CORDIAL

该数据集名为CORDIAL，它是一个评估多模态话语分析（MDA）的基准，通过在不同的话语领域内，跨越不同粒度层次上对连贯性关系进行分析。该基准包含了四个难度递增的设置：DisRel、Tweet字幕、CLUE单标签（SL）和CLUE多标签（ML）。它对来自9个不同模型家族的12个多语言模型的表现进行了评估。数据集覆盖了不同话语领域内广泛的连贯性关系，其任务包括连贯性关系预测和验证。

Named CORDIAL, this dataset is a benchmark for evaluating Multimodal Discourse Analysis (MDA). It analyzes coherence relations across diverse discourse domains and at multiple granularity levels, and includes four settings with increasing difficulty: DisRel, Tweet Captions, CLUE Single-Label (SL), and CLUE Multi-Label (ML). The benchmark evaluates the performance of 12 multilingual models from 9 distinct model families. The dataset covers a wide range of coherence relations across different discourse domains, with its tasks including coherence relation prediction and verification.