Explainable Event Detection with Event Trigger Identification as Rationale Extraction

Proceedings of Recent Advances in Natural Language Processing, pages 507–518 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_056 507 Explainable Event Detection with Event Trigger Identification as Rationale Extraction Hansi Hettiarachchi Birmingham City University Birmingham, UK hansi.hettiarachchi@bcu.ac.uk Tharindu Ranasinghe Aston University Birmingham, UK t.ranasinghe@aston.ac.uk Abstract Most event detection methods act at the sentence-level and focus on identifying sen- tences related to a particular event. However, identifying certain parts of a sentence that act as event triggers is also important and more challenging, especially when dealing with lim- ited training data. Previous event detection at- tempts have considered these two tasks sepa- rately and have developed different methods. We hypothesise that similar to humans, suc- cessful sentence-level event detection models rely on event triggers to predict sentence-level labels. By exploring feature attribution meth- ods that assign relevance scores to the inputs to explain model predictions, we study the be- haviour of state-of-the-art sentence-level event detection models and show that explanations (i.e. rationales) extracted from these models can indeed be used to detect event triggers. We, therefore, (i) introduce a novel weakly-super- vised method for event trigger detection; and (ii) propose to use event triggers as an explain- able measure in sentence-level event detection. To the best of our knowledge, this is the first ex- plainable machine learning approach to event trigger identification. 1 Introduction Every day, numerous socio-political protest events occur worldwide, targeting various decisions made by governments or authorities (Hutter, 2014; Weng and Lee, 2021). These events hold significant importance for political scientists, policymakers, democracy watchdogs, and other stakeholders (Raleigh et al., 2010) due to their potential to pro- vide insights into multiple aspects (Tarrow, 2022). These include analysing the nature, scope, and mag- nitude of such events, shaping public opinion re- garding different causes, assessing the status of freedom and democracy in different nations, and more (H¨urriyeto˘glu et al., 2021b). Due to the continuous and abundant data flow over time, news media outlets serve as invaluable sources for social and political scientists who seek to establish comprehensive knowledge bases of protest events (Chenoweth and Lewis, 2013). Early approaches to creating these knowledge bases re- lied on manual event detection methods (Wang et al., 2016), which can be expensive and slow. Therefore, to cope with the volume of news me- dia, researchers have experimented with automatic event detection methods (Leetaru and Schrodt, 2013). The organisation of the recent shared tasks such as CASE: Challenges and Applications of Au- tomated Extraction of Socio-political Events from Text (H¨urriyeto˘glu et al., 2021a, 2022) has pro- moted automatic event detection research within the natural language processing (NLP) community. Automated event detection tools are designed as pipelines that receive news articles and yield records of events. The first step of these pipelines is discriminating between relevant and irrelevant sentences (Croicu and Weidmann, 2015). In this research, we refer to this as sentence-level event detection. Once event-related sentences are deter- mined, the next task is to extract event information on the token level (Doddington et al., 2004). One such …

Content truncated to 500 words