The Good and Bad in Economic News: Comparing (Automatic) Measurements of Sentiment in Dutch Economic News

• In Event: News Analysis Using Computational Methods

Abstract

Scholars have repeatedly investigated the tone of economic news coverage over the course of many years or even decades. Because such studies, ideally, analyze more news items than feasibly is done with a manual content analysis, automatic measurements of tone have become increasingly popular and several classifiers have been developed and applied. However, similarities and differences between such automated sentiment classifiers in news items remain largely unknown, especially in smaller languages as Dutch. The current study compares four of such tools (SentiStrength, Pattern, Polyglot, DANEW) with a manual analysis of tone and a simple classifier (based on only 65 words). Results demonstrate that Polyglot and the simple classifier yield tone scores that are most comparable with manual coding. Remarkably, the different automatic classifiers only correlate very weakly among themselves. Overall, automatic sentiments measures are slightly better in capturing the tone of news headlines than that of complete articles.

Authors

• Mark Boukes, University of Amsterdam / ASCoR

• Bob Robbert Nicolai van de Velde, University of Amsterdam

• Rens Vliegenthart, U of Amsterdam