SeriesBench
收藏资源简介:
SeriesBench是一个包含105个精心策划的叙事驱动系列的数据集,涵盖了28个需要深度叙事理解的专门任务。该数据集由北京航空航天大学的研究团队创建,旨在评估多模态大型语言模型(MLLMs)对叙事驱动系列的理解能力。数据集内容丰富多样,包括日常生活、动漫、时空旅行、历史剧、奇幻等多种类型的叙事驱动系列。SeriesBench数据集的创建过程采用了新颖的长跨度叙事注释方法和全信息转换方法,将手动注释转换为多种任务格式。该数据集的应用领域包括系列推荐、互动媒体和自主视频摘要等,旨在解决现有模型在理解叙事驱动系列方面的挑战。
SeriesBench is a meticulously curated dataset consisting of 105 narrative-driven series, encompassing 28 specialized tasks that require deep narrative comprehension. Developed by a research team from Beihang University, this dataset aims to evaluate the narrative understanding capabilities of multimodal large language models (MLLMs) on narrative-driven series. The dataset covers a diverse range of narrative-driven content across genres including daily life, anime, time travel, historical dramas, fantasy, and more. The creation of SeriesBench adopts two novel methods: long-span narrative annotation and full-information conversion, which transform manual annotations into multiple task formats. Its application domains include series recommendation, interactive media, autonomous video summarization, and other related fields, with the goal of addressing the challenges faced by existing models in understanding narrative-driven series.




