Event Temporal Relation Extraction with Bayesian Translational Model

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1125–1138 May 2-6, 2023 ©2023 Association for Computational Linguistics Event Temporal Relation Extraction with Bayesian Translational Model Xingwei Tan1, Gabriele Pergola1, Yulan He1,2,3 1Department of Computer Science, University of Warwick, UK 2Department of Informatics, King’s College London, UK 3The Alan Turing Institute, UK {Xingwei.Tan, Gabriele.Pergola.1}@warwick.ac.uk yulan.he@kcl.ac.uk Abstract Existing models to extract temporal relations between events lack a principled method to in- corporate external knowledge. In this study, we introduce Bayesian-Trans, a Bayesian learning- based method that models the temporal rela- tion representations as latent variables and in- fers their values via Bayesian inference and translational functions. Compared to conven- tional neural approaches, instead of performing point estimation to find the best set parame- ters, the proposed model infers the parameters’ posterior distribution directly, enhancing the model’s capability to encode and express un- certainty about the predictions. Experimental results on the three widely used datasets show that Bayesian-Trans outperforms existing ap- proaches for event temporal relation extraction. We additionally present detailed analyses on un- certainty quantification, comparison of priors, and ablation studies, illustrating the benefits of the proposed approach.1 1 Introduction Understanding events and how they evolve in time has been shown beneficial for natural language un- derstanding (NLU) and for a growing number of related tasks (Cheng et al., 2013; Wang et al., 2018; Ning et al., 2020; Geva et al., 2021; Sun et al., 2022). Howeover, events often form complex struc- tures with each other through various temporal relations, which is challenging to track even for humans (Wang et al., 2020a). One of the main difficulties is the wide vari- ety of linguistic expressions of temporal relations across different contexts. Although many of them share some linguistic similarities, most of the top- ics in which they occur are characterized by some shared but unspoken knowledge that determines how temporal information is expressed. For exam- ple, when it comes to health, prevention is widely 1Experimental source code is available at https:// github.com/Xingwei-Warwick/Bayesian-Trans Figure 1: Comparison between with or without external knowledge incorporation on event relation extraction. practised, with many treatments (e.g., vaccinations) being effective only if administered before the on- set of a disorder. On the contrary, in the automo- tive industry, it is common that most people repair their car after a problem occurs. However, despite its simplicity, such commonsense knowledge is rarely stated explicitly in text and varies greatly across different domains. For example, in Figure 1, a detection model lacking the commonsense knowledge that vaccination can protect people from infection, tends to get confused by the com- plex linguistic structures in the excerpt and returns the wrong prediction entailing that ‘died’ happens after ‘vaccinated’. Instead, with the considera- tion of prior temporal knowledge involving the vac- cination event from an external knowledge source ATOMIC (Hwang et al., 2021), a model gives the correct prediction that ‘died’ occurs before ‘vac- cinated’. Methods proposed in recent studies for event relation extraction are mostly …

Content truncated to 500 words