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: 4
- Allyson Ettinger, Ahmed Elgohary, Colin Phillips, and Philip Resnik. 2018. Assessing composition in sentence vector representations. In Proceedings of the 27th International Conference on Computational Linguistics , pages 1790-1801.
- Mario Giulianelli, Jack Harding, Florian Mohnert, Dieuwke Hupkes, and Willem Zuidema. 2018. Under the hood: Using diagnostic classifiers to investigate and improve how language models track agreement information. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP , pages 240-248. · DOI
- Yair Lakretz, Germán Kruszewski, Théo Desbordes, Dieuwke Hupkes, Stanislas Dehaene, and Marco Baroni. 2019. The emergence of number and syntax units in LSTM language models. In Proceedings of the 2019 Conference of the North American Chapter of the Association for
- Ian Tenney, Patrick Xia, Berlin Chen, Alex Wang, Adam Poliak, R. Thomas McCoy, Najoung Kim, Benjamin Van Durme, Samuel R. Bowman, Dipanjan Das, and Ellie Pavlick. 2019b. What do you learn from context? Probing for sentence structure in contextualized word representations. In Proceedings of the International Conference on Learning Representations .
Cited by
Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., & Neubig, G. (2023) · ACM Computing Surveys · doi:10.1145/3560815 · Cited by 4,361
“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].”
BackgroundOn the Opportunities and Risks of Foundation Models
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. (2021) · arXiv · doi:10.48550/arxiv.2108.07258 · Cited by 2,676
Pre-trained models for natural language processing: A survey
Qiu, X., Sun, T., Xu, Y., Shao, Y., Dai, N., & Huang, X. (2020) · Science China Technological Sciences · doi:10.1007/s11431-020-1647-3 · Cited by 1,591
“zed the roles of BERT’s layers in dierent tasks and found that BERT solves tasks in a similar order to that in NLP pipelines. Furthermore, knowledge of subject-verb agreement [50] and semantic roles [44] are also confirmed to exist in BERT. Besides, Hewitt and Manning [59], Jawahar et al. [72], Kim et al. [85] proposed…”
MethodA primer in BERTology: What we know about how BERT works
Rogers, A., Kovaleva, O., & Rumshisky, A. (2020) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00349 · Cited by 1,552
“However, we do have evidence from an MLM probing study that BERT has some knowledge for semantic roles (Ettinger, 2019).”
BackgroundRecent Advances in Natural Language Processing via Large Pre-trained Language Models: A Survey
Min, B., Ross, H., Sulem, E., Veyseh, A. P. B., Nguyen, T. H., Sainz, O., Agirre, E., Heintz, I., & Roth, D. (2024) · ACM Computing Surveys · doi:10.1145/3605943 · Cited by 1,387
“Prompts can also be used to explore the linguistic knowledge of PLMs, focusing on diferent phenomena such as analogies [39], negation [43] or semantic similarity [42].”
Background
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.