Did the Model Understand the Question?
Open accessThe DOI matches its Crossref record; 36 of the 36 references it cites were found in a registry or an index.
Abstract
We analyze state-of-the-art deep learning models for three tasks: question answering on (1) images, (2) tables, and (3) passages of text.Using the notion of attribution (word importance), we find that these deep networks often ignore important question terms.Leveraging such behavior, we perturb questions to craft a variety of adversarial examples.Our strongest attacks drop the accuracy of a visual question answering model from 61.1% to 19%, and that of a tabular question answering model from 33.5% to 3.3%.Additionally, we show how attributions can strengthen attacks proposed by Jia and Liang (2017) on paragraph comprehension models.Our results demonstrate that attributions can augment standard measures of accuracy and empower investigation of model performance.When a model is accurate but for the wrong reasons, attributions can surface erroneous logic in the model that indicates inadequacies in the test data.
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.
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
Glove: Global Vectors for Word Representation
Pennington, J., Socher, R., & Manning, C. (2014) · Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.3115/v1/d14-1162 · Cited by 34,071 · cited 1×
“nearest word in GloVe word vector space (Pennington et al., 2014) and (c) crowdsourcing to check that the new sentence is grammatically correct.”
Background"Why Should I Trust You?": Explaining the Predictions of Any Classifier
Ribeiro, M. T., Singh, S., & Guestrin, C. (2016) · Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining · doi:10.1145/2939672.2939778 · Cited by 17,555
“Why Should I Trust You?”: Explaining the Predictions of Any Classifier
Ribeiro, M., Singh, S., & Guestrin, C. (2016) · Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations · doi:10.18653/v1/n16-3020 · Cited by 6,682
SQuAD: 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 · cited 1×
“The SQuAD dataset (Rajpurkar et al., 2016) for machine reading comprehension contains 107.”
Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
Simonyan, K., Vedaldi, A., & Zisserman, A. (2013) · arXiv · doi:10.48550/arxiv.1312.6034 · Cited by 5,370 · cited 1×
“We could use many other methods instead of Integrated Gradients (IG) to attribute a deep network’s prediction to its input features (Baehrens et al., 2010; Simonyan et al., 2013; Shrikumar et al., 2016; Binder et al., 2016; Springenberg et al., 2014).”
MethodAxiomatic attribution for deep networks
Sundararajan, M., Taly, A., & Yan, Q. (2017) · arXiv · doi:10.48550/arxiv.1703.01365 · Cited by 4,087 · cited 5×
“First, we apply Integrated Gradients (hence-forth, IG) (Sundararajan et al., 2017) to attribute the systems’ predictions to words in the questions.”
BackgroundMethodInfluentialSelf-citationStriving for Simplicity: The All Convolutional Net
Springenberg, J. T., Dosovitskiy, A., Brox, T., & Riedmiller, M. (2014) · arXiv · doi:10.48550/arxiv.1412.6806 · Cited by 3,447 · cited 1×
“We could use many other methods instead of Integrated Gradients (IG) to attribute a deep network’s prediction to its input features (Baehrens et al., 2010; Simonyan et al., 2013; Shrikumar et al., 2016; Binder et al., 2016; Springenberg et al., 2014).”
MethodMaking the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering
Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., & Parikh, D. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.670 · Cited by 2,287
Multimodal Compact Bilinear Pooling for Visual Question Answering and Visual Grounding
Fukui, A., Park, D. H., Yang, D., Rohrbach, A., Darrell, T., & Rohrbach, M. (2016) · Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d16-1044 · Cited by 1,582 · cited 2×
“1 % accuracy on the validation set (the state of the art (Fukui et al., 2016) achieves 66 .”
BackgroundAdversarial Examples for Evaluating Reading Comprehension Systems
Jia, R., & Liang, P. (2017) · Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d17-1215 · Cited by 1,410 · cited 9×
“Additionally, using insights derived from attributions we were able to improve the attack success rate of Jia and Liang (2017) (section 6.2).”
