Causal MRC: Mitigating Position Bias Based on Causal Graph

Jiazheng Zhu, Linjuan Wu, Shaojuan Wu, Xiaowang Zhang, Yuexian Hou, Zhiyong Feng

Abstract: Extractive Machine Reading Comprehension (MRC) requires models to obtain start and end positions of answers from a given passage. MRC models may tend to rely on position bias as a shortcut, and thus they fail to learn the multi-source knowledge from both passages and questions sufficiently. Recent debiasing methods proposed to exclude the position prior during inference. However, they cannot distinguish the “good” position context and “bad” position bias from the whole prior. In this paper, we propose a novel MRC framework CausalMRC based on causal graph to mitigate position bias. Motivated by causal inference, we design a causal graph for MRC to formulate the position bias as the direct causal effect of passages on answers. Specifically, we mitigate the position bias by subtracting the direct position effect from the total causal effect. Experiments demonstrate that our proposed CausalMRC achieves competitive performance on the biased SQuAD dataset while performing robustly on the original SQuAD.