Graph Knowledge Transfer for Offensive Language Identification with Graph Neural Networks

Identifying offensive language (OL) has become ever more important with the rise of online social media (OSM). Most works on OL identification have applied sequential models to learn offensive semantics. In a different light, recent popular graph neural networks (GNNs) model text in a word graph and learn local word-level usages for each document. This work aims to explore text GNNs on learning OL usages in a non-sequential view yet there are two barriers: (1) Similar to sequential models, GNNs also suffer from the out-of-vocabulary (OOV) issue due to the informal usages of OSM texts. (2) Lacking of appropriate edge weights derived from short content to the word graph. Motivated by the rich resources of OSM, this work leverages existing OL corpus to build a global reference graph and proposes a graph knowledge transfer mechanism (GKtran) to tackle the limited extent of learning from short texts. To embed OL knowledge, GKtran infers edge weights from a reference graph and transfers weights to each document's word graph. Edges that connect OOV words are assigned weights through weight propagation. Thus, each document graph is fully weighted for graph classification. Experimental results show that the additional information from training or external data transferred by GKtran effectively improves OL identification with graphs.