rl-bandits-lab/SEGALE-WMT24

SEGALE-WMT24是一个为SEGALE方案设计的长上下文机器翻译评估数据集，包含三种语言方向和四种合成设置。该数据集用于长上下文MT基准测试和元评估，研究度量标准和系统对欠翻译/过翻译及边界变化的敏感性。数据集包含英德、英西和日中三种语言对，每种语言对都有四种合成设置：drop source（模拟欠翻译）、drop target（模拟过翻译）、merge（模拟句子融合）和raw（基线系统预测）。每个翻译段包含源文本、系统翻译、系统标识符、文档ID和文档内的段索引。

SEGALE-WMT24 is a long-context machine translation evaluation dataset designed for the SEGALE scheme, containing three language directions and four synthetic setups. It is used for long-context MT benchmarking and meta-evaluation, studying the sensitivity of metrics and systems to under/over-translation and boundary variations. The dataset includes English-German, English-Spanish, and Japanese-Chinese language pairs, each with four synthetic setups: drop source (simulating under-translation), drop target (simulating over-translation), merge (simulating sentence fusion), and raw (baseline system predictions). Each translation segment includes source text, system translation, system identifier, document ID, and segment index within the document.