Skip to content

Repository files navigation

Sentence-Level Text Reuse in Multilingual Journalism

This repository accompanies a paper accepted to the SoCon-NLPSI 2026 workshop, co-located with LREC 2026. The work, titled “Rewrite the News: Tracing Editorial Reuse Across News Agencies,” investigates sentence-level text reuse in multilingual newswire workflows. It introduces a weakly supervised framework for detecting content reuse between English-language articles from the national press organization STA (Slovenian News Agency) and 15 Foreign Agencies across 7 languages.


📝 Project Overview

This work addresses the following research questions:

  1. Where in the article is reuse most likely to occur?
  2. Which parts of foreign articles are most frequently reused?
  3. How is content merged from multiple foreign sources?

Our findings suggest that reused content is most often found in the middle or end of news articles, with many-to-many and paraphrased alignments being dominant. These insights contribute to understanding journalistic workflows and content transformation across languages.


📁 Repository Structure

project-root/
├── preprocessing.ipynb
├── experiment_with_wiki_corpus.ipynb
├── experiment_with_webis_paraphrase_corpus.ipynb
├── similarity_approach.ipynb
├── data/
│   ├── Webis-CPC-11_1000_sample.json
│   ├── wiki_test_corpus_full.json
│   ├── x_full_data.json

📄 File Descriptions

File Description
x_full_data.json The core aligned dataset containing sentence-level reuse instances between STA (Slovenian News Agency) and foreign agencies. Includes language, article IDs, sentence indices, and reuse labels.
wiki_test_corpus_full.json A clean multilingual test set of Wikipedia article sentences used for evaluating paraphrase detection baselines.
Webis-CPC-11_1000_sample.json A 1,000-example subset from the Webis Cross-lingual Paraphrase Corpus, used for benchmarking under noisy, multilingual conditions.
similarity_approach.ipynb Main notebook for computing similarity scores using embedding-based model like Sentence-BERT.
experiment_with_wiki_corpus.ipynb Evaluation using high-quality Wikipedia data to establish baseline performance.
experiment_with_webis_paraphrase_corpus.ipynb Robustness evaluation on real-world paraphrases from noisy web sources.
preprocessing.ipynb Data preparation pipeline: article filtering, sentence segmentation, alignment prep, and cleaning.

📄 x_full_data.json Schema

Each entry in x_full_data.json represents a sentence with metadata. Fields include:

Field Description
sen_id Unique identifier for the sentence (UUID).
article_id ID of the source article.
seq_nr Sentence position in the article.
lang Language of the sentence (e.g., en, sl).
text Raw sentence text.
reused Binary reuse label (1 = reused in foreign article, 0 = not reused).
partner_id ID of the foreign article with reused content (if applicable).
partner_sentence_id ID of the matched sentence in the foreign article (if applicable).

DISCLAIMER:

The data provided is intended solely for research and academic purposes. Any use of this data for commercial purposes, including but not limited to marketing, product development, or resale, is strictly prohibited. By accessing or using this data, you agree to abide by these terms and acknowledge that any unauthorized commercial use may result in legal action.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages