Evaluation criteria for discourse relations semantic affinity

The paper presents an overview of structured definitions of discourse relations created based on classification principles and criteria for evaluating their semantic affinity. The authors point out the shortcomings of existing classification approaches that are sometimes inconsistent or contradictory and outline the benefits of an alternative approach to classification of discourse relations which is based on their structured definition. The paper provides examples of such definitions created within the Supracorpora Database of Connectives and discusses the criteria for evaluating their semantic closeness. As the structured definitions are represented by sets of distinguishing features, the authors discuss the problem of identifying proximity factors for each of these features. The gathered data suggest a hypothesis that among the three groups of features: “Level,” “Basic operation,” and “Feature family”, it is the last one that should have the most impact. Finally, directions for further research of this problem are considered, namely, the option of taking into account such factors as compatibility of discourse relations, correspondence of relations between the source text and its translation, and such cases where certain relation markers may express different discourse relations in various contexts.