Exploring Artificial Image Generation for Stance Detection

ACL ARR 2025 February Submission832 Authors

Abstract:

Stance detection is a task aimed at identifying and analyzing the author's stance from textual data.
Previous studies primarily rely on analyzing the text itself, which may not fully capture the implicit stance conveyed by the author.
To address this limitation, we propose a novel approach that involves transforming the original text into an artificially generated image and using this visual representation to aid in stance detection.
Our approach begins by employing a large vision-language model to generate potential images for a given text. These images are carefully crafted to adhere to three specific criteria: relevance to the text, consistency with the target of the stance, and consistency with the stance itself.
Next, we introduce a comprehensive evaluation framework to select the optimal image from the set of generated candidates.
Once the optimal image has been selected, we introduce a multimodal stance detection model that leverages both the original textual content and the generated image to identify the author's stance.
The experimental results demonstrate the effectiveness of our proposed approach, and also indicates the importance of artificially generated images for stance detection.

Paper Type: Long

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

Research Area Keywords: Stance Detection, Artificial Image Generation, Multimodal Learning

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 832