Combining BERT with Contextual Linguistic Features for Identification of Propaganda Spans in News Articles

摘要

Recent endeavours at detection of propaganda in news articles treat this as a fine-grained problem of detecting it within fragments; and hence, transformer based embeddings perform decently in such detection. We build our propaganda detection framework on top of a transformer model simultaneously enriching it with contextual linguistic information of surrounding part-of-speech tags and LIWC categories the word itself belongs to. The evaluation outcomes being encouraging indicate a huge potential for this line of reasoning in natural language processing of news text.