A recurrent stick breaking topic model for argument stance detection

Debate websites are valuable social media platforms for discussing controversial issues and gaining insights into diverse perspectives. However, with thousands of arguments on popular topics, browsing through entire discussions can be too time-consuming to extract a summary of the main viewpoints. To address this challenge, natural language processing techniques have been proposed to automatically identify argument structures and determine a user's stance on an issue. In addition to detecting stances, identifying topics that remain contentious between factions is especially important for policy-makers. However, the topics discussed by users with the same stance may vary in importance. To extract potential text topics, models based on probability and neural topic models have been proposed, with the latter being more effective due to improved computational cost and the ability to extract more consistent topics. This research proposes a two-stage stance detection model, which combines a neural topic model based on variational autoencoder and a recurrent neural network to learn the characteristics of an argument and enhance the model with subtopic features. The experimental results show that the model can obtain more coherent subtopics with a coherence of 0.42, and achieve an accuracy of nearly 70% in stance detection, demonstrating the effectiveness of the subtopic feature detection process.