Robustness to Spurious Correlations in Text Classification via Automatically Generated Counterfactuals

Despite the remarkable performance machine learning models have achieved in various tasks, studies have shown that statistical models typically learn correlational associations between features and classes, and model validity and reliability are threatened by spurious correlations. Examples include: a sentiment classifier learns that “Spielberg” is correlated with positive movie reviews (Wang and Culotta 2020); a toxicity classifier learns that “gay” is correlated with toxic comments (Wulczyn, Thain, and Dixon 2017); a medical system learns that the disease is associated with patient ID (Kaufman, Rosset, and Perlich 2011); an object detection system recognizes a sheep based on the grass in the background (Ghorbani, Abid, and Zou 2019). If these kinds of spurious correlations are built into the model during training time, the model could fail when test data has a different distribution or even on samples with minor changes, and the predictions will be problematic and suffer from algorithm fairness or trust issues.

One solution to achieve robustness is to learn causal associations between features and classes. E.g., in the sentence “This was a free book that sounded boring to me”, the word most responsible for the label being negative is “boring” instead of “free”. Identifying causal associations provides a way to build more robust and generalizable models.

Recent works try to achieve robustness with the aid of human-in-the-loop systems. Srivastava, Hashimoto, and Liang (2020) present a framework to make models robust to spurious correlations by leveraging human common sense of causality. They augment training data with crowd-sourced annotations about reasoning of possible shifts in unmeasured variables and finally conduct robust optimization to control worst-case loss. Similarly, Kaushik, Hovy, and Lipton (2020) ask humans to revise documents with minimal edits to change the class label, then augment the original training data with the counterfactual samples. Results show that the robust classifier is less sensitive to spurious correlations. While these prior works show the potential of using human annotations to improve model robustness, collecting such annotations can be costly.

In this paper, we propose to train a robust classifier with automatically generated counterfactual samples. Specifically, we first identify likely causal features using the closest opposite matching approach and then generate counterfactual training samples by substituting causal features with their antonyms and assigning opposite labels to the newly generated samples. Finally, we combine the original training data with counterfactual data to train a more robust classifier.

We experiment with sentiment classification tasks on two datasets (IMDB movie reviews and Amazon kindle reviews). For each dataset, we have the original training data and testing data, and additional human-generated counterfactual testing data. We first train a traditional classifier using the original data, which performs poorly on the counterfactual testing data (i.e., 10%-37% drop in accuracy). Then, we train a robust classifier with the combination of original training data and automatically-generated counterfactual training data, and it performs well on both the original testing data and the counterfactual testing data (i.e., 12% - 25% absolute improvement over the baseline). Additionally, we consider limited human supervision in the form of human-provided causal features, which we then use to generate counterfactual training samples. We find that a small number of causal features (e.g., 50) results in accuracy that is comparable to a model trained with $1.7K$ human-generated counterfactual training samples from previous work.

Spurious correlations are problematic and could be introduced in many ways. Sagawa et al. (2020) investigate how overparameterization exacerbates spurious correlations. They compare overparameterized models with underparameterized models and show that overparameterization encodes spurious correlations that do not hold in worst-group data. Kiritchenko and Mohammad (2018) showed that training data imbalances can lead to unintended bias and unfair applications (e.g., bias towards gender, race). Besides that, data leakage (Roemmele, Bejan, and Gordon 2011) and distribution shift between training data and testing data (Quionero-Candela et al. 2009) are particularly challenging and hard to detect as they introduce spurious correlations during model training and hurt model performance when deployed. Another new type of threat is backdoor attack (Dai, Chen, and Li 2019), where an attacker intentionally poisons a model by injecting spurious correlations into training data and manipulating model performance by specific triggers.

A growing line of research explores the challenges and benefits of using causal inference to improve model robustness. Wood-Doughty, Shpitser, and Dredze (2018) uses text classifiers in causal analyses to address issues of missing data and measurement error. Keith, Jensen, and O’Connor (2020) introduce methods to remove confounding from causal estimates. Paul (2017) proposes a propensity score matching method to learn meaningful causal associations between variables. Jia et al. (2019) consider label preserving transformations to improve model robustness to adversarial perturbations with Interval Bound Propagation.  Landeiro and Culotta (2018) address the issue of spurious correlations by doing back-door adjustment to control for known confounders. Wang and Culotta (2020) train a classifier to distinguish between spurious features and genuine features, and gradually remove spurious features to improve worst-case accuracy of minority groups.

Recent works investigate how additional human supervision can reduce spurious correlations and improve model robustness. Roemmele, Bejan, and Gordon (2011) and Sap et al. (2018) show that humans achieve high performance on commonsense causal reasoning and counterfactual tasks. Zaidan and Eisner (2008) ask annotators to provide rationales as hints to guide classifiers paying attention to relevant features. Lu et al. (2018) and Zmigrod et al. (2019) use counterfactual data augmentation to mitigate bias. Ribeiro et al. (2020) evaluate model robustness using generated counterfactuals that requires significant human intervention (either by specifying substitution terms or generating templates and labeling examples). Garg et al. (2019) presume a predefined set of 50 counterfactually fair tokens and augment the training data with counterfactuals to improve toxicity classifier fairness.

