MECI: A Multilingual Dataset for Event Causality Identification

Proceedings of the 29th International Conference on Computational Linguistics, pages 2346–2356 October 12–17, 2022. 2346 MECI: A Multilingual Dataset for Event Causality Identification 1Viet Dac Lai, 1Amir Pouran Ben Veyseh, 1Minh Van Nguyen 2Franck Dernoncourt, 1Thien Huu Nguyen 1Dept. of Computer Science, University of Oregon, OR, USA 2Adobe Research, Seattle, WA, USA {vietl,apouranb,minhnv,thien}@cs.uoregon.edu franck.dernoncourt@adobe.com Abstract Event Causality Identification (ECI) is the task of detecting causal relations between events mentioned in the text. Although this task has been extensively studied for English ma- terials, it is under-explored for many other languages. A major reason for this issue is the lack of multilingual datasets that provide consistent annotations for event causality rela- tions in multiple non-English languages. To address this issue, we introduce a new mul- tilingual dataset for ECI, called MECI. The dataset employs consistent annotation guide- lines for five typologically different languages, i.e., English, Danish, Spanish, Turkish, and Urdu. Our dataset thus enable a new research direction on cross-lingual transfer learning for ECI. Our extensive experiments demonstrate high quality for MECI that can provide am- ple research challenges and directions for fu- ture research. We will publicly release MECI to promote research on multilingual ECI. The dataset is available at https://github. com/nlp-uoregon/meci-dataset. 1 Introduction Event Causality Identification (ECI) is an impor- tant Information Extraction (IE) task that aims to identify causal relations between event mentions in text. For example, in the sentence “After inspec- tion of his computer , officers found that he was interested...”, a ECI system should detect a causal relation between two events “inspection” cause −−−→ “found”. ECI can provide valuable information for various applications such as event timeline con- struction (Shahaf and Guestrin, 2010), question- answering (Oh et al., 2016), future event forecast- ing (Hashimoto, 2019), and machine reading com- prehension (Berant et al., 2014). Due to its applications, ECI has been extensively studied in the natural language processing com- munity over the past decade. The vast majority of methods for ECI involve feature engineering models (Do et al., 2011; Hu and Walker, 2017; Hashimoto, 2019; Ning et al., 2018; Gao et al., 2019) and recent deep learning architectures (Kad- owaki et al., 2019; Zuo et al., 2021b; Liu et al., 2021; Zuo et al., 2021a; Man et al., 2022a). As such, the creation of large annotated datasets, e.g., EventStoryLine (Caselli and Vossen, 2017), has been critical to the development of ECI study. However, existing datasets for ECI only annotate causal relations between event mentions in data of a single language, i.e., mainly for English (Caselli and Vossen, 2017; Cybulska and Vossen, 2014; O’Gorman et al., 2016). On the one hand, this leaves many other languages unexplored for ECI, posing an important question about the generaliza- tion ability of existing methods to other languages. For instance, Spanish, Danish, and Turkish are not covered in those separated datasets for ECI. More- over, the current single-language datasets for ECI tend to employ different annotation guidelines that prevent their combination into a larger corpus and cross-lingual transfer learning research …

Content truncated to 500 words