What makes a document’s sentiment positive? What makes a loan applicant creditworthy? What makes a job candidate qualified? What about a photograph truly makes it depict a dolphin? Moreover, what does it mean for a feature to be relevant to such a determination?
Statistical learning offers one framework for approaching these questions. First, we swap out the semantic question for a more readily answerable associative question. For example, instead of asking what comprises a document’s sentiment, we recast the question as which documents are likely to be labeled as positive (or negative)? Then, in this associative framing, we interpret as relevant, those features that are most predictive
of the label. However, despite the rapid adoption and undeniable commercial success of associative learning, this framing seems unsatisfying.
Alongside deep learning’s predictive wins, critical questions have piled up concerningspuriousness, artifacts, reliability, and discrimination12, 5] or clues in the image’s background to recognize foreground objects even when that seems both unnecessary and somehow wrong: the beach is not what makes a seagull a seagull. And yet researchers struggle to articulate precisely why models should not rely on such patterns.
In NLP, these issues have emerged as central concerns in the literature on annotation artifacts and bias (in the societal sense). Across myriad tasks, researchers have demonstrated that models tend to rely on spurious associations [23, 8, 14, 16]. Notably, some models for question-answering tasks may not actually be sensitive to the choice of the question , while in Natural Language Inference (NLI), classifiers trained on hypotheses only (vs hypotheses and premises) perform surprisingly well [23, 8]. However, papers seldom make clear what, if anything, spuriousness
means within the standard supervised learning framework. ML systems are trained to exploit the mutual information between features and a label to make accurate predictions. Statistical learning does not offer a conceptual distinction between between spurious and non-spurious associations.
Causality, however, offers a coherent notion of spuriousness. Spurious associations owe to common cause rather than to a (direct or indirect) causal path. We might consider a factor of variation to be spuriously correlated with a label of interest if intervening upon it (counterfactually) would not impact the applicability of the label or vice versa. While our paper does not rely on the mathematical machinery of causality, we draw inspiration from the underlying philosophy to design a new dataset creation procedure in which humans counterfactually augment datasets.
Returning to NLP, even though the raw data does not come neatly disentangled into manipulable factors, people nevertheless speak colloquially of editing documents to manipulate specific aspects . For example, the following interventions seem natural: (i) Revise the letter to make it more positive; (ii) Edit the second sentence so that it appears to contradict the first. The very notion of targeted revisions like (i) suggests a generative process in which the sentiment is but one (manipulable) cause of the final document. These edits might be thought of as intervening on sentiment while holding all upstream features constant. However even if some other factor has no influence on sentiment, if they share some underlying common cause (confounding), then we might expect aspects of the final document to be predictive of sentiment owing to spurious association.
In this exploratory paper, we design a human-in-the-loop system for counterfactually manipulating documents. Our hope is that by intervening only upon the factor of interest, we might disentangle the spurious and non-spurious associations, yielding classifiers that hold up better when spurious associations do not transport out of sample. We employ crowd workers not to label documents, but rather to edit them, manipulating the text to make a targeted (counterfactual) class apply. For sentiment analysis, we direct the worker: revise this negative movie review to make it positive, without making any gratuitous changes. We might regard the second part of this directive as a sort of least action principle, ensuring that we perturb only those spans necessary to alter the applicability of the label. For NLI, a -class classification task (entailment, contradiction, neutral), we ask the workers to modify the premise while keeping the hypothesis intact, and vice versa, seeking two sets of edits corresponding to each of the (two) counterfactual classes. Using this platform, we collect thousands of counterfactually-manipulated examples for both sentiment analysis and NLI, extending the IMDb  and SNLI  datasets, respectively. The result is two new datasets (each an extension of a standard resource) that enable us to both probe fundamental properties of language and train classifiers less reliant on spurious signal.
