Research Papers 论文研究 2d ago Updated 1d ago 更新于 1天前 42

FrenchNews-7: Benchmarking Cross-Publisher French News Editorial Desk Classification FrenchNews-7:跨出版社法语新闻编辑部门分类基准测试

FrenchNews-7 introduces a cross-publisher benchmark for classifying French-language news articles into seven editorial desk categories using a multi-outlet corpus and URL-derived taxonomy A hybrid labeling pipeline combines publisher URL slugs with LLM annotation for ambiguous cases, achieving strong inter-rater reliability (human-human κ = 0.806, pairwise κ ≥ 0.766) CamemBERT-base fine-tuned on full article text emerges as the strongest model, achieving 0.799 overall recall and outperforming ze FrenchNews-7是首个跨出版商法国新闻编辑部门分类基准,结合多出版商语料库、URL派生七类分类法和微调CamemBERT分类器 采用混合标注管道(出版商URL slugs + LLM标注),经2人+2LLM评分者间研究验证,配对κ≥0.766,人-人κ=0.806 CamemBERT-base在完整文章文本上表现最佳,整体召回率0.799,超越GPT-OSS-120B、Mistral Small 3.2、Llama-3.3-70B等零样本LLM基线 跨出版商评估揭示边界稳定性不均:Sport/Culture & Loisirs/International转移干净,Economie(召回

55
Hot 热度
68
Quality 质量
58
Impact 影响力

Analysis 深度分析

TL;DR

  • FrenchNews-7 introduces a cross-publisher benchmark for classifying French-language news articles into seven editorial desk categories using a multi-outlet corpus and URL-derived taxonomy
  • A hybrid labeling pipeline combines publisher URL slugs with LLM annotation for ambiguous cases, achieving strong inter-rater reliability (human-human κ = 0.806, pairwise κ ≥ 0.766)
  • CamemBERT-base fine-tuned on full article text emerges as the strongest model, achieving 0.799 overall recall and outperforming zero-shot LLM baselines (GPT-OSS-120B, Mistral Small 3.2, Llama-3.3-70B)
  • Cross-publisher generalization is uneven: Sport, Culture & Loisirs, and International transfer cleanly, while Economie (recall = 0.517) and Societe (precision = 0.577) expose inherent editorial boundary ambiguity
  • The work demonstrates that some classification failures reflect genuine editorial convention disagreements rather than model limitations, establishing a reliability-tier guidance framework

Why It Matters

This benchmark addresses a critical gap in multilingual NLP by providing a rigorously evaluated resource for French-language news classification, a domain where cross-publisher generalization remains underexplored. The finding that certain categories hit a performance ceiling due to ambiguous editorial boundaries—not model capacity—offers practitioners a realistic framework for setting expectations when deploying text classification systems in real-world newsroom environments.

Technical Details

  • Dataset: FrenchNews-7 is a large multi-outlet French-language news corpus with a seven-class taxonomy derived from publisher URL structures, covering categories such as Sport, Culture & Loisirs, International, Economie, and Societe
  • Labeling pipeline: Hybrid approach combining deterministic publisher URL slug extraction with LLM-based annotation for structurally ambiguous cases, validated through an inter-rater study involving 2 humans and 2 LLMs
  • Model evaluation: Compares lexical, multilingual, and French-specific classifiers (including CamemBERT variants) under both in-distribution and held-out-publisher settings, with zero-shot LLM baselines tested on the held-out pool
  • Key results: CamemBERT-base on full article text achieves 0.799 overall recall, surpassing headline-only input and all zero-shot LLM baselines; Economie and Societe categories show the largest performance gaps, with Economie recall (0.517) closely matching blinded human agreement (0.55)
  • Reproducibility: Fine-tuned model, labeled manifest, collection scripts, and reliability-tier guidance table are publicly released

Industry Insight

  • News organizations and media analytics firms should treat editorial desk classification as a bounded problem: categories like Economie and Societe may inherently resist high-accuracy automation due to genuine structural ambiguity in French journalism conventions
  • The cross-publisher evaluation methodology provides a template for assessing model generalization in other language-specific and domain-specific classification tasks where training and deployment outlets differ
  • Practitioners should adopt reliability-tier frameworks when deploying classifiers, explicitly communicating to stakeholders which categories are production-ready versus which require human oversight due to irreducible boundary ambiguity

TL;DR

  • FrenchNews-7是首个跨出版商法国新闻编辑部门分类基准,结合多出版商语料库、URL派生七类分类法和微调CamemBERT分类器
  • 采用混合标注管道(出版商URL slugs + LLM标注),经2人+2LLM评分者间研究验证,配对κ≥0.766,人-人κ=0.806
  • CamemBERT-base在完整文章文本上表现最佳,整体召回率0.799,超越GPT-OSS-120B、Mistral Small 3.2、Llama-3.3-70B等零样本LLM基线
  • 跨出版商评估揭示边界稳定性不均:Sport/Culture & Loisirs/International转移干净,Economie(召回率0.517)和Societe(精确率0.577)接近人工盲评水平,反映编辑惯例而非模型瓶颈
  • 模型、标注清单、数据收集脚本和可靠性分层指南已开源

为什么值得看

本文提供了法语新闻分类任务的系统性基准,填补了多语言新闻自动化领域的空白。其混合标注管道和跨出版商评估框架为低资源语言NLP研究提供了可复用的方法论参考。

技术解析

  • 数据集与分类体系:FrenchNews-7构建于法国多出版商语料库,采用基于URL的七类分类法(Sport、Culture & Loisirs、International、Economie、Societe等),覆盖法国主流新闻机构。
  • 混合标注管道:标签分配结合出版商URL slugs(结构化来源)与LLM标注(处理结构模糊案例),通过2名人类标注员+2个LLM的评分者间研究验证可靠性,人-人κ=0.806,人机配对κ≥0.766。
  • 模型评估:比较词汇、多语言和法语特定训练分类器,在分布内和保留出版商设置下测试。最强模型CamemBERT-base使用完整文章文本,优于仅标题输入,跨出版商泛化能力显著。
  • 零样本LLM对比:在保留出版商集上对比GPT-OSS-120B、Mistral Small 3.2、Llama-3.3-70B,CamemBERT在整体召回率上超越所有基线,差距集中在Economie和Societe等边界模糊类别。
  • 边界稳定性分析:Sport/Culture & Loisirs/International类别跨出版商转移干净;Economie召回率0.517接近人工盲评0.55,Societe精确率0.577吸收边界歧义,表明这些类别的分类上限受编辑惯例限制而非算法瓶颈。

行业启示

  • 专用模型优于通用LLM:在特定语言/领域任务中,微调的专用模型(如CamemBERT)在召回率和泛化能力上显著超越零样本大模型,新闻机构应优先投资领域适配而非依赖通用LLM。
  • 标注质量决定任务上限:Economie和Societe类别的模型性能接近人工盲评水平,提示分类任务的理论上限由标注一致性决定,应优先优化标注协议而非盲目扩大模型规模。
  • 跨出版商评估是必要标准:单一出版商评估会高估模型泛化能力,行业应建立跨来源基准测试,以识别编辑惯例差异带来的系统性偏差。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Benchmark 基准测试 Dataset 数据集 LLM 大模型 Fine-tuning 微调 Evaluation 评测