Abstract
Online news platforms opened discussion forums under their articles to promote reader interaction and foster constructive discussion. But what makes a discussion constructive, and how can platforms encourage this behavior? Additionally, can computational tools help? This question is especially relevant due to the growing activity on these platforms. This thesis
... read more
addresses the ambiguity surrounding the concept of constructive discussion and explores how online news platforms currently try to encourage it, providing a foundation for computational tools to assist in moderating user comments.
The thesis adopts two anthropological perspectives to define constructive discussion: an etic perspective, where the concept is described objectively by external researchers, and an emic perspective, based on the insights of moderators who regularly distinguish between constructive and non-constructive comments.
In Chapter 2, research examines how five news platforms (The Guardian, Die Zeit, New York Times, El País, and NU.nl) moderate content and encourage constructive debate. While AI increasingly removes undesirable comments, moderators have more time to promote constructive conversations manually.
Chapters 3 and 4 focus on the etic perspective, emphasizing interactivity and argument diversity. A constructive discussion includes both supporting and opposing arguments. Formulas to calculate the balance between 'pro' and 'con' arguments were applied to social media platforms. The results show that argument diversity varies: on a platform like Gab, echo chambers dominate, while other platforms feature more diverse discussions.
In Chapter 4, comments from the Dutch platform Nujij (on climate change) were annotated in terms of pro and con arguments. Initial models failed to recognize rare arguments due to their infrequency. However, an active learning approach improved classification by adding more examples of these rare arguments, aiding moderators in filtering nuanced discussions.
The thesis then describes an emic perspective in Chapter 5, using moderator decisions to model what constitutes a constructive comment. A group recommender system was developed, ranking comments based on their likelihood of being featured. These models help moderators by showing them only the top 5 or 10 comments, reducing the need to review entire discussions. The system was tested on data from 2020 and 2023, showing that the model was robust across different news topics. Moderator evaluations confirmed that the system often suggested comments that would be featured, though variation among moderators remained.
Chapter 6 investigates whether featuring individual comments improved discussion quality or participation. The study compared discussions with and without highlighted comments. While the presence of highlighted comments did not improve overall quality (measured by likes and flagged comments), there was a potential reduction in activity in discussions where moderators featured comments. Further research is needed to fully understand this effect.
This thesis highlights the importance of considering multiple perspectives when defining constructive discussion. Computational tools can support moderators by pre-selecting comments, allowing them to focus on contextual factors that influence what is deemed constructive. The combination of computational models and fieldwork ensures that these tools are adapted to the complex realities of online moderation on news platforms, with moderators ultimately deciding what counts as constructive.
show less
