T.M. Scanlon at SemEval-2023 Task 4: Leveraging Pretrained Language Models for Human Value Argument Mining with Contrastive Learning

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 603–608 July 13-14, 2023 ©2023 Association for Computational Linguistics T.M. Scanlon at SemEval-2023 Task 4: Leveraging Pretrained Language Models for Human Value Argument Mining with Contrastive Learning Milad Molazadeh Oskuee Department of Computer Science Iran University of Science and Technology Tehran, Iran molazadeh_milad@comp.iust.ac.ir Mostafa Rahgouy Department of Computer Science Auburn University, Alabama, USA mzr0108@auburn.edu Hamed Babaei Giglou TIB Leibniz Information Centre for Science and Technology Hannover, Germany hamed.babaei@tib.eu Cheryl D Seals Department of Computer Science Auburn University, Alabama, USA sealscd@auburn.edu Abstract Human values are of great concern to social sciences which refer to when people have dif- ferent beliefs and priorities of what is generally worth striving for and how to do so. This paper presents an approach for human value argument mining using contrastive learning to leverage the isotropy of language models. We fine-tuned DeBERTa-Large in a multi-label classification fashion and achieved an F1 score of 49% for the task, resulting in a rank of 11. Our proposed model provides a valuable tool for analyzing ar- guments related to human values and highlights the significance of leveraging the isotropy of large language models for identifying human values. 1 Introduction Argument mining is a promising area of research in natural language processing that aims to identify and analyze arguments from the text. Identifying human values in arguments is significant because different beliefs and priorities can cause conflicts or alignment between individuals or cultures, leading to disagreements on controversial issues. Under- standing the values behind different perspectives can improve communication and decision-making. The automatic identification of values in written arguments is valuable for argument faceted search, value-based argument generation, and value-based personality profiling (Kiesel et al., 2023, 2022). The ValueEval1, a task 4 of the shared task on Se- mEval 2023 is designed to automatically identify human values in arguments. The task has been defined as follows: Given a textual argument and a human 1https://touche.webis.de/semeval23/touche23-web/ value category, classify whether or not the argument draws on that category. Touché23-ValueEval dataset (Mirzakhmedova et al., 2023) for this task consists of a valuable collection of argumentative texts, where each ar- gument is labeled with one or more of the human values. Arguments are given as premise text, con- clusion text, and binary stance of the premise to the conclusion whatever it is "in favor of" or "against" the premise text. The task uses a set of 20 value categories compiled from the social science litera- ture. The dataset for this task consists of two differ- ent sets called Main, and Supplementary datasets, where in this research we focus on all 20 value cat- egories of the Main dataset for identifying human values behind arguments in the English language. Contrastive learning (CL) (Gao et al., 2021) has shown promising results in various natural lan- guage processing tasks, including text classifica- tion, sentiment analysis, and named entity recog- nition. CL is a technique used to learn representa- tions that are invariant to changes in …

Content truncated to 500 words