ExCAR: Event Graph Knowledge Enhanced Explainable Causal Reasoning

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 2354–2363 August 1–6, 2021. ©2021 Association for Computational Linguistics 2354 ExCAR: Event Graph Knowledge Enhanced Explainable Causal Reasoning Li Du, Xiao Ding∗, Kai Xiong, Ting Liu, and Bing Qin Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China {ldu, xding, kxiong, tliu,qinb}@ir.hit.edu.cn Abstract Prior work infers the causation between events mainly based on the knowledge induced from the annotated causal event pairs. However, ad- ditional evidence information intermediate to the cause and effect remains unexploited. By incorporating such information, the logical law behind the causality can be unveiled, and the interpretability and stability of the causal rea- soning system can be improved. To facilitate this, we present an Event graph knowledge en- hanced explainable CAusal Reasoning frame- work (ExCAR). ExCAR first acquires addi- tional evidence information from a large-scale causal event graph as logical rules for causal reasoning. To learn the conditional probabilis- tic of logical rules, we propose the Conditional Markov Neural Logic Network (CMNLN) that combines the representation learning and struc- ture learning of logical rules in an end-to-end differentiable manner. Experimental results demonstrate that ExCAR outperforms previ- ous state-of-the-art methods. Adversarial eval- uation shows the improved stability of Ex- CAR over baseline systems. Human evaluation shows that ExCAR can achieve a promising explainable performance. 1 Introduction Causal reasoning aims at understanding the gen- eral causal dependency between the cause and ef- fect (Luo et al., 2016). Causality is commonly ex- pressed by humans in the text of natural language, and is of great value for various Artificial Intelli- gence applications, such as question answering (Oh et al., 2013), event prediction (Li et al., 2018), and decision making (Sun et al., 2018). Previous work mainly learns causal knowledge from manually annotated causal event pairs, and achieves promising performances (Luo et al., 2016; ∗Corresponding author Figure 1: (a) Without the evidence event i, we can hardly reveal the implicit causation between a and b. (b) The absent of evidence events may restrict the performance of event-pair based methods. (c) Given an event pair, the ExCAR framework obtains evidence events from an event graph and conducts causal reasoning using the additional evidence events. (d) The causal strength (cs) of the same rule can vary with different antecedents. We define such phenomenon as superimposed causal effect. Xie and Mu, 2019a; Li et al., 2019). However, re- cent works have questioned the seemingly superb performance for some of these studies (McCoy et al., 2019; Poliak et al., 2018; Gururangan et al., 2018). Specifically, training data may contain ex- ploitable superficial cues that are correlative of the expected output. The main concern is that these works have not learned the underlying mechanism of causation so that their inference models are not stable enough and their results are not explainable. While we notice that there is plentiful evidence information outside the given corpus that can pro- vide …

Content truncated to 500 words