Abstract: The abductive natural language inference task (Abductive NLI) seeks to select more plausible hypothetical events based on given antecedent events and consequent events. However, inherent biases such as “logical defects” and “single-sentence label leakage” stemming from mediator and confounding variables in the inference process pose challenge. To address these issues, this paper proposes a novel knowledge-aware debiased inference model integrating intervention and counterfactual (KDIC). The model comprises three key modules: the mediator modulation module, the hypothesis-only bias module, and the external knowledge fusion module. Firstly, the mediator modulation module consists of causal graph intervention and hypothesis contrast learning. Causal graph intervention constructs a potential causal graph from given events and then extracts mediator variables, standing for the potential feature of unobserved events, via self-attention mechanism and graph convolutional network for guiding deep encoding. Concurrently, hypothesis contrast learning encourages the model to discern key factors affecting hypothesis judgment, rectifying logical inconsistencies. Secondly, the hypothesis-only bias module addresses the counterfactual problem by proactively identifying the inference biases arising from “single-sentence label leakage”. This module reduces the model’s reliance on specific words or phrases in the hypothesis, thereby enhancing robustness. Finally, this paper leverages a pre-trained common sense knowledge graph encoder, ComET, within the external knowledge fusion module. This integration enriches the model’s understanding of observed events’ motivations and potential outcomes, bolstering logical coherence across events. Experiments results on the αNLI dataset demonstrate that  the inference ability of KDIC is second only to Electra-large-discriminator trained via self-supervised learning. However, KDIC exhibits greater robustness to alleviate biases in the inference process.

Key words: natural language inference, debiased inference, intervention, counterfactual

摘要： 溯因自然语言推理任务旨在根据给定前提事件和结果事件选择更加合理的假设事件。针对推理过程中由中介变量与混杂变量导致的“逻辑缺陷”与“单句标签泄露”偏差问题，提出一种基于干预与反事实原理推理模型（KDIC）。该模型中包含了中介变量调节器、单句假设偏差、外部知识融合三个模块。中介变量调节器由因果图干预与假设对比学习组成。因果图干预旨在基于给定事件构建一个因果图，基于注意力机制和图卷积网络提取中介变量指导深度网络编码，去捕获未观察事件的潜在特征。同时，引入假设对比学习，激励模型主动区分影响假设判断的关键因素并弥补“逻辑缺陷”。基于反事实问题构建单据假设偏差模块，以主动识别“单句标签泄露”带来的推理偏差，减少模型对假设中某些特定词汇或短语的依赖。采用预训练常识知识图谱编码器ComET引入外部知识，确保模型全面理解观察事件发生的动机和可能结果，增强事件间的整体逻辑性。在αNLI数据集上进行了实验，证明KDIC的推理能力仅次于基于自监督训练得到的Electra-large-discriminator，但KDIC具有更强的鲁棒性以缓解推理过程中的偏差。

关键词: 自然语言推理, 去偏推理, 干预, 反事实