Two papers accepted to the conference "Natural Language Processing for Digital Humanities" (NLP4DH)

04/08/2025

Two papers accepted at NLP4DH, on argument persuasiveness, and on automatically structuring political statements in text.

BamNLP got two papers accepted at the conference "Natural Language Processing for Digital Humanities (NLP4DH)", collocated with the Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics.

Lynn Greschner and Roman Klinger. Fearful Falcons and Angry Llamas: Emotion Category Annotations of Arguments by Humans and LLMs. https://arxiv.org/abs/2412.15993
Dmitry Nikolaev and Sean Papay. Strategies for political-statement segmentation and labelling in unstructured text. https://arxiv.org/abs/2503.07179

In the first paper, the authors study if particular emotions (joy, fear, …) in argumentative texts affect how convincing an argument is. So far, work in argument mining focused mostly on the concept of emotionality – how emotional an argument is formulated, and less which emotion is expressed. The authors create the Emo-Defabel corpus with emotion annotations of arguments and find that positive emotions (e.g., joy, pride) correlate with higher convincingness, while negative emotions (e.g., anger) decrease the convincingness of arguments. Interestingly, three large language models are not able to correctly predict emotions in arguments. They are biased strongly towards negative emotions, when it comes to argumentative texts.

In the second paper, work towards facilitating research by political scientists. They are often interested in analyzing the positions taken by political parties across countries and over time. To do so, they identify and label the political statements present in party manifestos. In this work, the authors investigate different methods for using machine learning to find and label statements automatically, and how to interpret the results in terms of changing party positions. They find that, with a machine learning model capable of both finding and classifying political statements, automatic methods can reliably extract political statements from other political texts, including the transcripts of parliamentary speeches. Applied to the Hansard corpus, the record of parliamentary debate transcripts from the United Kingdom, they track parties' trajectories in the political spectrum based on the political statements made by party members in debates.