Scholarlink

<p>Generating movie review arguments that are both linguistically fluent and semantically consistent with a specific numerical rating remains a challenge for Large Language Models (LLMs), often resulting in rating drift. To address this, we propose a novel Rating-Aware Argument Generation framework designed to produce persuasive arguments aligned with fine-grained rating intensities. We first construct and release the Douban Movie Review Dataset (DMRD), a large-scale Chinese multimodal dataset containing posters, metadata, and rating-labeled reviews, filling a gap in non-English multimodal resources. Our framework integrates these modalities through a two-stage training pipeline: Supervised Fine-Tuning (SFT) with a rating consistency loss to regularize sentiment intensity, followed by Direct Preference Optimization (DPO) to better align with human quality standards. Experimental results demonstrate that our method is effective in improving rating alignment and mitigating rating drift, achieving a 72.16% rating accuracy while generating coherent and visually grounded arguments. Our DMRD dataset is publicly available and can be found at <a href="https://github.com/poriahcorvus/DMRD">https://github.com/poriahcorvus/DMRD</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Rating-aware argument generation for movie reviews with multimodal large language models and a new dataset

  • 期刊论文

  • 发表日期：

    2026年7月16日

查看全文
  • Wenjie Hua,
  • Quan Fang,
  • Muyi Sun,
  • Shibiao Xu,
  • Man Zhang

摘要

Generating movie review arguments that are both linguistically fluent and semantically consistent with a specific numerical rating remains a challenge for Large Language Models (LLMs), often resulting in rating drift. To address this, we propose a novel Rating-Aware Argument Generation framework designed to produce persuasive arguments aligned with fine-grained rating intensities. We first construct and release the Douban Movie Review Dataset (DMRD), a large-scale Chinese multimodal dataset containing posters, metadata, and rating-labeled reviews, filling a gap in non-English multimodal resources. Our framework integrates these modalities through a two-stage training pipeline: Supervised Fine-Tuning (SFT) with a rating consistency loss to regularize sentiment intensity, followed by Direct Preference Optimization (DPO) to better align with human quality standards. Experimental results demonstrate that our method is effective in improving rating alignment and mitigating rating drift, achieving a 72.16% rating accuracy while generating coherent and visually grounded arguments. Our DMRD dataset is publicly available and can be found at https://github.com/poriahcorvus/DMRD.