While recent works have been considerably improving the quality of the natural language explanations (NLEs) generated by a model to justify its predictions, there is very limited research in detecting and alleviating inconsistencies among generated NLEs. In this work, we leverage external knowledge bases to significantly improve on an existing adversarial attack for detecting inconsistent NLEs. We apply our attack to high-performing NLE models and show that models with higher NLE quality do not necessarily generate fewer inconsistencies. Moreover, we propose an off-the-shelf mitigation method to alleviate inconsistencies by grounding the model into external background knowledge. Our method decreases the inconsistencies of previous high-performing NLE models as detected by our attack.

Natural language free-text explanation generation is an efficient approach to train explainable language processing models for commonsense-knowledge-requiring tasks. The most predominant form of these models is the explain-then-predict (EtP) structure, which first generates explanations and uses them for making decisions. The performance of EtP models is highly dependent on that of the explainer by the nature of their structure. Therefore, large-sized explanation data are required to train a good explainer model. However, annotating explanations is expensive. Also, recent works reveal that free-text explanations might not convey sufficient information for decision making. These facts cast doubts on the effectiveness of EtP models. In this paper, we argue that the predict-then-explain (PtE) architecture is a more efficient approach in terms of the modelling perspective. Our main contribution is twofold. First, we show that the PtE structure is the most data-efficient approach when explanation data are lacking. Second, we reveal that the PtE structure is always more training-efficient than the EtP structure. We also provide experimental results that confirm the theoretical advantages.