Reasoner Outperforms: Generative Stance Detection with Rationalization for Social MediaThis short paper presents a generative stance detection framework that elicits rationales from GPT-3.5 conditioned on gold labels and distills them into small T5/Flan-T5 models, showing that multitask learning with rationale generation outperforms chain-of-thought and standard finetuning, beating th

Reasoner Outperforms: Generative Stance Detection with Rationalization for Social Media

Jiaqing Yuan[^yuan] (Amazon, New York, USA — jordanyuan111@gmail.com), Ruijie Xi[^xi] (Meta, Bellevue, WA, USA — rxi@ncsu.edu), and Munindar P Singh (Computer Science, North Carolina State University, Raleigh, NC, USA — mpsingh@ncsu.edu)

[^yuan]: Jiaqing Yuan's work was performed while he was at NC State University.

[^xi]: Ruijie Xi's work was performed while she was at NC State University.

Abstract

Stance detection is vital for promoting a trustworthy, human-centric Web by identifying biased or harmful narratives in user-generated content. Whereas recent LLM-based methods excel in accuracy, they often lack interpretability. We propose a generative stance detection approach that outputs explicit rationales and distills them into smaller language models (SLMs) via single-task and multi-task learning. Our method enables Flan-T5 to outperform GPT-3.5 zero-shot by up to 9.57%. We further show that rationales enhance multitask performance and improve distillation fidelity, advancing the development of transparent, fair, and trustworthy NLP systems.

Read the full paper

Full text is not republished here because this paper was not released under a Creative Commons license. Metadata and abstract only.

Do you like what you are reading? Subscribe to receive updates.

Unsubscribe anytime