XHotpotQA: A Benchmark for Cross-Lingual Knowledge Composition in Multi-Hop Question Answering

XHotpotQA：多跳问答中跨语言知识组合的基准测试

Abstract: Knowledge-intensive multi-hop question answering requires systems to select evidence and compose dependent facts, yet multilingual benchmarks usually translate an entire example into one language. This hides failures at language boundaries inside the reasoning chain.

摘要： 知识密集型多跳问答要求系统能够选择证据并组合相关事实，然而现有的多语言基准测试通常将整个示例翻译成单一语言。这掩盖了推理链中语言边界处的失败。

We introduce XHotpotQA, a controlled benchmark for cross-lingual knowledge composition over mixed-language evidence. Each instance is modeled as an evidence-dependency graph whose question, bridge evidence, answer-bearing evidence, and distractors have explicit language assignments.

我们引入了 XHotpotQA，这是一个针对混合语言证据进行跨语言知识组合的受控基准测试。每个实例都被建模为一个证据依赖图，其问题、桥接证据、包含答案的证据以及干扰项都有明确的语言分配。

The audited resource contains 15,661 training and 7,405 validation instances, with sentence-level support supervision and supplied distractors. In validation, 99.81% of items cross the question-to-gold-evidence language interface and 95.60% use gold paragraphs in different languages.

该经审计的资源包含 15,661 个训练实例和 7,405 个验证实例，并提供句子级的支持监督和干扰项。在验证集中，99.81% 的项目跨越了“问题到黄金证据”的语言接口，95.60% 的项目使用了不同语言的黄金段落。

Across three reader artifacts, full question-evidence mismatch is associated with 10.25 to 15.79 lower Unicode-aware answer F1 than partial alignment, and different-script evidence with deficits of 11.98 to 23.70 points; the corresponding adapted-selector contrasts are 1.71 and 1.78 points. Under this supplied-candidate design, the evaluated readers therefore show substantially larger condition-associated deficits than the selector.

在三个阅读器模型中，完全的问题-证据不匹配导致的 Unicode 感知答案 F1 分数比部分对齐低 10.25 到 15.79 个点，而不同脚本的证据则导致 11.98 到 23.70 个点的差距；相应的适配选择器对比差异仅为 1.71 和 1.78 个点。在这种提供候选的设计下，被评估的阅读器在条件相关性方面表现出的缺陷明显大于选择器。

XHotpotQA provides role-aware diagnostics, modular evaluation, and an audited test bed for knowledge-based systems that must integrate evidence across languages.

XHotpotQA 为必须跨语言整合证据的知识型系统提供了角色感知诊断、模块化评估以及经审计的测试平台。