Coordination between contributors in online communities, is known to rely on several concurrent modalities, including both code- or artifact-mediated interactions and direct dialectic exchange through forums, issues trackers or else instant messaging tools. Even though the latter form has been subject to various inquiries, direct approaches that would help communities and moderators distinguish between gossip exchanges and serious debates - notably in order to avoid "town council" discussions, as criticized by Cox - are still largely missing. However, although discussion topics vary, any discussion shape, seen as a graph, is typically a tree, regardless of its theme. Using this graph-theoretic framework, we attempt to identify categories of discussions taking into account not only the properties of underlying trees, issued from the graph theory, but also the expertise status of the participants and the relationships existing between them as well. We establish these relationships by detecting the direction of exchanges between two contributors and model them through an author graph, as the ones employed in studies of social networks. Furthermore, the motifs - triads - in author graphs together with quantitative data derived from discussion trees and status analysis allows us to determine a classification of discussions from a benchmark Reddit discussion forum. We apply a clustering algorithm based on the author networks of threads, which highlights tree distinct groups. To better understand the dynamics of those classes, we implement the relational event model, which emphasizes three effects, which influence in different ways the groups. The result obtained may possibly be used in decision-making by forum moderators and communities.