Quintilian at SemEval-2023 Task 4: Grouped BERT for Multi-Label Classification

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1609–1612 July 13-14, 2023 ©2023 Association for Computational Linguistics Quintilian at SemEval-2023 Task 4: Grouped BERT for Multi-Label classification Ajay Narasimha Mopidevi University of Colorado, Boulder Ajay.Mopidevi@colorado.edu Hemanth Chenna University of Colorado, Boulder Hemanth.Chenna@colorado.edu Abstract In this paper, we initially discuss about the ValueEval (Kiesel et al., 2023) task and the challenges involved in multi-label classifica- tion tasks. We tried to approach this task using Natural Language Inference and proposed a Grouped-BERT 1 architecture which leverages commonality between the classes for a multi- label classification task. 1 Introduction Everyone has a perspective on how they approach a problem and make a decision, from the simplest de- cisions such as choosing to lend a pen to a friend to decisions that affect their life and everyone around them. These decisions are made consciously based on some of the values that they strongly believe in. The task 4 of SemEval-2023 - ValueEval (Kiesel et al., 2023), is to understand which of these human values form the basis for someone’s decision in a textual argument. ValueEval (Kiesel et al., 2023) presents the task of understanding human values as a Natural Lan- guage Processing problem, by proposing an an- notated dataset of textual arguments, with the la- bels being the human values that would be drawn to make that decision. Each argument in this dataset is provided with a Premise, Conclusion and Stance. Lets consider this example argument from the dataset with premise "marriage is the ultimate commitment to someone, if people want it they should be allowed", the conclusion "We should abandon marriage" and the stance "against". The task is to figure out why someone takes a particu- lar stance (in favor of/ against) for the conclusion, given the premise - "What human values led to someone taking this particular stance?" The task studies the (often implicit) human val- ues behind natural language arguments, such as to 1https://github.com/ajaymopidevi/Grouped-BERT have freedom of thought or to be broadminded. Val- ues are commonly accepted answers to why some options are desirable in the ethical sense and are thus essential both in real-world argumentation and theoretical argumentation frameworks. However, their large variety has been a major obstacle to modeling them in argument mining. In their dataset, Mirzakhmedova et al. (2023) also provided each argument with a set of human values in the form of labels that are closely aligned with psychological research. For each argument, these 20 labels provide greater detail about which human values are inferred for making the decision. 2 Background This classification task becomes much more chal- lenging as it is a multi-label classification problem. In classification problems, deep learning architec- tures try to update their layer weights to emphasize the output of softmax layer of the correct class by making it closer to 1, while also making the out- puts of the other classes closer to 0. In a multi-label classification problem, that’s not possible as each input may have multiple …

Content truncated to 500 words