MaRVL (Multicultural Reasoning over Vision and Language)

我们创建 a_x000D_
多元文化推理的多语言数据集_x000D_
通过从母语注释者中引出陈述来超越视觉和语言（MaRVL）_x000D_
关于成对的图像。该任务包括区分每个有根据的语句是否是_x000D_
对或错。我们建立了一系列baselines_x000D_
使用最先进的模型并发现他们的_x000D_
跨语言迁移性能明显落后于英语监督性能。_x000D_
这些结果促使我们重新评估鲁棒性_x000D_
当前最先进模型的准确性和准确性_x000D_
超越狭隘的领域，还要开辟新的_x000D_
real_x000D_ 开发面临的激动人心的挑战
多语言和多文化系统

We present a multilingual dataset for multicultural reasoning, termed Multilingual Reasoning beyond Vision and Language (MaRVL), which is constructed by eliciting statements from native annotators regarding paired images. The task requires distinguishing whether each justified statement is correct or incorrect. We establish a series of baseline models using state-of-the-art methods, and find that their cross-language transfer performance significantly lags behind the performance achieved under English supervision. These results motivate us to re-evaluate the robustness and accuracy of current state-of-the-art models beyond narrow domains, and to explore exciting new challenges in the development of multilingual and multicultural systems.

搜集汇总

背景与挑战

背景概述

MaRVL是一个多语言多文化的视觉与语言推理数据集，通过母语注释者创建图像-语句对，要求判断语句的真假。它于2021年由多个学术机构发布，旨在评估模型在跨文化场景下的鲁棒性。