., Halse, S. E., & Kropczynski, J. (2022) · International Conference on Natural Language Processing and Information Retrieval (NLPIR) · doi:10.1145/3582768.3582789 · Cited by 8
“Despite these advances, studies [8] have shown that even these approaches have still struggled with negation.”
Backgrounddos Santos, A. M., Moura, F. R. T., Pinto, L. V. L., Alves, A. V. N., Figueiredo, K., Costa, F. A. R., & Seruffo, M. C. D. R. (2023) · IEEE Access · doi:10.1109/access.2023.3261339 · Cited by 9
“However, there are drawbacks in its use due to its complexity in training and testing [65].”
BackgroundTransparency Helps Reveal When Language Models Learn Meaning
Wu, Z., Merrill, W., Peng, H. M., Beltagy, I., & Smith, N. A. (2023) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00565 · Cited by 19
“Practically, through the lens of strong transparency, our results supplement prior studies that identified pretrained LMs’ insufficient semantic representations (Tenney et al., 2019; Yu and Ettinger, 2020, 2021; Wu et al., 2021; i.a. ).”
BackgroundResultFrom chocolate bunny to chocolate crocodile: Do Language Models Understand Noun Compounds?
Coil, A., & Shwartz, V. (2023) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2023.findings-acl.169 · Cited by 8
Can Pretrained Language Models (Yet) Reason Deductively?
Yuan, Z., Hu, S., Vulić, I., Korhonen, A., & Meng, Z. (2023) · Conference of the European Chapter of the Association for Computational Linguistics · doi:10.18653/v1/2023.eacl-main.106 · Cited by 15
Fan, C., Tian, J., Li, Y., Chen, W., He, H., & Jin, Y. (2023) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2023.emnlp-main.913 · Cited by 11
Can Language Models Be Tricked by Language Illusions? Easier with Syntax, Harder with Semantics
Zhang, Y., Gibson, E. A. F., & Davis, F. (2023) · Conference on Computational Natural Language Learning · doi:10.18653/v1/2023.conll-1.1 · Cited by 13
“…dependency between the li-censor and the NPI is captured by language models (Jumelet and Hupkes, 2018; Jumelet et al., 2021; Shin et al., 2023) but with more difficulty as compared to subject-verb agreement or other syntactic dependencies (Marvin and Linzen, 2018; Warstadt et al., 2019, 2020).”
BackgroundRethinking visual prompt learning as masked visual token modeling
Liao, N., Shi, B., Zhang, X., Cao, M., Yan, J., & Tian, Q. C. (2025) · Artificial Intelligence · doi:10.1016/j.artint.2025.104417 · Cited by 9
Evaluating Text Generation from Discourse Representation Structures
Wang, C., van Noord, R., Bisazza, A., & Bos, J. (2021) · Workshop on Generation, Evaluation and Metrics (GEM) · doi:10.18653/v1/2021.gem-1.8 · Cited by 8
“Interestingly, recent work in machine translation (Tang, 2020) and language modelling (Ettinger, 2020) has also shown that state-of-the-art neural models still struggle with handling negation.”
BackgroundInspecting the concept knowledge graph encoded by modern language models
Aspillaga, C., Mendoza, M., & Soto, Á. (2021) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2021.findings-acl.263 · Cited by 20
Exploring the Capacity of a Large-scale Masked Language Model to Recognize Grammatical Errors
Nagata, R., Kimura, M., & Hanawa, K. (2022) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2022.findings-acl.324 · Cited by 8
MacBehaviour: An R package for behavioural experimentation on large language models
Duan, X., Li, S., & Cai, Z. G. (2024) · Behavior Research Methods · doi:10.3758/s13428-024-02524-y · Cited by 8
Multilingual Negation Scope Resolution for Clinical Text
Hartmann, M., & Søgaard, A. (2021) · Research at the University of Copenhagen (University of Copenhagen) · Cited by 10
Same Sentence Prediction: A new Pre-training Task for BERT
Sonmezoz, K., & Amasyalı, M. F. (2021) · Innovations in Intelligent Systems and Applications Conference (ASYU) · doi:10.1109/asyu52992.2021.9598954 · Cited by 7
Life after BERT: What do Other Muppets Understand about Language?
