Stance Detection for Macro Topics based on Multi-factor Aggregation Analysis

Anonymous

Abstract: This paper presents an innovative exploration into stance detection, with a specific focus on subjects characterized by their inherently abstract and macroscopic nature, termed as ``macro topics.'' Due to the intricate complexity associated with these subjects, individuals often refrain from explicitly stating their opinions, thereby introducing challenges to stance detection when the target is implicit or unmentioned in the text. To address this complexity, we propose a tailored representation model designed to effectively encapsulate the nuanced aspects of macro topics. Our model relies on a comprehensive multidimensional analysis of sub-topics within a given macro topic, employing a specially designed discourse-based Latent Dirichlet Allocation (LDA) model. Utilizing this representation, an aggregation analysis is implemented to deduce stances on the macro topic by examining the array of sub-topic stances. The analysis of stances associated with sub-topics expressed in text is achieved by leveraging the semantic analysis capability of large language models (LLMs). Our approach attains superior stance detection accuracy, as validated through extensive experiments conducted on large-scale social media and finance text datasets.

Paper Type: long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English