基于反事实推理的事实验证去偏方法

Counterfactual Inference for Fact Verification Debiasing

事实验证是一项具有挑战性的任务,旨在使用来自可信赖语料库的多个证据句子来验证声明。为了促进研究,一些事实验证数据集被提出,极大地加速了事实验证技术的发展。然而,现有的事实验证数据集通常采用众包的方法构造,无可避免地引入偏差。已有事实验证去偏工作大致可以分为基于数据增强的方法和基于权重正则化的方法,前者不灵活,后者依赖于训练阶段的不确定输出。与已有工作不同,该文从因果关系出发,提出基于反事实推理的事实验证去偏方法。该文首先设计事实验证中的因果图,建模声明、证据以及它们之间的交互和预测结果的因果关系。接着,根据因果图提出事实验证去偏方法,通过总间接效应去除声明带来的偏差影响。我们使用多任务学习的方式来训练模型。训练时,该文采用多任务学习的方式建模各个因素的影响,同时在有偏和无偏测试集上评估模型的性能。实验结果表明,对比基准方法,该文模型在性能上获得了一致的提升。

Fact verification is a challenging task with the purpose to verify a claim using multiple evidence from a trustworthy corpus.In contrast to the existing data-augmentation-based methods and weight-regularization-based methods,we propose a debias model for fact verification based on counterfactual inference from the perspective of causality.Specifically,we first design the cause graph for fact verification,modeling the claim,evidence,interactions between them and the cause-effect relationship of the predicted results.Then,we propose a debias method based on the cause graph to remove the bias effect caused by the claim through the total indirect effect.During training,we use multi-task learning to model the influence of various factors.Experiments on both biased and unbiased test datasets demonstrate that our model can achieve consistent performance improvement compared with baselines.