BackgroundMethodInfluentialVQA: Visual Question Answering
Agrawal, A., Lu, J., Antol, S., Mitchell, M., Zitnick, C. L., Batra, D., & Parikh, D. (2015) · arXiv (Cornell University) · doi:10.48550/arxiv.1505.00468 · Cited by 900
Visual7W: Grounded Question Answering in Images
Zhu, Y., Groth, O., Bernstein, M., & Fei-Fei, L. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.540 · Cited by 794 · cited 1×
“The Visual Question Answering Task (Agrawal et al., 2015; Teney et al., 2017; Kazemi and Elqursh, 2017; Ben-younes et al., 2017; Zhu et al., 2016) requires a system to answer questions about images (fig.”
BackgroundMUTAN: Multimodal Tucker Fusion for Visual Question Answering
Ben-younes, H., Cadene, R., Cord, M., & Thome, N. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.285 · Cited by 716 · cited 2×
“The Visual Question Answering Task (Agrawal et al., 2015; Teney et al., 2017; Kazemi and Elqursh, 2017; Ben-younes et al., 2017; Zhu et al., 2016) requires a system to answer questions about images (fig.”
BackgroundSQuAD: 100,000+ Questions for Machine Comprehension of Text
Rajpurkar, P., Zhang, J., Lopyrev, K., & Liang, P. (2016) · arXiv · doi:10.48550/arxiv.1606.05250 · Cited by 678 · cited 1×
“The SQuAD dataset (Rajpurkar et al., 2016) for machine reading comprehension contains 107.”
VQA: Visual Question Answering: www.visualqa.org
Agrawal, A., Lu, J., Antol, S., Mitchell, M., Zitnick, C. L., Parikh, D., & Batra, D. (2017) · International Journal of Computer Vision · doi:10.1007/s11263-016-0966-6 · Cited by 570
Compositional Semantic Parsing on Semi-Structured Tables
Pasupat, P., & Liang, P. (2015) · Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) · doi:10.3115/v1/p15-1142 · Cited by 514
Not Just a Black Box: Learning Important Features Through Propagating Activation Differences
Shrikumar, A., Greenside, P., Shcherbina, A., & Kundaje, A. B. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1605.01713 · Cited by 496 · cited 1×
“We could use many other methods instead of Integrated Gradients (IG) to attribute a deep network’s prediction to its input features (Baehrens et al., 2010; Simonyan et al., 2013; Shrikumar et al., 2016; Binder et al., 2016; Springenberg et al., 2014).”
MethodMaking the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering
Goyal, Y., Khot, T., Agrawal, A., Summers-Stay, D., Batra, D., & Parikh, D. (2019) · International Journal of Computer Vision · doi:10.1007/s11263-018-1116-0 · Cited by 487 · cited 1×
“Goyal et al. (2016) propose the VQA 2.0 dataset, which has pairs of similar images that have different answers on the same question.”
BackgroundInfluentialLayer-Wise Relevance Propagation for Neural Networks with Local Renormalization Layers
Binder, A., Montavon, G., Lapuschkin, S., Müller, K.-R., & Samek, W. (2016) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-44781-0_8 · Cited by 418
Tips and Tricks for Visual Question Answering: Learnings from the 2017 Challenge
Teney, D., Anderson, P., He, X., & Hengel, A. V. D. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00444 · Cited by 396
Multimodal Compact Bilinear Pooling for Visual Question Answering and Visual GroundingPreprint
Fukui, A., Park, D. H., Yang, D., et al. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1606.01847 · Cited by 388 · cited 2×
“1 % accuracy on the validation set (the state of the art (Fukui et al., 2016) achieves 66 .”
BackgroundYin and Yang: Balancing and Answering Binary Visual Questions
Zhang, P., Goyal, Y., Summers-Stay, D., Batra, D., & Parikh, D. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.542 · Cited by 312 · cited 1×
“Kafle and Kanan (2017); Zhang et al. (2016) examine the VQA data, identify de-ficiencies, and propose data augmentation to reduce over-representation of certain question/answer types.”
BackgroundAnalyzing the Behavior of Visual Question Answering Models
Agrawal, A., Batra, D., & Parikh, D. (2016) · Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d16-1203 · Cited by 297 · cited 1×
“Agrawal et al. (2016) analyze several VQA models.”
