This article presents the results of an annotation experiment involving 36 participants tasked with rating the subjectivity of 150 excerpts from Belgian French press articles and with identifying linguistic indicators of subjectivity in the excerpts. We first explore inter-annotator agreement and correlations between the annotation results and the variables associated with the articles, then perform qualitative analysis of a sample of annotated texts. We introduce “textual heat maps”, a convenient method to visualize the results of the token-level annotations. The study reveals that the perception of subjectivity in text varies among readers, but that some tokens are widely considered as indicators of subjectivity. While expected indicators such as sentiment words and expressive punctuation were found to influence interpretations, less established markers such as discourse markers and modal verbs also played a role. The use of the pronoun on was identified as a significant subjectivity indicator, despite its common use to conceal the author’s presence in journalistic writing. These findings offer insights into subjectivity in journalistic discourse and have significant implications for the field of journalism and for the improvement of automated models for subjectivity detection.
Read full abstract