We show that classifiers trained on original IMDb reviews fail on counterfactually-revised data and vice versa, and spurious correlations in these datasets are picked up by even linear models, however, augmenting the revised examples breaks up these correlations (e.g., genre ceases to be predictive of sentiment). For a Bidirectional LSTM  trained on IMDb reviews, classification accuracy goes down from to when evaluated on original vs revised reviews. The same classifier trained on revised reviews achieves an accuracy of on original reviews compared to on their revised counterparts. These numbers go to and respectively when the classifier is retrained on the combined dataset. Similar behavior is observed for linear classifiers. We discovered that BERT  is more resilient to such drops in performance on sentiment analysis. Despite that, it appears to rely on spurious associations in SNLI hypotheses identified by Gururangan et al. . We show that if fine-tuned on SNLI sentence pairs, BERT fails on pairs with revised premise and vice versa, experiencing more than a point drop in accuracy. However, fine-tuned on the combined set, it performs much better across all datasets. Similarly, a Bi-LSTM trained on hypotheses alone can accurately classify of the SNLI dataset but performs worse than the majority class baseline when evaluated on the revised dataset. When trained on hypotheses only from the combined dataset, its performance is expectedly worse than simply selecting the majority class on both SNLI as well as the revised dataset. Our data will be made available at https://github.com/dkaushik96/bizarro-data.
2 Related Work
Several papers demonstrate cases where NLP systems appear not to learn what humans consider to be the difference that makes the difference. For example, otherwise state-of-the-art models have been shown to be vulnerable to synthetic transformations such as distractor phrases [11, 27], to misclassify paraphrased task [10, 22] and to fail on template-based modifications . Glockner et al.  demonstrate that simply replacing words by synonyms or hypernyms, which should not alter the applicable label, nevertheless breaks ML-based NLI systems. Gururangan et al.  and Poliak et al.  show that classifiers correctly classified the hypotheses alone in about
of SNLI corpus. They further discover that crowd workers adopted specific annotation strategies and heuristics for data generation.Chen et al.  identify similar issues exist with automatically-constructed benchmarks for question-answering . Kaushik & Lipton  discover that reported numbers in question-answering benchmarks could often be achieved by the same models when restricted to be blind either to the question or to the passages. Dixon et al. , Zhao et al.  and Kiritchenko & Mohammad  showed how imbalances in training data lead to unintended bias in the resulting models, and, consequently, potentially unfair applications. Shen et al.  substitute words to test the behavior of sentiment analysis algorithms in the presence of stylistic variation, finding that similar word pairs produce significant differences in sentiment score.
Several papers explore richer feedback mechanisms for classification. Some ask annotators to highlight rationales, spans of text indicative of the label [28, 29, 24]. For each document, Zaidan et al.  remove the rationales to generate contrast documents, learning classifiers to distinguish original documents from their contrasting counterparts. While this feedback is easier to collect than ours, how to leverage it for training deep NLP models, where features are not neatly separated, remains less clear.
Lu et al.  programmatically alter text to invert gender bias and combined the original and manipulated data yielding gender-balanced dataset for learning word embeddings. In the simplest experiments, they swap each gendered word for its other-gendered counterpart. For example, the doctor ran because he is late becomes the doctor ran because she is late. However, they do not substitute names even if they co-refer to a gendered pronoun. Building on their work, Zmigrod et al.  describe a data augmentation approach for mitigating gender stereotypes associated with animate nouns for morphologically-rich languages like Spanish and Hebrew. They use a Markov random field to infer how the sentence must be modified while altering the grammatical gender of particular nouns to preserve morpho-syntactic agreement. In contrast, Maudslay et al.  describe a method for probabilistic automatic in-place substitution of gendered words in a corpus. Unlike Lu et al., they propose an explicit treatment of first names by pre-defining name-pairs for swapping, thus expanding Lu et al.’s list of gendered word pairs significantly.
3 Data Collection
We use Amazon’s Mechanical Turk crowdsourcing platform to recruit editors to counterfactually revise each dataset. To ensure high quality of the collected data, we restricted the pool to U.S. residents that had already completed at least HITs and had an over HIT approval rate. For each HIT, we conducted pilot tests to identify appropriate compensation per assignment, receive feedback from workers and revise our instructions accordingly. A total of workers contributed throughout the whole process, of which contributed edits reflected in the final datasets.
