Prodicus at SemEval-2023 Task 4: Enhancing Human Value Detection with Data Augmentation and Fine-Tuned Language Models

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 193–198 July 13-14, 2023 ©2023 Association for Computational Linguistics Lauri Ingman at SemEval-2023 Task 4: A Chain Classifier for Identifying Human Values behind Arguments Spencer Paulissen and Caroline J. Wendt All authors have contributed equally University of Colorado Boulder Department of Computer Science spencer.paulissen@colorado.edu and caroline.wendt@colorado.edu Abstract Identifying expressions of human values in textual data is a crucial albeit complicated challenge, not least because ethics are highly variable, often implicit, and transcend circum- stance. Opinions, arguments, and the like are generally founded upon more than one guiding principle, which are not necessarily indepen- dent. As such, little is known about how to classify and predict moral undertones in nat- ural language sequences. Here, we describe and present a solution to ValueEval, our shared contribution to SemEval 2023 Task 4. Our re- search design focuses on investigating chain classifier architectures with pretrained contex- tualized embeddings to detect 20 different hu- man values in written arguments. We show that our best model substantially surpasses the clas- sification performance of the baseline method established in prior work. We discuss limita- tions to our approach and outline promising directions for future work. 1 Introduction 1.1 Motivation Human values are inextricably linked to the forma- tion of thoughts, opinions, and actions. They under- lie an individual’s moral, ethical, and philosophical convictions which are manifest in their worldview. In essence, public and private discourse are expres- sions of human values. As online communication platforms increasingly provide social alternatives to the "town square," understanding written con- tent, as it relates to the human values contained within it, is an interesting and important NLP chal- lenge. ValueEval presents a unique opportunity to investigate the human values behind the formation and presentation of opinions on controversial top- ics, particularly those with social, economic, and political implications (Kiesel et al., 2023). The main objective of ValueEval1 is to detect 1https://touche.webis.de/semeval23/ touche23-web/index.html#evaluation human values in arguments. Given an argument with three attributes and a finite set of human value categories, the system must classify whether or not the argument is associated with a given value for all values in the set. ValueEval is composed of a single task rather than sub-tasks; however, the committee specifies the option to focus on a subset of the 20 human value categories for classification. Here, we employ a classification model using all 20 human value categories. 1.2 Related Work Since its debut, BERT (Devlin et al., 2018) has be- come a bedrock of many NLP applications (Tenney et al., 2019). BERT enables researchers to expedite the task of creating contextual word embeddings. More recently, newer alternatives to BERT have become increasingly popular, including the GPT models created by OpenAI. Yet, BERT continues to transform the traditional architecture for NLP prob- lems: provided a sample as input to a pre-trained transformer language model, one can generate con- textualized embeddings to then be passed to down- stream tasks. Following in the steps of many others, …

Content truncated to 500 words