Document Classification via Stable Graph Patterns and Conceptual AMR Graphs

This paper proposes an approach and an associated system based on pattern structures, aimed at the classification of documents represented as graphs. The representation of documents relies on Abstract Meaning Representation (AMR) document graphs. Given a set of AMR document graphs, the system learns characteristic graph patterns, that can be reused by an aggregate rule classifier to predict the class of a document. The selection of the most stable graph patterns is based on the gSOFIA algorithm and the \(\varDelta -\) stability measure. In the experiments, two document datasets are considered for validating the approach. The first includes documents belonging to 10 different newsgroups and the second contains sports news articles belonging to 5 topical areas. The results in terms of the macro-averaged \(F_1\) scores, are quite satisfactory and show that the approach is well-founded and useful.