The original IMDb dataset consists of reviews divided equally across train and test splits. To keep the task of editing from growing unwieldy, we filter out the longest 20% of reviews, leaving reviews in the train split from which we randomly sample reviews, enforcing a : class balance. Following revision by the crowd workers, we partition this dataset into train/validation/test splits containing , and examples, respectively. We present each review to two workers, instructing to revise the review such that (a) the document remains coherent and (b) the new label (given) accurately describes the revised document. Moreover, we instruct the workers not to make gratuitous modifications.
Over a four week period, we manually inspected each generated review and rejected the ones that were outright wrong (sentiment was still the same or the review was a spam). After review, we rejected roughly of revised reviews. For original reviews, we did not approve any among the counterfactually-revised counterparts supplied by the workers. To construct the new dataset, we chose one revised review (at random) corresponding to each original review. In qualitative analysis, we identified eight common patterns among the edits (Table 2).
For each review, having access to its counterfactually-revised counterpart enables us to isolate which parts the review humans believe are truly indicative of sentiment. These are the parts that were removed, replaced, or inserted into the original review to generate a new review that has the opposite sentiment. We identify the position indices where such replacements or insertions were made and create a binary vector representing the edits in each original review. To analyze inter-editor agreement, we compute the Jaccard similarity between the vectors corresponding to each revised review (Table1). We observe that there is a higher agreement between two workers on smaller reviews and it decreases with the length of the review.
|Number of tokens|
|Types of revisions||Examples|
|Recasting fact as hoped for||The world of Atlantis, hidden beneath the earth’s core, is fantastic
The world of Atlantis, hidden beneath the earth’s core is supposed to be fantastic
|Suggesting sarcasm||thoroughly captivating thriller-drama, taking a deep and realistic view
thoroughly mind numbing “thriller-drama”, taking a “deep” and “realistic” (who are they kidding?) view
|Inserting modifiers||The presentation of simply Atlantis’ landscape and setting
The presentation of Atlantis’ predictable landscape and setting
|Replacing modifiers||“Election” is a highly fascinating and thoroughly captivating thriller-drama
“Election” is a highly expected and thoroughly mind numbing “thriller-drama”
|Inserting phrases||Although there’s hardly any action, the ending is still shocking.
Although there’s hardly any action (or reason to continue watching past 10 minutes), the ending is still shocking.
|Diminishing via qualifiers||which, while usually containing some reminder of harshness, become more and more intriguing.
which, usually containing some reminder of harshness, became only slightly more intriguing.
|Differing perspectives||Granted, not all of the story makes full sense, but the film doesn’t feature any amazing new computer-generated visual effects.
Granted, some of the story makes sense, but the film doesn’t feature any amazing new computer-generated visual effects.
|Changing ratings||one of the worst ever scenes in a sports movie. 3 stars out of 10.
one of the wildest ever scenes in a sports movie. 8 stars out of 10.
Natural Language Inference
Unlike sentiment analysis, SNLI is -way classification task, with inputs consisting of two sentences, a premise and a hypothesis and the three possible labels being entailment, contradiction, and neutral. The label is meant to describe the relationship between the facts stated in each sentence. We randomly sampled , , and pairs from the train, validation, and test sets of SNLI respectively, constraining the new data to have balanced classes. In one HIT, we asked workers to revise the hypothesis while keeping the premise intact, seeking edits corresponding to each of the two counterfactual classes. We refer to this data as Revised Hypothesis (RH). In another HIT, we asked workers to revise the original premise, while leaving the original hypothesis intact, seeking similar edits, calling it Revised Premise (RP).
Following data collection, we employed a different set of workers to verify whether the given label accurately described the relationship between each premise-hypothesis pair. We presented each pair to three workers and performed a majority vote. When all three reviewers were in agreement, we approved or rejected the pair based on their decision, else, we verified the data ourselves. Finally, we only kept premise-hypothesis pairs for which we had valid revised data in both RP and RH, corresponding to both counterfactual labels. As a result, we discarded data. RP and RH, each comprised of pairs in train, in validation, and in test, leading to a total of pairs in train, in validation, and in test in the revised dataset.
