Tool for journalists to edit the text generation logic of an automated journalist

Abstract

Automated journalism means writing fact-based articles based on structured data using algorithms or software. The advantages of automated journalism are scalability, speed and lower costs. The limitations of it are fluency, quality of writing and limited perception. In this thesis, the different implementation methods of automated journalism were compared. These implementation methods were templates, decision trees, fact ranking method and different machine learning solutions. It was found out that no implementation method was strictly better than others but all had distinct advantages and disadvantages. When selecting an implementation method these factors should be taken into account and weighed. Finnish national broadcasting company Yle’s automated journalist Voitto-robot was discussed. Voitto’s implementation is based on templates and decision trees. While Voitto’s text generation is easily modifiable and transparent due to its implementation method, this was only available to programmers. The decision trees were implemented directly in the code which made them hard to understand and the template files were too complex to be easily edited. In this thesis, a proof-of-concept web application was made to allow journalists and other content creators the possibility to edit the templates and decision trees of Voitto independently. The created software was analysed and it was found that it helped journalists understand the text generation and modify it as they wanted. Even in its proof-of-concept state, it was good enough to be used to automate election reporting for the Finnish parliamentary election of 2019.