SUTNLP at SemEval-2023 Task 4: LG-Transformer for Human Value Detection

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 340–346 July 13-14, 2023 ©2023 Association for Computational Linguistics SUTNLP at SemEval-2023 Task 4: LG-Transformer for Human Value Detection Hamed Hematian Hemati, Sayed Hesam Alavian, Hossein Sameti, Hamid Beigy AI Group, Computer Engineering Department, Sharif University of Technology, Tehran, Iran { hematian.hemati.hamed, hesam.alavian70}@gmail.com, {sameti, beigy}@sharif.edu Abstract When we interact with other humans, human values guide us to consider the human element. As we shall see, value analysis in NLP has been applied to personality profiling but not to argument mining. As part of SemEval-2023 Shared Task 4, our system paper describes a multi-label classifier for identifying human val- ues. Human value detection requires multi- label classification since each argument may contain multiple values. In this paper, we pro- pose an architecture called Label Graph Trans- former (LG-Transformer). LG-Transformer is a two-stage pipeline consisting of a trans- former jointly encoding argument and labels and a graph module encoding and obtaining further interactions between labels. Using ad- versarial training, we can boost performance even further. Our best method scored 50.00 % using F1 score on the test set, which is 7.8 % higher than the best baseline method. Our code is publicly available on Github.1 1 Introduction Most, if not all, social sciences stress the impor- tance of human values (Rokeach, 1973) and have integrated them into computational frameworks of argumentation (Bench-Capon, 2003). The anal- ysis of values in NLP has been applied to per- sonality profiling but hasn’t been applied to argu- ment mining (Maheshwari et al., 2017). Values are commonly accepted answers to why some op- tion is desirable in the ethical sense and are thus essential both in real-world argumentation and the- oretical argumentation frameworks (Kiesel et al., 2022). For Identifying Human Values behind Ar- guments, Touché23-ValueEval collected 9,324 ar- guments from 6 diverse sources, including reli- gious texts, political discussions, free-text argu- ments, newspaper editorials, and online democracy platforms (Mirzakhmedova et al., 2023b). 1https://github.com/SUTNLP/LG_Transformer The purpose of this system paper is to present SUTNLP’s (David Gauthier on leaderboard) work on the SemEval-2023 Shared Task 4 which is fo- cused on developing a classifier to classify human values (Kiesel et al., 2023). Detecting human val- ues requires multi-label classification since each argument may contain several values. To tackle this problem, we propose a two-stage pipeline con- sisting of (1) a transformer (Vaswani et al., 2017) which jointly encodes the argument and labels and (2) a graph module (Wu et al., 2021) which fur- ther encodes and gets interactions between labels. Lastly, we employ adversarial training (Bai et al., 2021) to further enhance the model’s performance. The best performing method achieves F1 score of 50.00 % on the test set, outperforming the best baseline method by 7.8 %. Using our method, we ranked 8th out of 40 competing teams. Results are submitted through TIRA (Fröbe et al., 2023). 2 Background In human value detection task, a textual argument and multiple human value categories are given and the goal …

Content truncated to 500 words