Mitigating Spurious Correlation in Natural Language Understanding with Counterfactual Inference

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11308 - 11321 December 7-11, 2022 ©2022 Association for Computational Linguistics Mitigating Spurious Correlation in Natural Language Understanding with Counterfactual Inference Can Udomcharoenchaikit†, Wuttikorn Ponwitayarat†, Patomporn Payoungkhamdee†, Kanruethai Masuk‡, Weerayut Buaphet†, Ekapol Chuangsuwanich♣, Sarana Nutanong† †School of Information Science and Technology, VISTEC, Thailand ‡VISAI AI, Thailand ♣Department of Computer Engineering, Chulalongkorn University, Thailand {canu_pro,wuttikornp.p_s22,patomporn.p_s21,kanruethaim_pro, weerayut.b_s20,snutanon}@vistec.ac.th, ekapolc@cp.eng.chula.ac.th Abstract Despite their promising results on standard benchmarks, NLU models are still prone to make predictions based on shortcuts caused by unintended bias in the dataset. For example, an NLI model may use lexical overlap as a shortcut to make entailment predictions due to repetitive data generation patterns from anno- tators, also called annotation artifacts. In this paper, we propose a causal analysis framework to help debias NLU models. We show that (1) by defining causal relationships, we can intro- spect how much annotation artifacts affect the outcomes. (2) We can utilize counterfactual inference to mitigate bias with this knowledge. We found that viewing a model as a treatment can mitigate bias more effectively than viewing annotation artifacts as treatment. (3) In addi- tion to bias mitigation, we can interpret how much each debiasing strategy is affected by annotation artifacts. Our experimental results show that using counterfactual inference can improve out-of-distribution performance in all settings while maintaining high in-distribution performance. 1 1 Introduction Mitigating spurious correlations is crucial to the robustness of any learning method. Although deep learning models can perform well on con- ventional natural language understanding (NLU) benchmarks, researchers have found that these models leverage superficial patterns to produce cor- rect predictions rather than learning the underlying tasks (Gururangan et al., 2018; McCoy et al., 2019). As a result, these models perform poorly when ap- plied on out-of-distribution datasets, particularly on challenge sets that highlight models’ reliance on spurious patterns by testing them on counterexam- ples. For example, McCoy et al. (2019) show that 1The work was performed while Kanruethai Masuk was an intern at VISTEC. The code is available at https://github. com/c4n/debias_nlu. ANNOTATION PROCESS When creating an input sample for the entailment class, an annotator tends to use the same words in the premise to write the hypothesis. This is an example of lexical overlap annotation artifact. P H The singer walked to the theater. The singer walked. The singer walked to the theater. P Entailment Class CAUSAL GRAPH P H The nurse near the singer walked. The singer walked. We use a causal graph to represent how an annotation artifact affects the prediction directly and how an annotation artifact affects a training input, which in turn affects the prediction. Annotation Artifact + - E C N Annotator Main Model Bias Model Probability Answer COUNTERFACTUAL INFERENCE P = Premise H = Hypothesis Bias Model Main Model E C N Input: Text - E C N Input: Biased Features +e.g., word overlap - = Debiased Answer E C N ×Prediction: E ×Prediction: E Prediction: N ✓ E = Entailment …

Content truncated to 500 words