Modeling Complex Event Scenarios via Simple Entity-focused Questions

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2468–2483 May 2-6, 2023 ©2023 Association for Computational Linguistics Modeling Complex Event Scenarios via Simple Entity-focused Questions Mahnaz Koupaee1, Greg Durrett2, Nathanael Chambers3, Niranjan Balasubramanian1 1 Stony Brook University, 2 The University of Texas at Austin, 3 United States Naval Academy 1{mkoupaee,niranjan}@cs.stonybrook.edu 2gdurrett@cs.utexas.edu, 3nchamber@usna.edu Abstract Event scenarios are often complex and involve multiple event sequences connected through different entity participants. Exploring such complex scenarios requires an ability to branch through different sequences, something that is difficult to achieve with standard event lan- guage modeling. To address this, we propose a question-guided generation framework that models events in complex scenarios as answers to questions about participants. At any step in the generation process, the framework uses the previously generated events as context, but generates the next event as an answer to one of three questions: what else a participant did, what else happened to a participant, or what else happened. The participants and the ques- tions themselves can be sampled or be provided as input from a user, allowing for controllable exploration. Our empirical evaluation shows that this question-guided generation provides better coverage of participants, diverse events within a domain, comparable perplexities for modeling event sequences, and more effective control for interactive schema generation1. 1 Introduction Event scripts (Schank and Abelson, 1977), also known as event schemas, describe a sequence of events in a particular context. Representing and modeling such schemas is central to applications in AI such as question answering, discourse under- standing, and information extraction (Balasubrama- nian et al., 2013). Early work used hand-crafted event schemas as a starting point (Schank and Abel- son, 1977; Mooney and DeJong, 1985), but modern techniques attempt to extract these at a large scale from unlabeled data (Chambers and Jurafsky, 2008; Chambers, 2013; Pichotta and Mooney, 2016; We- ber et al., 2018b). 1The code is available at https://github.com/ StonyBrookNLP/qa-event-lms Figure 1: Question-guided event sequence generation. The user can interact with the system by asking ques- tions regarding the entities and the system will generate corresponding events. Different questions can lead to different paths as shown with different colored boxes. Event language models can also be used to approximate schematic knowledge via event se- quences. They can be trained to generate a se- quence of events with their participating roles de- scribing a real-life scenario. However, these sce- narios can often be complex and don’t always fit as simple sequences (Weber et al., 2018b). For exam- ple, suppose we have the following event: police arrested suspect on several charges, with police, suspect and charges as entities. The sce- nario can be described in multiple ways depending on which entities we want to focus on and what roles they play in the subsequent events. We may be interested in knowing what the police did, what the suspect did or what happened to the charges, each of which can be explored as its own sequence, 2468 yielding many interconnected sequences …

Content truncated to 500 words