Feedback-Verified Stance-controlled Text Generation Model

Stance-controlled text generation aims to produce texts with predefined stance labels (support or oppose). While modern large language models (LLMs) can generate naturally fluent texts through the recent advances in AI technology, they still exhibit limitations when addressing domain-specific requirements, necessitating fine-tuning on specialized datasets. To address this challenge, we propose a feedback-verified stance-controlled text generation model. The model fine-tunes open-source LLMs using stance detection datasets and incorporates a feedback validation mechanism to enhance the stance controllability of generated texts. Experimental results demonstrate that our method achieves an optimal balance between stance controllability and generation quality. Compared with several mainstream approaches, our model shows superior performance in stance controllability with a 2.5% improvement, while maintaining near-optimal performance in both text relevance and fluency, closely approaching the best results in these metrics.