ReCo: Reliable Causal Chain Reasoning via Structural Causal Recurrent Neural Networks

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6426 - 6438 December 7-11, 2022 ©2022 Association for Computational Linguistics ReCo: Reliable Causal Chain Reasoning via Structural Causal Recurrent Neural Networks Kai Xiong1∗ Xiao Ding1† Zhongyang Li2 Li Du1 Ting Liu1 Bing Qin1 Yi Zheng2 Baoxing Huai2 1Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China 2Huawei Cloud, China {kxiong, xding, ldu, tliu, qinb}@ir.hit.edu.cn {lizhongyang6, zhengyi29, huaibaoxing}@huawei.com Abstract Causal chain reasoning (CCR) is an essen- tial ability for many decision-making AI sys- tems, which requires the model to build reli- able causal chains by connecting causal pairs. However, CCR suffers from two main transi- tive problems: threshold effect and scene drift. In other words, the causal pairs to be spliced may have a conflicting threshold boundary or scenario. To address these issues, we propose a novel Reliable Causal chain reasoning frame- work (ReCo), which introduces exogenous vari- ables to represent the threshold and scene fac- tors of each causal pair within the causal chain, and estimates the threshold and scene contradic- tions across exogenous variables via structural causal recurrent neural networks (SRNN). Ex- periments show that ReCo outperforms a series of strong baselines on both Chinese and En- glish CCR datasets. Moreover, by injecting reli- able causal chain knowledge distilled by ReCo, BERT can achieve better performances on four downstream causal-related tasks than BERT models enhanced by other kinds of knowledge. 1 Introduction Causal chain reasoning aims at understanding the long-distance causal dependencies of events and building reliable causal chains. Here, reliable means that events in the causal chain can naturally occur in the order of causal evolution within some circumstance based on the commonsense (Roem- mele et al., 2011). Causal chain knowledge is of great importance for various artificial intelligence applications, such as question answering (Asai et al., 2019), and abductive reasoning (Du et al., 2021a). Many studies focus on the reliability of causal pair knowledge but ignore that of causal chain knowledge, especially in the natural language processing (NLP) community. †Corresponding Author *This work was conducted during the internship of Kai Xiong at Huawei Cloud (a) Threshold effect problem in causal chain. (b) Scene drift problem in causal chain. ✘ swimming' in'the'sea salt'water' intake dehydration 🏊 💧 💦 😵💫‍ salt'water' intake Swimming'in'the'sea'causes'a% few%sips%of'saltwater'intake A%large%bottle%of'salt'water' intake'can'cause'dehydration swimming' in'the'sea salt'water' intake dehydration Causal Chain: Causal Pairs: Contexts: A B B C A B C playing'Dota dispute gets a red card dispute 🎮 playing Dota might result in a dispute ⚽ David&Silva&gets'a'red&card&for' the'dispute'with'the'referee playing'Dota dispute gets a red card A B B C A B C Causal Chain: Causal Pairs: Contexts: ✘ Figure 1: Causal chains with (a) threshold effect and (b) scene drift problems, which can be estimated by the contradictions of threshold and scene factors in the contexts, respectively. Previous works mainly acquire causal chain knowledge by first extracting precise causal pairs from text with rule-based (Heindorf et al., 2020; Li et al., 2020) or neural-based (Ding et al., 2019; Zhang …

Content truncated to 500 words