C2L: Causally Contrastive Learning for Robust Text Classification

Abstract

Despite the super-human accuracy of recent deep models in NLP tasks, their robustness is reportedly limited due to their reliance on spurious patterns. We thus aim to leverage contrastive learning and counterfactual augmentation for robustness. For augmentation, existing work either requires humans to add counterfactuals to the dataset or machines to automatically matches near-counterfactuals already in the dataset. Unlike existing augmentation is affected by spurious correlations, ours, by synthesizing “a set” of counterfactuals, and making a collective decision on the distribution of predictions on this set, can robustly supervise the causality of each term. Our empirical results show that our approach, by collective decisions, is less sensitive to task model bias of attribution-based synthesis, and thus achieves significant improvements, in diverse dimensions: 1) counterfactual robustness, 2) cross-domain generalization, and 3) generalization from scarce data.

LaNiO3 as a Novel Anode for Lithium-Ion Batteries

“…Some work also introduces counterfactual and causal ideas. Among the existing counterfactual contrastive learning methods, C2l (Choi et al, 2022) is similar to ours. It synthesizes "a set" of counterfactuals and makes a collective decision on the distribution of predictions on this set, which can robustly supervise the causality of each term.…”

Graph topology enhancement for text classification

“…Therefore, the human-in-the-loop process is designed to take advantage of human knowledge to modify text and obtain opposite labels for counterfactual augmentation [15]. But due to the high cost of human labor, many methods of automatic counterfactual augmentation have also been developed [14], [16], [17]. Edits against auto-mined causal features are used to obtain counterfactual samples.…”

“…[16] learns the rules by logical reasoning and gives faithful counterfactual predictions. C2L make a collective decision based on a set of counterfactuals to overcome shortcut learning [17]. AutoCAD guides controllable generative models to automatically generate counterfactual data [58].…”

“…It uses tokens with high LPLMs attention scores as possible shortcuts. C2L [17], which monitors the causality of each word collectively through a set of automatically generated counterfactual samples and uses contrastive learning to improve the robustness of the model.…”

A Study on the Perception of SNS Users on Artificial Intelligence Design Using Text Mining

“…This is because auto-tagging has a potentially very large label space, ranging from subject topics to knowledge components (KC) (Zhang et al, 2015;Koedinger et al, 2012;Mohania et al, 2021;Viswanathan et al, 2022). The resulting data scarcity decreases performance on rare labels during training (Chalkidis et al, 2020;Lu et al, 2020;Snell et al, 2017;Choi et al, 2022).…”