Incorporating Lexical Knowledge via WordNet to Latent Dirichlet Allocation in Offensive Message Detection

摘要

We propose a model to offensive messages detection for political discourse that combines topic modeling and lexicon-based approaches for knowledge extraction. We develop an extension to the LDA suitable for offensive message detection by leveraging on lexical and semantic word features. Our model employs an externally supplied lexicon and WordNet, a lexical database, to incorporate prior knowledge to the LDA. At the document-level, we model the semantic relationship between a limited list of concepts with political orientation and corpus-determined themes. At the topic-level, we incorporate lexical word prior based on the WordNet lexical relationship between an externally supplied list of offensive words and topics generated from the corpus. Our model presumes a set of preselected labels that document themes should fit. We test our model against different sets of datasets and compare its performance against several baselines. The experiments confirm the effectiveness of our approach in both prediction and classification tasks.