Abstract
Scientific articles have semantic contents that are usually quite specific to their disciplinary origins. To characterize such semantic contents, topic-modeling algorithms make it possible to identify topics that run throughout corpora. However, they remain limited when it comes to investigating the extent to which topics are jointly used together in specific documents and form particular associative patterns. Here, we propose to characterize such patterns through the identification of "topic associative rules" that describe how topics are associated within given sets of documents. As a case study, we use a corpus from a subfield of the humanities-the philosophy of science-consisting of the complete full-text content of one of its main journals: Philosophy of Science. On the basis of a pre-existing topic modeling, we develop a methodology with which we infer a set of 96 topic associative rules that characterize specific types of articles depending on how these articles combine topics in peculiar patterns. Such rules offer a finer-grained window onto the semantic content of the corpus and can be interpreted as "topical recipes" for distinct types of philosophy of science articles. Examining rule networks and rule predictive success for different article types, we find a positive correlation between topological features of rule networks (connectivity) and the reliability of rule predictions (as summarized by the F-measure). Topic associative rules thereby not only contribute to characterizing the semantic contents of corpora at a finer granularity than topic modeling, but may also help to classify documents or identify document types, for instance to improve natural language generation processes.