While recent works have proposed the idea of generating and augmenting with counterfactuals for robust classifications, the main contributions of this paper are as follows:

• •

  We propose to discover likely causal features using statistical matching techniques.

• •

  Using these features, we automatically generate counterfactual samples by substituting causal features with antonyms, which significantly reduces human effort.

• •

  We conduct experiments demonstrating the improved robustness of the resulting classifier to spurious features.

• •

  We conduct additional analyses to show how the robust classifier increases the importance of causal features and decreases the importance of spurious features.

To train a classification model, we fit a function $f(\cdot)$ with a set of labeled data and learn a map between input features and output labels. We consider a binary text classification task with the simple approach of logistic regression model¹¹1Our approach is model agnostic. We focus on logistic regression for interpretability and clarity.: $f(x;\theta)=\frac{1}{1+e^{-\langle x,\theta\rangle}}$ using bag-of-words features. Specifically, each document is a sequence of words $d=\langle w_{1}\ldots w_{k}\rangle$ that is transformed into a feature vector $x$ via one-hot representation $x=\langle x_{1}\ldots x_{V}\rangle$ ($V$ is the vocabulary size), and has a binary label $y\in\{-1,1\}$. The model is fit on a set of labeled documents $\mathcal{D}=\{(d_{1},y_{1})\ldots(d_{n},y_{n})\}$, and parameters are estimated by minimizing the loss function $\mathcal{L}$: $\theta^{*}\leftarrow\arg\min_{\theta}\mathcal{L}(\cal{D},\theta)$. We can examine the (partial) correlations between features and labels by model coefficients.

Spurious correlations are very common in statistical models and they could mislead classifiers. For example, in our experimental dataset of Amazon kindle reviews, the classifier learns that “free” has a strong correlation with negative sentiment because “free” has a high frequency in negative book reviews (e.g., “This was a free book that sounded boring to me”), and thus the classifier makes errors when predicting positive documents that contain “free”.

Previous works have tried various methods to reduce spurious correlations (e.g., regularization, feature selection, back-door adjustment (Hoerl and Kennard 1970; Forman 2003; Landeiro and Culotta 2018)). However, a more direct solution is to learn meaningful causal associations between features and classes. While expressing causality in the context of text classification can be challenging, we follow the previous work (Paul 2017) to operationalize the definition of a causal feature as follows: term $w$ is a causal feature in document $d$ if, all else being equal, one would expect $w$ to be a determining factor in assigning a label to $d$. For example, in the sentence “This was a free book that sounded boring to me”, “boring” is primarily responsible for the negative sentiment. In contrast, the term “free” itself does not convey negative sentiment. We consider “boring” as a causal term and “free” as a non-causal term (term refers to word feature).

Our approach in this paper is to first identify such causal features and then use them to automatically generate counterfactual training samples. Specifically, for a sample $(d,y)$, we get the corresponding counterfactual sample $(d^{\prime},y^{\prime})$ by (i) substituting causal terms in $d$ with their antonyms to get $d^{\prime}$, and (ii) assigning an opposite label $y^{\prime}$ to $d^{\prime}$. Let’s consider the previous example to see how augmenting with counterfactual samples might work. Traditional classifiers trained on original data learns that “free” is correlated with the negative class due to its high frequency in negative book reviews. For every negative document containing “free”, we generate one corresponding counterfactual document. The counterfactual sample for “This was a free book that sounded boring to me”(neg) would be “This was a free book that sounded interesting to me”(pos). When augmenting the original training data with counterfactual data, “free” would get equal frequency in both classes for the ideal case (i.e, if we could generate counterfactual samples for all documents containing “free”). Thus, a classifier fit on the combined dataset should have a reduced coefficient for “free” and increased coefficients for “boring” and “interesting.”

Our approach is a two-stage process: we first identify likely causal features and then generate counterfactual training data using causal features. To identify causal features, we consider the counterfactual framework of causal inference (Winship and Morgan 1999). If word $w$ in document $d$ were replaced with some other word $w^{\prime}$, how likely is it that the label $y$ would change? Since conducting randomized control trials to answer this question is infeasible, we instead use matching methods (Imbens 2004; King and Nielsen 2019). The intuition is as follows: if $w$ is a reliable piece of evidence to determine the label of $d$, we should be able to find a very similar document $d^{\prime}$ that (i) does not contain $w$, and (ii) has the opposite label of $d$. For example, $(d,y)$ = (“This was a free book that sounded boring to me”, neg) and $(d^{\prime},y^{\prime})$ = (“This was a free book that sounded interesting to me”, pos) would be an ideal match where substituting causal term “boring” with another term “interesting” flips the label. While this is not a necessary condition of a causal feature (there may not be a good match in a limited training set), in the experiments below we find this to be a fairly precise approach to generate a small number of high-quality causal features.

The full steps of our approach are as follows:

1. 1.