We collected all data after IRB approval and measured the time taken to complete each HIT to ensure that all workers were paid more than the federal minimum wage. During our pilot studies, workers spent roughly minutes per revised review, and minutes per revised sentence (for NLI). We paid workers per revision, and per verification, totalling for the study.
We use scikit-learn  implementations of SVMs, Naïve Bayes and Random Forests for sentiment analysis. We train these models on TF-IDF bag of words feature representations of the reviews. We use trees for Random Forests and identify parameters for all classifiers using grid search conducted over the validation set.
When training Bi-LSTMs for sentiment analysis, we restrict the vocabulary to the most frequent tokens, replacing out of vocabulary tokens by UNK. We fix the maximum input length at
tokens and pad smaller reviews. Each token is represented by a randomly-initialized-dimensional embedding. Our model consists of a bidirectional LSTM (hidden size
) with recurrent dropout (probability20] activation (hidden size ), and then a fully-connected output layer with softmax activation. We train all models for a maximum of epochs using Adam , with a learning rate of e- and a batch size of . We apply early stopping when validation loss does not decrease for epochs. We use the architecture described in Poliak et al.  to evaluate hypothesis-only baselines.111https://github.com/azpoliak/hypothesis-only-NLI
We use an off-the-shelf uncased BERT Base model, fine-tuning for each task.222https://github.com/huggingface/pytorch-transformers To account for BERT’s sub-word tokenization, we set the maximum token length is set at for sentiment analysis and for NLI. We fine-tune BERT up to epochs with same early stopping criteria as for Bi-LSTM, using the Adam optimizer with a batch size of (to fit on a Tesla V- GPU). We found learning rates of e- and e- to work best for sentiment analysis and NLI respectively.
5 Experimental Results
We find that for sentiment analysis, linear models trained on the original reviews achieve accuracy when evaluated on original reviews but only (level of random guessing) on revised reviews (Table 3). Linear models trained on revised reviews achieve accuracy on revised reviews but only on the original test set. We see similar pattern for Bi-LSTMs where accuracy drops substantially in both directions. Interestingly, while BERT models suffer drops too, they are less pronounced, perhaps a benefit of the exposure to a larger dataset where the spurious patterns may not have held. Classifiers trained on combined datasets perform well on both, often within pts of models trained on the same amount of data taken only from the original data. Thus, there may be a price to pay for breaking the reliance on spurious associations, but it may not be substantial.
To gain intuition about what is learnable absent the edited spans, we tried training several models on passages where the edited spans have been removed from training set sentences (but not test set). SVM, Naïve Bayes, and Bi-LSTM achieve accuracy, respectively, on this task, suggesting that there is substantial signal in these potentially immaterial sections. However, BERT performs worse than random guessing.
In one simple demonstration of the benefits of our approach, we note that seemingly irrelevant words such as genre names are picked up by linear models trained on either original or revised reviews as top predictors. However, because humans never edit the genre during revision owing to its lack of semantic relevance, combining the original and revised datasets breaks these associations and genre ceases to be predictive of sentiment (Fig 3). Models trained on original data but at the same scale as combined data are able to perform slightly better on the original test set but still fail on the revised reviews. All models trained on original reviews receive a slight boost in accuracy on revised data (except Naïve Bayes), yet their performance significantly worse compared to specialised models. Retraining models on a combination of the original reviews with revised reviews leads to significant increases in accuracy for all models on classifying revised reviews, while slightly improving the accuracy on classifying the original reviews. This underscores the importance of including counterfactually-revised examples in training data.
