Event Causality Identification via Derivative Prompt Joint Learning

Proceedings of the 29th International Conference on Computational Linguistics, pages 2288–2299 October 12–17, 2022. 2288 Event Causality Identification via Derivative Prompt Joint Learning Shirong Shen1 and Heng Zhou1 and Tongtong Wu1 and Guilin Qi1∗ 1School of Computer Science and Engineering, Southeast University, China {ssr, zhouheng2020, wutong8023, gqi}@seu.edu.cn Abstract This paper studies event causality identification, which aims at predicting the causality relation for a pair of events in a sentence. Regarding event causality identification as a supervised classification task, most existing methods suf- fer from the problem of insufficient annotated data. In this paper, we propose a new derivative prompt joint learning model for event causality identification, which leverages potential causal knowledge in the pre-trained language model to tackle the data scarcity problem. Specifi- cally, rather than external data or knowledge augmentation, we derive two relevant prompt tasks from event causality identification to en- hance the model’s ability to identify explicit and implicit causality. We evaluate our model on two benchmark datasets and the results show that our model has great advantages over previ- ous methods. 1 Introduction Event causality identification (ECI) is an impor- tant natural language processing (NLP) task, which aims at identifying causality between events in sen- tences. Event causality identification supports a variety of NLP applications, e.g., machine read- ing comprehension (Berant et al., 2014) and event prediction (Radinsky et al., 2012). (Berant et al., 2014; Radinsky et al., 2012). As shown in Fig- ure 1, an ECI model identifies the causalities in sentences S1 and S2: (i) practice cause −→ won in S1; (ii) attack cause −→ killed in S2. The causality between events in a sentence mainly contains two types: explicit causality and implicit causality. For instance, the causality practice cause −→ won in S1 is an explicit causality, which is triggered by the explicit cue words in the sentence. ECI models can take causal cue words as the shortcut for explicit causality identification. As a contrast, the causality * Corresponding author. Figure 1: Examples of different causalities. S1 contains explicit causality between practice and won. S2 con- tains implicit causality between killed and attack. S3 has the same patterns with S2, but it does not contain any causality. attack cause −→ killed in S2 is an implicit causality, because none of explicit cue words is mentioned in S2. And as shown in Figure 1, comparing S2 and S3, we may not always derive the existence of causality from two highly similar expressions without explicit cue words. In other words, Implicit causality must be inferred from the semantics and contexts of events. Most existing methods regard ECI as a classi- fication task, and train customized ECI models on annotated data by supervised learning (Cheng and Miyao, 2017; Choubey and Huang, 2017). However, the large-scale annotated datasets of ECI are relatively hard to collect, referring to that the so far largest dataset EventStoryLine (Caselli and Vossen, 2017) only contains 258 documents and 4316 sentences. Therefore, ECI models are challenged by the data scarcity problem in su- …

Content truncated to 500 words