   We first train an initial classifier and extract strongly correlated terms $\langle t_{1}\ldots t_{k}\rangle$ as candidate causal features. E.g., for logistic regression model, we would extract features with high magnitude coefficients. For more complex models, other transparency algorithms may be used (Martens and Provost 2014).

2. 2.

   For each top term $t$ and a set of documents containing $t$: $D_{t}=\langle d_{1}\ldots d_{n}\rangle$, we search for a set of matched documents $D^{\prime}_{t}=\langle d_{1}^{\prime}\ldots d_{n}^{\prime}\rangle$ and get ${D_{match}}=\{(d_{1},d_{1}^{\prime},score_{1})\ldots(d_{n},d_{n}^{\prime},score_{n}^{\prime})\}$, where the score for each match is the context similarity of $d_{i}$ and $d_{i}^{\prime}$. The matched documents have opposite labels.

3. 3.

   Then for each term $t$ and its corresponding matching set $D_{match}$, we pick the tuple $(d_{i},d_{i}^{\prime},score_{i})$ that has the highest similarity score as the closest opposite match. We then identify likely causal features by picking those whose closest opposite matches have scores greater than a threshold (0.95 is used below).

4. 4.

   We use PyDictionary²²2https://github.com/geekpradd/PyDictionary to get antonyms for causal terms.

5. 5.

   For each training sample, we generate its counterfactual sample by substituting causal terms with antonyms and assigning an opposite label to the counterfactual sample.

6. 6.

   Finally, we train a robust classifier using the combination of original training data and counterfactual data.

   We provide more details on these steps below.

We perform sentiment classification experiments on the following two datasets.³³3Code and data available at: https://github.com/tapilab/aaai-2021-counterfactuals Each dataset has human-generated counterfactual testing samples to provide benchmark performance for classifier robustness.

IMDB movie reviews: This dataset is sampled from the original IMDB dataset (Pang and Lee 2005) and the conterfactual part is collected and published by Kaushik, Hovy, and Lipton (2020). They randomly sampled $2.5K$ reviews with balanced class distributions and partition them into 1707 training, 245 validation, and 488 testing samples. Then they instruct Amazon Mechanical Turk workers to revise each document with minimum changes towards a counterfactual label, and finally collected $2.5K$ counterfactually-manipulated samples.

Each document of this dataset is a long paragraph. We are both interested in exploring classifier performance for long texts and short texts. So, we additionally create a version of this dataset segmented into single sentences. To do so, we first fit a binary classifier on the original data and identify strongly correlated terms as keywords. Then we split each original document into single sentences and keep those containing at least one keyword. Sentence labels are inherited from the original document labels. To justify the validity of this approach, we randomly sampled 500 sentences and manually checked their labels. The inherited labels were correct for 484 sentences (i.e., 96.8% accuracy). We differentiate the IMDB dataset with long texts as IMDB-L and short texts as IMDB-S.

Amazon Kindle reviews (Kindle): This dataset contains book reviews from the Amazon Kindle Store and each review has a rating ranges from 1-5 (He and McAuley 2016). We label reviews with ratings {4,5} as positive and reviews with ratings {1,2} as negative, and then process this dataset to be single sentences following the approach used in IMDB.

Human edited counterfactuals: For the IMDB dataset, we have the human-generated counterfactual training data and counterfactual testing data. For kindle dataset, we randomly select 500 samples as test data (comparable size with the test data from IMDB-L) and manually edit them to be counterfactual samples with the minimum edits.

Ground truth causal terms: We manually annotated a set of ground truth causal terms for each dataset. Specifically, we asked two student annotators to label a term as causal if, all else being equal, this term is a determining factor in assigning a label to a document. While there is some subjectivity in the annotation, we did a round of training to resolve disagreements prior to annotation and the final agreement was generally high for this task (e.g., 96% raw agreement by fraction of labels that agree).

Table 3 shows the basic data statistics. For the top terms, we select them by thresholding on the magnitude of coefficients. For IMDB-L, we use threshold 0.4, and for IMDB-S and Kindle, we use threshold 1.0.

We have presented a framework to automatically generate counterfactual training samples from causal terms and then train a robust classifier using the combination of original data and counterfactual data. Using this framework, we can easily improve classifier robustness even with few causal terms. If enough causal terms are annotated (e.g., 100 in our experiments), it is possible to achieve performance comparable to using human-generated counterfactuals. In future work, we will investigate extensions to increase the precision and recall of causal term identification to further reduce the reliance on human supervision. Additionally, it would be interesting to extend this framework to other tasks such as topic classification robustness. To do so, we would need to generalize the notion of “antonyms” to include terms that indicate a different topic (e.g., to convert a sports news story to a political news story, we might change the sentence “watch the game” to “watch the debate”). Then we could generate “counterfactuals” by substituting topic-related terms with terms that are not semantically related to the current topic (or related to other topics).

This research was funded in part by the National Science Foundation under grant #1618244. Zhao Wang was funded in part by a Dissertation Fellowship from the Computer Science department at Illinois Tech. We would also like to thank the anonymous reviewers for useful feedback.