Safety classifiers are critical in mitigating toxicity on online forums such as social media and in chatbots. Still, they continue to be vulnerable to emergent, and often innumerable, adversarial attacks.Traditional automated adversarial data generation methods, however, tend to produce attacks that are not diverse, but variations of previously observed harm types.We formalize the task of automated adversarial discovery for safety classifiers - to find new attacks along previously unseen harm dimensions that expose new weaknesses in the classifier.We measure progress on this task along two key axes (1) adversarial success: does the attack fool the classifier? and (2) dimensional diversity: does the attack represent a previously unseen harm type?Our evaluation of existing attack generation methods on the CivilComments toxicity task reveals their limitations: Word perturbation attacks fail to fool classifiers, while prompt-based LLM attacks have more adversarial success, but lack dimensional diversity.Even our best-performing prompt-based method finds new successful attacks on unseen harm dimensions of attacks only 5% of the time.Automatically finding new harmful dimensions of attack is crucial and there is substantial headroom for future research on our new task.

This paper formulates a new task of extracting privacy parameters from a privacy policy, through the lens of Contextual Integrity (CI), an established social theory framework for reasoning about privacy norms. Through extensive experiments, we further show that incorporating CI-based domain-specific knowledge into a BERT-based SRL model results in the highest precision and recall, achieving an F1 score of 84%. With our work, we would like to motivate new research in building NLP applications for the privacy domain.

Counterfactual Data Augmentation (CDA) is a commonly used technique for improving robustness in natural language classifiers. However, one fundamental challenge is how to discover meaningful counterfactuals and efficiently label them, with minimal human labeling cost. Most existing methods either completely rely on human-annotated labels, an expensive process which limits the scale of counterfactual data, or implicitly assume label invariance, which may mislead the model with incorrect labels. In this paper, we present a novel framework that utilizes counterfactual generative models to generate a large number of diverse counterfactuals by actively sampling from regions of uncertainty, and then automatically label them with a learned auxiliary classifier. Our key insight is that we can more correctly label the generated counterfactuals by training a pairwise classifier that interpolates the relationship between the original example and the counterfactual. We demonstrate that with a small amount of human-annotated counterfactual data (10%), we can generate a counterfactual augmentation dataset with learned labels, that provides an 18-20% improvement in robustness and a 14-21% reduction in errors on 6 out-of-domain datasets, comparable to that of a fully human-annotated counterfactual dataset for both sentiment classification and question paraphrase tasks.

Developing robust NLP models that perform well on many, even small, slices of data is a significant but important challenge, with implications from fairness to general reliability. To this end, recent research has explored how models rely on spurious correlations, and how counterfactual data augmentation (CDA) can mitigate such issues. In this paper we study how and why modeling counterfactuals over multiple attributes can go significantly further in improving model performance. We propose RDI, a context-aware methodology which takes into account the impact of secondary attributes on the model’s predictions and increases sensitivity for secondary attributes over reweighted counterfactually augmented data. By implementing RDI in the context of toxicity detection, we find that accounting for secondary attributes can significantly improve robustness, with improvements in sliced accuracy on the original dataset up to 7% compared to existing robustness methods. We also demonstrate that RDI generalizes to the coreference resolution task and provide guidelines to extend this to other tasks.

Learning contextual text embeddings that represent causal graphs has been useful in improving the performance of downstream tasks like causal treatment effect estimation. However, existing causal embeddings which are trained to predict direct causal links, fail to capture other indirect causal links of the graph, thus leading to spurious correlations in downstream tasks. In this paper, we define the faithfulness property of contextual embeddings to capture geometric distance-based properties of directed acyclic causal graphs. By incorporating these faithfulness properties, we learn text embeddings that are 31.3% more faithful to human validated causal graphs with about 800K and 200K causal links and achieve 21.1% better Precision-Recall AUC in a link prediction fine-tuning task. Further, in a crowdsourced causal question-answering task on Yahoo! Answers with questions of the form “What causes X?”, our faithful embeddings achieved a precision of the first ranked answer (P@1) of 41.07%, outperforming the existing baseline by 10.2%.

We propose a new framework to uncover the relationship between news events and real world phenomena. We present the Predictive Causal Graph (PCG) which allows to detect latent relationships between events mentioned in news streams. This graph is constructed by measuring how the occurrence of a word in the news influences the occurrence of another (set of) word(s) in the future. We show that PCG can be used to extract latent features from news streams, outperforming other graph-based methods in prediction error of 10 stock price time series for 12 months. We then extended PCG to be applicable for longer time windows by allowing time-varying factors, leading to stock price prediction error rates between 1.5% and 5% for about 4 years. We then manually validated PCG, finding that 67% of the causation semantic frame arguments present in the news corpus were directly connected in the PCG, the remaining being connected through a semantically relevant intermediate node.

We describe our experiences in using an open domain question answering model (Chen et al., 2017) to evaluate an out-of-domain QA task of assisting in analyzing privacy policies of companies. Specifically, Relevant CI Parameters Extractor (RECIPE) seeks to answer questions posed by the theory of contextual integrity (CI) regarding the information flows described in the privacy statements. These questions have a simple syntactic structure and the answers are factoids or descriptive in nature. The model achieved an F1 score of 72.33, but we noticed that combining the results of this model with a neural dependency parser based approach yields a significantly higher F1 score of 92.35 compared to manual annotations. This indicates that future work which in-corporates signals from parsing like NLP tasks more explicitly can generalize better on out-of-domain tasks.

In this paper, we propose a new unsupervised learning framework to use news events for predicting trends in stock prices. We present Word Influencer Networks (WIN), a graph framework to extract longitudinal temporal relationships between any pair of informative words from news streams. Using the temporal occurrence of words, WIN measures how the appearance of one word in a news stream influences the emergence of another set of words in the future. The latent word-word influencer relationships in WIN are the building blocks for causal reasoning and predictive modeling. We demonstrate the efficacy of WIN by using it for unsupervised extraction of latent features for stock price prediction and obtain 2 orders lower prediction error compared to a similar causal graph based method. WIN discovered influencer links from seemingly unrelated words from topics like politics to finance. WIN also validated 67% of the causal evidence found manually in the text through a direct edge and the rest 33% through a path of length 2.