|Accuracy on original reviews|
|Original & Revised ()|
|Original () & Revised ()|
|Original w/o Edited Spans ()|
|Accuracy on revised reviews|
|Original & Revised ()|
|Original () & Revised ()|
|Train/Eval||Original||RP||RH||RP & RH|
|Revised Premise (RP; )|
|Revised Hypothesis (RH; )|
|RP & RH ()|
|Original w/ RP & RH ()|
|Train/Test||Original||RP||RH||RP & RH|
|RP & RH ()|
|Original w/ RP & RH ()|
Natural Language Inference
Fine-tuned on original sentence pairs, BERT achieves accuracy on SNLI dataset but it is only able to accurately classify sentence pairs from the RP set (Table 4). Fine-tuning BERT on the full SNLI training set ( sentence pairs) results in similar behavior. Fine-tuning it on RP sentence pairs improves its accuracy to on RP but causes a drop of roughly pts on SNLI. On RH sentence pairs, this results in an accuracy of on RH and on SNLI test set but on the RP set. To put these numbers in context, each individual hypothesis sentence in RP is associated with two labels, each in the presence of a different premise. A model that relies on hypotheses only would at best perform slightly better than choosing the majority class when evaluated on this dataset. However, fine-tuning BERT on a combination of RP and RH leads to consistent performance on all datasets as the dataset design forces models to look at both premise and hypothesis. Combining original sentences with RP and RH improves these numbers even further. We compare this with the performance obtained by fine-tuning it on sentence pairs sampled from SNLI training set, and show that while the two perform roughly within pts of each other when evaluated on SNLI, the former outperforms latter on both RP and RH.
To further isolate this effect, Bi-LSTM trained on SNLI hypotheses only achieves accuracy on SNLI test set, which drops to if it is retrained on combination of original, RP and RH data (Table 5). Note that this combined dataset consists of five variants of each original premise-hypothesis pair. Of these five pairs, three consist of the same hypothesis sentence, each associated with different truth value given the respective premise. Using these hypotheses only would provide conflicting feedback to a classifier during training, thus causing the drop in performance. Further, we notice that the gain of the latter over majority class baseline comes primarily from the original data, as the same model retrained only on RP and RH data experiences a further drop of in accuracy, performing worse than just choosing the majority class at all times.
One reasonable concern might be that our models would simply distinguish whether an example were from the original or revised dataset and thereafter treat them differently. The fear might be that our models would exhibit a hypersensitivity (rather than insensitivity) to domain. To test the potential for this behavior, we train several models to distinguish between original and revised data (Table 6). BERT identifies original reviews from revised reviews with accuracy. In case of NLI, BERT and Naïve Bayes perform roughly within pts of the majority class baseline () whereas SVM performs substantially worse.
By leveraging humans not only to provide labels but also to intervene upon the data, revising documents to alter the applicability of various labels, we are able to derive insights about the underlying semantic concepts. Moreover we can leverage the augmented data to train classifiers less dependent on spurious associations. Our study demonstrates the promise of leveraging human-in-the-loop feedback to disentangle the spurious and non-spurious associations, yielding classifiers that hold up better when spurious associations do not transport out of sample. Our methods appear useful on both sentiment analysis and NLI, two contrasting tasks. In sentiment analysis, expressions of opinion matter more than stated facts, while in NLI this is reversed. SNLI poses another challenge in that it is a -class classification task using two input sentences. In future work, we plan to extend these techniques, finding ways to leverage humans in the loop to build more robust systems for question answering and summarization (among others).
The authors are grateful to Amazon AWS for providing GPUs to conduct the experiments, Salesforce Research and Facebook AI for their generous grants that made the data collection possible, Sivaraman Balakrishnan, Shruti Rijhwani, Shruti Palaskar, Aishwarya Kamath, Michael Collins, Rajesh Ranganath and Sanjoy Dasgupta for their valuable feedback, and Tzu-Hsiang Lin for their generous help in creating the data collection platform.
Bowman et al. 
Bowman, S. R., Angeli, G., Potts, C., & Manning, C. D. (2015).
A large annotated corpus for learning natural language inference.
Empirical Methods in Natural Language Processing (EMNLP).
