COLA: Contextualized Commonsense Causal Reasoning from the Causal Inference Perspective

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5253–5271 July 9-14, 2023 ©2023 Association for Computational Linguistics COLA: Contextualized Commonsense Causal Reasoning from the Causal Inference Perspective Zhaowei Wang1, Quyet V. Do1, Hongming Zhang2, Jiayao Zhang3, Weiqi Wang1, Tianqing Fang1, Yangqiu Song1, Ginny Y. Wong4, & Simon See4 1Department of Computer Science and Engineering, HKUST 2Tencent AI Lab, Bellevue, USA 3University of Pennsylvania 4NVIDIA AI Technology Center (NVAITC), NVIDIA, Santa Clara, USA {zwanggy,yqsong}@cse.ust.hk, {gwong,ssee}@nvidia.com Abstract Detecting commonsense causal relations (cau- sation) between events has long been an es- sential yet challenging task. Given that events are complicated, an event may have different causes under various contexts. Thus, exploit- ing context plays an essential role in detecting causal relations. Meanwhile, previous works about commonsense causation only consider two events and ignore their context, simplifying the task formulation. This paper proposes a new task to detect commonsense causation between two events in an event sequence (i.e., context), called contextualized commonsense causal rea- soning. We also design a zero-shot frame- work: COLA (Contextualized Commonsense Causality Reasoner) to solve the task from the causal inference perspective. This framework obtains rich incidental supervision from tem- porality and balances covariates from multiple timestamps to remove confounding effects. Our extensive experiments show that COLA1 can detect commonsense causality more accurately than baselines. 1 Introduction Commonsense Causal Reasoning (CCR) aims at identifying plausible causes and effects of events in natural language that are typically reasonable by an average person (Zhang et al., 2022c). To solve the task, existing efforts devoted by the community mainly rely on language models wholeheartedly with supervised learning approaches (Stali¯unait˙e et al., 2021; Sap et al., 2019; Tamborrino et al., 2020; He et al., 2020; Raffel et al., 2020). Those ingenious engineering works have brought signifi- cant progress in recent years. However, recent stud- ies (Kavumba et al., 2019; Han and Wang, 2021) found that pure engineering designs are inadequate to seize commonsense causation, as language mod- 1The code and data are available at https://github. com/HKUST-KnowComp/COLA. Cause 1. Emma felt hungry. Cause 2. Emma was doing her job. Emma made a steak in the kitchen. 🥩 Context-Free 🤔 ✅ Which cause is more plausible? Cause 1. Emma felt hungry. Cause 2. Emma was doing her job. Emma made a steak in the kitchen. 🥩 Context 1: Emma exercised for a while. ✅ 🏃 Cause 1. Emma felt hungry. Cause 2. Emma was doing her job. Emma made a steak in the kitchen. 🥩 Context 2: Emma is a chef at a steakhouse. ✅ 👩🍳 ❌ ❌ Figure 1: An illustration of leveraging context to con- duct commonsense causal reasoning. Both causes could be plausible under different contexts, while only the frequent one (i.e., Cause 1) is plausible without context. els tend to reach higher scores by exploiting super- ficial artifacts in data. Recently, Zhang et al. (2022c) first studied grasp- ing commonsense causation from the causal infer- ence perspective, by drawing analogies between ob- servational …

Content truncated to 500 words