Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging Digitized Newspapers to Collect Summarization Data in Low-Resource Languages

Noam Dahan, Omer Kidron, Gabriel Stanovsky

TL;DR

This paper addresses the scarcity of high-quality summarization data in low-resource languages by leveraging front-page teasers from digitized historical newspapers as naturally annotated summaries. It introduces a two-step data-collection approach to extract teaser-based (teaser, article) pairs and validates the method across seven languages, culminating in HebTeaseSum, a 7,774-sample Hebrew multi-document corpus built from a single title. The authors also develop an automatic teaser-article extraction pipeline, evaluating teaser identification and matching methods (TF-IDF, sentence-transformer, and zero-shot LLMs) and demonstrating the feasibility of large-scale data generation. The findings show that while LLMs can produce coherent summaries, coverage gaps persist, especially in lower-resource languages, underscoring the need for curated datasets and OCR-corrected data to enable robust evaluation and fine-tuning for multilingual summarization."

Abstract

High quality summarization data remains scarce in under-represented languages. However, historical newspapers, made available through recent digitization efforts, offer an abundant source of untapped, naturally annotated data. In this work, we present a novel method for collecting naturally occurring summaries via Front-Page Teasers, where editors summarize full length articles. We show that this phenomenon is common across seven diverse languages and supports multi-document summarization. To scale data collection, we develop an automatic process, suited to varying linguistic resource levels. Finally, we apply this process to a Hebrew newspaper title, producing HEBTEASESUM, the first dedicated multi-document summarization dataset in Hebrew.

Leveraging Digitized Newspapers to Collect Summarization Data in Low-Resource Languages

TL;DR

This paper addresses the scarcity of high-quality summarization data in low-resource languages by leveraging front-page teasers from digitized historical newspapers as naturally annotated summaries. It introduces a two-step data-collection approach to extract teaser-based (teaser, article) pairs and validates the method across seven languages, culminating in HebTeaseSum, a 7,774-sample Hebrew multi-document corpus built from a single title. The authors also develop an automatic teaser-article extraction pipeline, evaluating teaser identification and matching methods (TF-IDF, sentence-transformer, and zero-shot LLMs) and demonstrating the feasibility of large-scale data generation. The findings show that while LLMs can produce coherent summaries, coverage gaps persist, especially in lower-resource languages, underscoring the need for curated datasets and OCR-corrected data to enable robust evaluation and fine-tuning for multilingual summarization."

Abstract

High quality summarization data remains scarce in under-represented languages. However, historical newspapers, made available through recent digitization efforts, offer an abundant source of untapped, naturally annotated data. In this work, we present a novel method for collecting naturally occurring summaries via Front-Page Teasers, where editors summarize full length articles. We show that this phenomenon is common across seven diverse languages and supports multi-document summarization. To scale data collection, we develop an automatic process, suited to varying linguistic resource levels. Finally, we apply this process to a Hebrew newspaper title, producing HEBTEASESUM, the first dedicated multi-document summarization dataset in Hebrew.

Paper Structure

This paper contains 39 sections, 3 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Extensive Digitization Efforts.
  4. Terminology.
  5. Front-Page Teasers are Common, High-Quality Summaries
  6. Constructing Summarization Data from Teasers
  7. Human Evaluation
  8. Front-Page Teasers are Common in a Wide Range of Languages
  9. Experiments
  10. Metrics and LLM-as-a-Judge
  11. Results
  12. Generated summaries miss aspects of the gold summary.
  13. There is bigger difference in model performance in lower resource languages.
  14. Automatic Extraction of Teaser Summaries and Corresponding Articles
  15. Automatically Identifying Teasers and Matching them to Articles
  16. ...and 24 more sections

Figures (3)

  • Figure 1: Newspaper's front-page teasers are a natural source for high-quality, expert-written summaries of full articles appearing inside the paper. We show that they are common across languages, and lend themselves to straightforward data collection.
  • Figure 2: Front-Page Teasers are common in a variety of languages. Highlighted text shows the reference to relevant pages where the corresponding articles can be found, and serves as a useful signal for identifying summaries. From Left-to-Right: teasers from Rana Blad (Norway), Fréttablaðið (Iceland) and Stampa Sera (Italy). Links to newspapers and translations are in Appendix \ref{['sec:appendixn']}.
  • Figure 3: Our approach for extracting a summarization dataset from printed newspapers. We first find teasers and corresponding page numbers on the front page based on newspaper-specific keywords (e.g., "Full articles on Pages 8-9"). Then, we turn to these pages and find the articles most relevant to the content of the teaser, thus resulting in (teaser summary, relevant articles) pairs.