Lialin, V., Zhao, K., Shivagunde, N., & Rumshisky, A. (2022) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/2022.acl-long.227 · Cited by 10
“One drawback of datasets from Ettinger (2020) that we have noticed was the ambiguity of answers.”
BackgroundMethodResultInfluentialCan Negation Be Depicted? Comparing Human and Machine Understanding of Visual Representations
Sato, Y., Mineshima, K., & Ueda, K. (2023) · Cognitive Science · doi:10.1111/cogs.13258 · Cited by 7
Disentangling Societal Inequality from Model Biases: Gender Inequality in Divorce Court Proceedings
Dutta, S., Srivastava, P., Solunke, V., Nath, S., & KhudaBukhsh, A. R. (2023) · International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2023/661 · Cited by 7
Hypert: hypernymy-aware BERT with Hearst pattern exploitation for hypernym discovery
Yun, G., Lee, Y., Moon, A.-S., & Lee, J. (2023) · Journal Of Big Data · doi:10.1186/s40537-023-00818-0 · Cited by 7
“To use a PLM for identifying hypernym relationships, [16] evaluated the ability of BERT through human language experiments called prompting and demonstrated proficient results in hypernym retrieval.”
MethodProbing Quantifier Comprehension in Large Language Models: Another Example of Inverse Scaling
Gupta, A. (2023) · BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP) · doi:10.18653/v1/2023.blackboxnlp-1.4 · Cited by 12
Mosbach, M., Degaetano‐Ortlieb, S., Krielke, M.-P., Abdullah, B. M., & Klakow, D. (2020) · International Conference on Computational Linguistics · doi:10.18653/v1/2020.coling-main.67 · Cited by 7
NwQM: A neural quality assessment framework for Wikipedia
Guda, B. P. R., Seelaboyina, S. B., Sarkar, S., & Mukherjee, A. (2020) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2020.emnlp-main.674 · Cited by 10
Davis, F., & Altmann, G. T. M. (2021) · Cognition · doi:10.1016/j.cognition.2021.104651 · Cited by 9
“Although this has still to be systematically tested (see Ettinger, 2020, for evidence suggesting that BERT lacks event knowledge; and Tran, Vusazza, & Monz, 2018, and Abnar, Dehghani, & Zuidema, 2020, for further elucidation of the role of recurrence within NLP), we did briefly explore whether the…”
Situation and Behavior Understanding by Trope Detection on Films
Chang, C.-H., Su, H.-T., Hsu, J.-H., Wang, Y.-S., Chang, Y., Liu, Z. Y., Chang, Y.-L., Cheng, W., Wang, K.-J., & Hsu, W. H. (2021) · International World Wide Web Conference (WWW) · doi:10.1145/3442381.3449806 · Cited by 7
“Ettinger [10] highlighted that one of the limitations of BERT model is being insensitive to role-based events.”
BackgroundInfluentialUncovering Constraint-Based Behavior in Neural Models via Targeted Fine-Tuning
Davis, F., & van Schijndel, M. (2021) · doi:10.18653/v1/2021.acl-long.93 · Cited by 12
Laifa, A., Gautier, L., & Cruz, C. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-86534-4_13 · Cited by 7
Using Computational Cognitive Modeling in Usage‐Based Linguistics
Divjak, D., & Milin, P. (2023) · Wiley eBooks · doi:10.1002/9781119839859.ch17 · Cited by 7
PENDAMPINGAN LITERASI MELALUI DISEMINASI KEGIATAN PENGEMBANGAN MEDIA PEMBELAJARAN
Azzajjad, M. F., & Ahmar, D. S. (2024) · DISEMINASI Jurnal Pengabdian kepada Masyarakat · doi:10.33830/diseminasiabdimas.v6i1.6454 · Cited by 5
“Indonesia memiliki banyak bahasa dan dialek yang berbeda, dan bahasa pengantar dalam pendidikan dapat memengaruhi tingkat literasi (Ettinger, 2020).”
BackgroundBenchmarking Gaslighting Negation Attacks Against Reasoning Models
Zhu, B., Yin, H., Chen, J., & Jiang, Y.-G. (2026) · Lecture notes in computer science · doi:10.1007/978-981-95-6960-1_14 · Cited by 6
Towards Semantics-Enhanced Pre-Training: Can Lexicon Definitions Help Learning Sentence Meanings?
Ren, X., Sun, X., Wang, H., & Liu, Q. (2021) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v35i15.17619 · Cited by 6
“As syntax learning and semantics learning are highly coupled, this stage also promotes the model in some semantic aspects, e.g., selectional preferences (Ettinger 2020; Tenney et al. 2019), which have prominent syntactic clues.”
BackgroundModelling General Properties of Nouns by Selectively Averaging Contextualised Embeddings
Li, N., Bouraoui, Z., Camacho-Collados, J., Espinosa-Anke, L., Gu, Q., & Schockaert, S. (2021) · International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2021/530 · Cited by 11
About Time: Do Transformers Learn Temporal Verbal Aspect?
Metheniti, E., de Cruys, T. V., & Hathout, N. (2022) · Workshop on Cognitive Modeling and Computational Linguistics (CMCL) · doi:10.18653/v1/2022.cmcl-1.10 · Cited by 17
“However, it remains unclear whether the success of these models is due to a genuine capability to accurately model linguistic meaning, or whether the models are just very good at picking up statistical correlations, but fail to capture fine-grained semantic distinctions (Ettinger, 2020).”
BackgroundBetz, G., & Richardson, K. (2023) · PLoS ONE · doi:10.1371/journal.pone.0281372 · Cited by 12
“…successes of NLMs, studies that have carefully probed current systems equally point out major limitations that are pertinent to the Rationality Hypothesis , namely: NLMs’ sheer inability to produce coherent and consistent output [47, 48], along with a failure to properly handle negations [49–51].”
Quantifying Structural and Non‐structural Expectations in Relative Clause Processing
Chen, Z., & Hale, J. (2021) · Cognitive Science · doi:10.1111/cogs.12927 · Cited by 8
Beyond Distributional Hypothesis: Let Language Models Learn Meaning-Text Correspondence
Jang, M., Mtumbuka, F., & Lukasiewicz, T. (2022) · Findings of the Association for Computational Linguistics: NAACL · doi:10.18653/v1/2022.findings-naacl.156 · Cited by 13
“…conducted various probing tasks and observed that PLMs exhibit faulty behaviours, such as insensitiveness to sentence ordering (Pham et al., 2021; Gupta et al., 2021; Sinha et al., 2021b), incomprehension on number-related representations (Wallace et al., 2019; Lin et al., 2020; Nogueira et al.,…”
Metadata
- Authors
- Ettinger, A.
- Affiliations
- University of Chicago (US)
- Corresponding author
- Ettinger, Allyson
- Venue
- Transactions of the Association for Computational Linguistics
- Publisher
- MIT Press
- Published
- 2020-12
- Volume
- 8
- Pages
- 34-48
- ISSN
- 2307-387X (ISSN-L)
- OpenAlex
- W2965647350
- Cited by
- 511
- References
- 38
- Citations by year
- 2026: 16 · 2025: 32 · 2024: 39 · 2023: 122 · 2022: 92 · 2021: 151 · 2020: 53
- FWCI
- 41.25
- 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, cc-by-sa)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Social Sciences › Arts and Humanities › Language and Linguistics › Syntax, Semantics, Linguistic Variation
- Keywords
- BERT, language model pretraining, role reversal, negation, event prediction, language model evaluation
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.