Interactive News Summarization

Newsletters and the opening paragraph in news articles (i.e. lead text) aim at summarizing the most important information while maintaining the reader’s attention. This important task is done manually, since text quality of state-of-the-art summarization models is not sufficient. A lead or newsletter text that contains all information from an article would fail at capturing the reader’s interest in continuing reading. Additionally, current summarization models are highly prone to hallucinate content, that is, to fabricate facts that are not supported by the original text. The goal of this project is to build a demo application to assist journalists in writing lead paragraphs or newsletters in real time using a summarization model. Specifically, we focus on single-document summarization in German and aim at leveraging the user’s input to guide and improve the generated summaries.

Goal

The objective of this project is to perform research on guided single-document summarization of German text. In addition, we aim at implementing a demo application to support the journalists with the task of writing lead texts or newsletters. The application will use our guided summarization models to ensure the generated summary aligns with specific information from the news article. This would allow journalists to write leads or newsletters more efficiently while maintaining their quality and the user’s interest.

Outcome

At the core of this project, we built guided summarization models to support the generation of lead news articles in German. We also developed a demo application that is designed to assist users in crafting engaging lead paragraphs for news articles in German. Our demo application leverages our guided summarization models to identify relevant information in the source article and uses it as a guidance to generate a concise summary. Additionally, the user interface provides manual highlighting capabilities, allowing users to emphasize key information according to their preferences.

external page Public Open-Source Repository