SWAGex at SemEval-2020 Task 4: Commonsense Explanation as Next Event Prediction

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 422–429 Barcelona, Spain (Online), December 12, 2020. 422 SWAGex at SemEval-2020 Task 4: Commonsense Explanation as Next Event Prediction Wiem Ben Rim School of Computing Tokyo Institute of Technology Tokyo, Japan wiem.benrim@nlp.c.titech.ac.jp Naoaki Okazaki School of Computing Tokyo Institute of Technology Tokyo, Japan okazaki@c.titech.ac.jp Abstract We describe the system submitted by the SWAGex team to the SemEval-2020 Commonsense Validation and Explanation Task. We used multiple methods on the pre-trained language model BERT (Devlin et al., 2018) for tasks that require the system to recognize sentences against commonsense and justify the reasoning behind this decision. Our best performing model is BERT trained on SWAG fine-tuned for the task. We investigate the ability to transfer commonsense knowledge from SWAG to SemEval-2020 by training a model for the Explanation task with Next Event Prediction data. 1 Introduction Commonsense reasoning has long been a challenge in the field of Natural Language Processing (NLP). Recently, the emergence of large pre-trained language models achieving state-of-the-art results in several NLP tasks, such as BERT and GPT (Devlin et al., 2018; Radford et al., 2019) has also influenced commonsense tasks to regain popularity. Natural Language Inference (NLI) is one of the most popular tasks in commonsense reasoning research. It consists of determining if a hypothesis is correct, given a premise. Situations With Adversarial Generation (SWAG) task is closely related to NLI (Zellers et al., 2018). It consists of determining the most appropriate ending given a start phase as a context. The start phrases and endings in SWAG dataset are video captions. We show an example in Figure 1. The task is to find the ending that describes a possible (future) world that can follow the given sentence even when it is not strictly entailed. Making such inference necessitates a rich understanding of everyday physical situations (Zellers et al., 2018). Figure 1: SWAG dataset example. The startphrase offers one sentence as context, and a beginning part of the second sentence. The task is to predict its appropriate ending. In this paper, we explore the ability of SWAG to infer commonsense explanation to tackle Semeval 2020 Task 4: Commonsesense Validation and Explanation (ComVE). To this end, we followed the directions given by the organizers of the task (Wang et al., 2020) and attempted a contribution to their two main tasks. The first main task is concerned with the commonsense validation of a sentence, which is also called Sen- making. It consists of providing two sentences that have similar wordings, one being against commonsense This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/. 423 and the other not. The sentences given do not require specialized knowledge and are trivial for humans, but remain challenging for machines. The second main task is Commonsense Explanation that aims at finding the key reason why a statement does not make sense. While the first task tests the system’s performance in choosing which sentence is more likely to be true, …

Content truncated to 500 words