EVENT DOSSIER
On September 7, 2026, a paper titled “Discourse Dependency: A Continuous Criterion for Translation Difficulty” was published on the arXiv cs.CL preprint platform. This study proposes using “discourse dependency” as a continuous criterion for evaluating translation difficulty, aiming to go beyond traditional discrete classification methods and more accurately quantify the complexity of text translation at the discourse level.
本文提出“话语依赖（Discourse Dependency，DDP）”作为衡量翻译难度的新标准。该指标基于命名实体重提和代词指称计算，无需参考金标准即可识别需长距离上下文解决的片段。在 WMT24++ 和 WMT25 测试中，高难度片段占比低且现有领域标签无法区分。实验显示，随着 DDP 值增加（如≥15），五种上下文注入策略均无法匹配人类后编辑水平，而自动指标无差异。研究认为，面对前沿系统分数饱和，评估重点应从模型得分转向其上下文检索能力。