Function Words as Stable Features for German Opinion Articles Classification

The paper addresses the challenge of stable classification of neutral news reports and opinion articles under shifts in publisher and domain. We apply experiments using different lexical and grammatical word categories on a hand curated data set of German online news and opinion articles that are labeled in terms of topic and publisher trustworthiness. The findings demonstrate that function words, typically used as stop words, are effective markers for opinion article classification that provide stable predictions under domain shifts and across different publishers. The results can help with the development of suitable feature sets to distinguish between opinion-free news articles and opinion articles, for example as a tool to flag online articles with a high level of subjectivity.