Mb-courage @ exist: Gcn classification for sexism identification in social networks?

We describe our approach (MB-Courage team) in the Sexism Identification in Social Networks Shared Task (EXIST). We submitted three runs for each task, two of them based on Graph Convolutional Neural Networks (GCN) exploring different edge creation strategies and one combining graph embeddings from different GCN through ensemble methods. In addition, we explored different GCN models and text-to- graph strategies. We identified that in Task 2 the models take advantage of the syntactic relationship between words encoded in the graph, while it did not strongly impact Task 1. Moreover, the models generalized the task while maintaining similar (in some cases better) results in the social network that was not used in training. On average, our best models performed similarly across languages and social media, ranking 37th (out of 72 runs) for Task 1 and 40th (out of 63) for Task 2.