- Chen et al.  Chen, D., Bolton, J., & Manning, C. D. (2016). A thorough examination of the cnn/daily mail reading comprehension task. In Association for Computational Linguistics (ACL).
- Devlin et al.  Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT).
- Dixon et al.  Dixon, L., Li, J., Sorensen, J., Thain, N., & Vasserman, L. (2018). Measuring and mitigating unintended bias in text classification. In AAAI/ACM Conference on AI, Ethics, and Society. ACM.
- Geirhos et al.  Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F. A., & Brendel, W. (2018). Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231.
- Glockner et al.  Glockner, M., Shwartz, V., & Goldberg, Y. (2018). Breaking nli systems with sentences that require simple lexical inferences. In Association for Computational Linguistics (ACL).
- Graves & Schmidhuber  Graves, A., & Schmidhuber, J. (2005). Framewise phoneme classification with bidirectional lstm and other neural network architectures. Neural networks, 18(5-6).
- Gururangan et al.  Gururangan, S., Swayamdipta, S., Levy, O., Schwartz, R., Bowman, S., & Smith, N. A. (2018). Annotation artifacts in natural language inference data. In North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT).
- Hermann et al.  Hermann, K. M., Kocisky, T., Grefenstette, E., Espeholt, L., Kay, W., Suleyman, M., & Blunsom, P. (2015). Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems (NeurIPS).
- Iyyer et al.  Iyyer, M., Wieting, J., Gimpel, K., & Zettlemoyer, L. (2018). Adversarial example generation with syntactically controlled paraphrase networks. In North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT).
- Jia & Liang  Jia, R., & Liang, P. (2017). Adversarial examples for evaluating reading comprehension systems. In Empirical Methods in Natural Language Processing (EMNLP).
- Jo & Bengio  Jo, J., & Bengio, Y. (2017). Measuring the tendency of cnns to learn surface statistical regularities. arXiv preprint arXiv:1711.11561.
- Jurafsky et al.  Jurafsky, D., Chahuneau, V., Routledge, B., & Smith, N. (2014). Narrative framing of consumer sentiment in online restaurant reviews. First Monday, 19(4).
- Kaushik & Lipton  Kaushik, D., & Lipton, Z. C. (2018). How much reading does reading comprehension require? a critical investigation of popular benchmarks. In Empirical Methods in Natural Language Processing (EMNLP).
- Kingma & Ba  Kingma, D. P., & Ba, J. (2015). Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR).
- Kiritchenko & Mohammad  Kiritchenko, S., & Mohammad, S. (2018). Examining gender and race bias in two hundred sentiment analysis systems. In Joint Conference on Lexical and Computational Semantics (*SEM).
- Lu et al.  Lu, K., Mardziel, P., Wu, F., Amancharla, P., & Datta, A. (2018). Gender bias in neural natural language processing. arXiv preprint arXiv:1807.11714.
- Maas et al.  Maas, A. L., Daly, R. E., Pham, P. T., Huang, D., Ng, A. Y., & Potts, C. (2011). Learning word vectors for sentiment analysis. In Association for Computational Linguistics: Human Language Technologies (ACL-HLT).
- Maudslay et al.  Maudslay, R. H., Gonen, H., Cotterell, R., & Teufel, S. (2019). It’s all in the name: Mitigating gender bias with name-based counterfactual data substitution. arXiv preprint arXiv:1909.00871.
- Nair & Hinton  Nair, V., & Hinton, G. E. (2010). Rectified linear units improve restricted boltzmann machines. In International Conference on Machine Learning (ICML).
- Pedregosa et al.  Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., & Duchesnay, E. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research (JMLR), 12.
- Pfeiffer et al.  Pfeiffer, J., Kamath, A., Gurevych, I., & Ruder, S. (2019). What do deep networks like to read? arXiv preprint arXiv:1909.04547.
- Poliak et al.  Poliak, A., Naradowsky, J., Haldar, A., Rudinger, R., & Van Durme, B. (2018). Hypothesis Only Baselines in Natural Language Inference. In Joint Conference on Lexical and Computational Semantics (*Sem).
