A Research Pathway for Stance Detection in Speech

Stance detection is a Natural Language Processing (NLP) task that involves identifying an individual’s standpoint on a specific topic and determining their stance as either in favor of or against it. It has various potential applications for governments and organizations, helping them understand how public opinion evolves across multiple social media platforms. Although humans use both written and spoken language for communication, most research in the field primarily focuses on text-based data for building models. This paper aims to put forth a novel research pathway for developing a stance detection model that is specifically trained on audio-based data. To establish a direction for our approach, we analyze several foundational factors, including the utterance of the stance and the emotional elements present in human speech and conversations. We subsequently construct a flow to process the data by extracting text- and audio-based features, as well as supplementary attributes such as the author’s profile. A two-phase modeling approach is proposed to integrate text and speech into a single ensemble model, aiming to enhance the accuracy of the predictions. Finally, several improvement opportunities are also identified, providing a baseline for future endeavors in audio-based stance detection.