Adversarial Examples for Evaluating Reading Comprehension SystemsPreprint
Jia, R., & Liang, P. (2017) · arXiv (Cornell University) · doi:10.48550/arxiv.1707.07328 · Cited by 259 · cited 9×
“Additionally, using insights derived from attributions we were able to improve the attack success rate of Jia and Liang (2017) (section 6.2).”
BackgroundMethodInfluentialAn Analysis of Visual Question Answering Algorithms
Kafle, K., & Kanan, C. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.217 · Cited by 247 · cited 1×
“Kafle and Kanan (2017); Zhang et al. (2016) examine the VQA data, identify de-ficiencies, and propose data augmentation to reduce over-representation of certain question/answer types.”
BackgroundNeural Semantic Parsing with Type Constraints for Semi-Structured Tables
Krishnamurthy, J., Dasigi, P., & Gardner, M. (2017) · Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d17-1160 · Cited by 229 · cited 4×
“(2018) note that the state-of-the-art strongly supervised6 model on WikiTableQuestions (Krishnamurthy et al., 2017) enjoys a 7% gain in its final accuracy by leveraging perturbed tables during training.”
BackgroundInfluentialShow, Ask, Attend, and Answer: A Strong Baseline For Visual Question Answering
Kazemi, V., & Elqursh, A. (2017) · arXiv (Cornell University) · doi:10.48550/arxiv.1704.03162 · Cited by 130 · cited 5×
“Table 1: VQA network (Kazemi and Elqursh, 2017): Accuracy for prefix attacks; original accuracy is 61.”
BackgroundInfluentialCompositional Semantic Parsing on Semi-Structured Tables
Pasupat, P., & Liang, P. (2015) · arXiv (Cornell University) · doi:10.48550/arxiv.1508.00305 · Cited by 91 · cited 1×
“We now analyze question answering over tables based on the WikiTableQuestions benchmark dataset (Pasupat and Liang, 2015).”
MethodNothing Else Matters: Model-Agnostic Explanations By Identifying Prediction Invariance
Ribeiro, M. T., Singh, S. K., & Guestrin, C. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1611.05817 · Cited by 59 · cited 2×
“One could also use model agnostic techniques like Ribeiro et al. (2016b).”
MethodInferring Logical Forms From Denotations
Pasupat, P., & Liang, P. (2016) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/p16-1003 · Cited by 56 · cited 2×
“We quantify this weakness by evaluating NP on the set of perturbed tables generated by Pasupat and Liang (2016) and find that its accuracy drops from 33 .”
MethodHow to Explain Individual Classification Decisions
(2010) · Journal of Machine Learning Research · doi:10.5555/1756006.1859912 · Cited by 54 · cited 1×
“We could use many other methods instead of Integrated Gradients (IG) to attribute a deep network’s prediction to its input features (Baehrens et al., 2010; Simonyan et al., 2013; Shrikumar et al., 2016; Binder et al., 2016; Springenberg et al., 2014).”
MethodLayer-wise Relevance Propagation for Neural Networks with Local Renormalization LayersPreprint
Binder, A., Montavon, G., Bach, S., et al. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1604.00825 · Cited by 24 · cited 1×
“We could use many other methods instead of Integrated Gradients (IG) to attribute a deep network’s prediction to its input features (Baehrens et al., 2010; Simonyan et al., 2013; Shrikumar et al., 2016; Binder et al., 2016; Springenberg et al., 2014).”
MethodAn Analysis of Visual Question Answering AlgorithmsPreprint
Kafle, K., & Kanan, C. (2017) · arXiv (Cornell University) · doi:10.48550/arxiv.1703.09684 · Cited by 22 · cited 1×
“Kafle and Kanan (2017); Zhang et al. (2016) examine the VQA data, identify de-ficiencies, and propose data augmentation to reduce over-representation of certain question/answer types.”
BackgroundInferring Logical Forms From DenotationsPreprint
Pasupat, P., & Liang, P. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1606.06900 · Cited by 13 · cited 2×
“We quantify this weakness by evaluating NP on the set of perturbed tables generated by Pasupat and Liang (2016) and find that its accuracy drops from 33 .”
