Author Unknown: Evaluating Performance of Author Extraction Libraries on Global Online News Articles
Sriharsha Hatwar, Virginia Partridge, Rahul Bhargava, Fernando Bermejo
TL;DR
Evaluation of five existing software packages and one customized model for author extraction shows evidence for Go-readability and Trafilatura as the most consistent solutions, but all packages produce highly variable results across languages.
Abstract
Analysis of large corpora of online news content requires robust validation of underlying metadata extraction methodologies. Identifying the author of a given web-based news article is one example that enables various types of research questions. While numerous solutions for off-the-shelf author extraction exist, there is little work comparing performance (especially in multilingual settings). In this paper we present a manually coded cross-lingual dataset of authors of online news articles and use it to evaluate the performance of five existing software packages and one customized model. Our evaluation shows evidence for Go-readability and Trafilatura as the most consistent solutions for author extraction, but we find all packages produce highly variable results across languages. These findings are relevant for researchers wishing to utilize author data in their analysis pipelines, primarily indicating that further validation for specific languages and geographies is required to rely on results.