- Poulis & Dasgupta  Poulis, S., & Dasgupta, S. (2017). Learning with feature feedback: from theory to practice. In Artificial Intelligence and Statistics (AISTATS).
- Ribeiro et al.  Ribeiro, M. T., Singh, S., & Guestrin, C. (2018). Semantically equivalent adversarial rules for debugging nlp models. In Association for Computational Linguistics (ACL).
- Shen et al.  Shen, J. H., Fratamico, L., Rahwan, I., & Rush, A. M. (2018). Darling or babygirl? investigating stylistic bias in sentiment analysis. 5th Workshop on Fairness, Accountability, and Transparency in Machine Learning (FATML)..
- Wallace et al.  Wallace, E., Feng, S., Kandpal, N., Gardner, M., & Singh, S. (2019). Universal adversarial triggers for nlp. arXiv preprint arXiv:1908.07125.
- Zaidan et al.  Zaidan, O., Eisner, J., & Piatko, C. (2007). Using “annotator rationales” to improve machine learning for text categorization. In North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT).
- Zaidan & Eisner  Zaidan, O. F., & Eisner, J. (2008). Modeling annotators: A generative approach to learning from annotator rationales. In Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics.
- Zhao et al.  Zhao, J., Wang, T., Yatskar, M., Ordonez, V., & Chang, K.-W. (2018). Gender bias in coreference resolution: Evaluation and debiasing methods. In North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT).
- Zmigrod et al.  Zmigrod, R., Mielke, S. J., Wallach, H., & Cotterell, R. (2019). Counterfactual data augmentation for mitigating gender stereotypes in languages with rich morphology. In Association for Computational Linguistics (ACL).
Appendix A Appendix
|Revision||Removed words||Inserted words|
|Positive to Negative||movie, film, great, like, good, really, would, see, story, love||movie, film, one, like, bad, would, really, even, story, see|
|Negative to Positive||bad, even, worst, waste, nothing, never, much, would, like, little||great, good, best, even, well, amazing, much, many, watch, better|
|Revision||Removed words||Inserted words|
|Entailment to Neutral||woman, walking, man, blue, sitting, men, girl, standing, looking, running||person, near, child, something, together, people, tall, vehicle, wall, holding|
|Neutral to Entailment||man, street, black, water, little, front, young, playing, woman, two||waiting, couple, playing, running, getting, making, tall, game, black, happily|
|Entailment to Contradiction||blue, people, standing, girl, front, street, red, young, sitting, band||sitting, standing, inside, young, women, child, red, men, sits, one|
|Contradiction to Entailment||sitting, man, walking, black, blue, people, red, standing, white, street||man, sitting, sleeping, woman, sits, eating, playing, park, two, standing|
|Neutral to Contradiction||man, woman, people, boy, black, red, standing, young, two, water||man, woman, boy, men, alone, sitting, girl, dog, three, one|
|Contradiction to Neutral||man, sitting, black, blue, walking, red, standing, street, white, street||man, sitting, woman, people, person, near, something, something, sits, black|
|Entailment to Neutral||man, wearing, white, blue, black, shirt, one, young, people, woman||people, there, playing, man, person, wearing, outside, two, old, near|
|Neutral to Entailment||white, wearing, shirt, black, blue, man, two, standing, young, red||playing, wearing, man, two, there, woman, people, men, near, person|
|Entailment to Contradiction||man, wearing, white, blue, black, two, shirt, one, young, people||people, man, woman, playing, no, inside, person, two, wearing, women|
|Contradiction to Entailment||wearing, blue, black, man, white, two, red, shirt, young, one||people, there, man, two, wearing, playing, people, men, woman, outside|
|Neutral to Contradiction||white, man, wearing, shirt, black, blue, two, standing, woman, red||woman, man, there, playing, two, wearing, one, men, girl, no|
|Contradiction to Neutral||wearing, blue, black, man, white, two, red, sitting, young, standing||people, playing, man, woman, two, wearing, near, tall, men, old|