PolyglotFakeFacts-ITPS: Reproducibility Package for Multilingual Information Threat Prioritization

This repository provides the reproducibility package associated with the experimental evaluation of the Information Threat Priority Score (ITPS), an interpretable framework for multilingual information-threat prioritization. The package is based on a controlled subset of the PolyglotFakeFacts V2 dataset and contains 2,000 news articles (1,000 Fake and 1,000 Real) with identical language distributions across classes. The subset was constructed using a deterministic language-balanced sampling procedure with random seed 20260819. The experimental ITPS evaluated in this study combines three components: leakage-controlled authenticity risk (A), embedding-derived security relevance (S), and cross-lingual narrative co-presence (P). Security relevance and cross-lingual semantic similarity were computed using the sentence-transformers/paraphrase-multilingual-mpnet-base-v2 multilingual embedding model. The package includes the experimental subset, the selection and provenance manifest, the executed experiment notebook, leakage and robustness audits, precomputed out-of-fold authenticity probabilities, final article-level A/S/P/ITPS results, ranking-reversal analyses, sensitivity analyses, and documentation of the experimental environment and methodology. The authenticity component was evaluated using source/domain-disjoint and near-duplicate-aware five-fold cross-validation with boilerplate/source-fingerprint stripping. The final leakage-controlled authenticity probabilities used by the ITPS experiment achieved ROC-AUC = 0.943520. This package is intended to support transparency, verification, and reproducibility of the reported ITPS experiment. It is a derived experimental research artifact and does not replace or constitute a new version of the parent PolyglotFakeFacts dataset. Parent dataset: Ciobanu, Alexandru (2026), “PolyglotFakeFacts: A Multilingual Dataset of Fake and Real News across Politics, Security, and Social Domains,” Mendeley Data, V2, DOI: 10.17632/gff8bmr4ff.2.本仓库提供了与信息威胁优先级评分（Information Threat Priority Score，简称ITPS）实验评估相关的可复现性套件，该评分是一种面向多语言信息威胁优先级排序的可解释框架。 本套件基于PolyglotFakeFacts V2数据集的受控子集，包含2000篇新闻文章（其中虚假新闻、真实新闻各1000篇），且两类样本的语言分布完全一致。 该子集通过确定性语言平衡采样流程构建，随机种子设为20260819。 本研究评估的实验版ITPS包含三个核心组件：可控泄露真实性风险（leakage-controlled authenticity risk，简称A）、嵌入衍生安全相关性（embedding-derived security relevance，简称S）以及跨语言叙事共存度（cross-lingual narrative co-presence，简称P）。安全相关性与跨语言语义相似度均通过sentence-transformers/paraphrase-multilingual-mpnet-base-v2多语言嵌入模型计算得到。 本套件包含实验子集、样本选择与来源说明文档、已执行的实验笔记本、泄露与鲁棒性审计报告、预计算的折外真实性概率、最终的文章级A/S/P/ITPS评分结果、排名反转分析、敏感性分析，以及实验环境与方法论说明文档。 真实性组件的评估采用了源/域不相交且感知近重复的五折交叉验证，并移除了文本模板与源指纹。本ITPS实验所使用的最终可控泄露真实性概率达到了ROC-AUC=0.943520的性能。 本套件旨在提升已报道的ITPS实验的透明度、可验证性与可复现性。其属于衍生的实验研究成果，并非原始PolyglotFakeFacts数据集的替代版本或新版本。 原始数据集：Ciobanu, Alexandru (2026), 《PolyglotFakeFacts：涵盖政治、安全与社会领域的多语言虚假与真实新闻数据集》，Mendeley Data，V2，DOI：10.17632/gff8bmr4ff.2.