MethodIt was the training data pruning too!Preprint
Mudrakarta, P. K., Taly, A., Sundararajan, M., & Dhamdhere, K. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1803.04579 · Cited by 6 · cited 1×
“Indeed, Mudrakarta et al. (2018) note that the state-of-the-art strongly supervised 6 model on WikiTableQuestions (Krishnamurthy et al., 2017) enjoys a 7% gain in its final accuracy by leveraging perturbed tables during training.”
BackgroundSelf-citationA Novel Framework for Robustness Analysis of Visual QA ModelsPreprint
Huang, J.-H., Dao, C. D., Alfadly, M., et al. (2017) · arXiv (Cornell University) · doi:10.48550/arxiv.1711.06232 · Cited by 2 · cited 1×
“Huang et al. (2017) evaluate robustness of VQA models by appending questions with semantically similar questions.”
Method
Research works printed but not linked: 6
- Ruth C Fong and Andrea Vedaldi. 2017. Interpretable explanations of black boxes by meaningful pertur- bation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3429-3437.
- Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. 2015. Explaining and harnessing adversar- ial examples. In International Conference on Learn- ing Representations.
- Robin Jia and Percy Liang. 2017. Adversarial ex- amples for evaluating reading comprehension sys- tems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Process- ing, EMNLP 2017, Copenhagen, Denmark, Septem- ber 9-11, 2017.
- Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In European confer- ence on computer vision, pages 740-755. Springer.
- Arvind Neelakantan, Quoc V Le, and Ilya Sutskever. 2016. Neural programmer: Inducing latent pro- grams with gradient descent. In International Con- ference on Learning Representations ICLR.
- Adams Wei Yu, David Dohan, Quoc Le, Thang Luong, Rui Zhao, and Kai Chen. 2018. Fast and accurate reading comprehension by combining self-attention and convolution. In International Conference on Learning Representations.
Cited by
Explainable AI: a review of machine learning interpretability methods
Linardatos, P., Papastefanopoulos, V., & Kotsiantis, S. (2020) · Entropy · doi:10.3390/e23010018 · Cited by 2,922
“Another work, closely related to the integrated gradients method, was proposed in [30], where attributions are used in order to help identify weaknesses of three question-answer models better than the conventional models, while also to provide workflow superiority.”
Model Cards for Model Reporting
Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I. D., & Gebru, T. (2019) · Proceedings of the Conference on Fairness, Accountability, and Transparency · doi:10.1145/3287560.3287596 · Cited by 2,681
“Additional details may include, for example, interpretability approaches, such as saliency maps, TCAV [33], and Path-Integrated Gradients [38, 43]); stakeholder-relevant explanations (e.”
BackgroundTechniques for interpretable machine learning
Du, M., Liu, N., & Hu, X. (2019) · Communications of the ACM · doi:10.1145/3359786 · Cited by 1,268
“ist among training data. For instance, a post-hoc attribution approach has been utilized to analyze three question answering models and attributes the models' prediction to words in the question [24]. The attribution visualizations show that these models often ignore the important part of the questions and rely o…”
MethodTechniques for Interpretable Machine Learning
Du, M., Liu, N., & Hu, X. (2018) · arXiv · doi:10.48550/arxiv.1808.00033 · Cited by 1,261
Analysis Methods in Neural Language Processing: A Survey
Belinkov, Y., & Glass, J. (2019) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00254 · Cited by 477
Metadata
- Authors
- Mudrakarta, P. K., Taly, A., Sundararajan, M., & Dhamdhere, K.
- Affiliations
- University of Chicago (US)
- Google (United States) (US)
- Venue
- Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
- Published
- 2018
- Pages
- 1896-1906
- ISSN
- 0736-587X (ISSN-L)
- OpenAlex
- W2798312813
- Cited by
- 166
- References
- 26
- Citations by year
- 2026: 2 · 2025: 6 · 2024: 8 · 2023: 14 · 2022: 26 · 2021: 40 · 2020: 37 · 2019: 29 · 2018: 4
- FWCI
- 10.94
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- Locations
- Proceedings of the conference - Association for Computational Linguistics. Meeting (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Keywords
- adversarial examples, question answering, visual question answering, table question answering, text comprehension, word importance, attribution, model interpretability, model robustness
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.