Stay on Topic, Please: Aligning User Comments to the Content of a News Article

Social scientists have shown that up to $$50\%$$ of the comments posted a news article no relation its journalistic content. In this study we propose classification algorithm categorize user based on their alignment The seeks match an similarity content, entities in discussion, and topics. We BERTAC, BERT-based approach learns jointly article-comment embeddings infers relevance class comments. introduce ordinal loss penalizes difference between predicted true labels. conduct thorough show influence proposed learning process. results five representative outlets our can learn comment with $$36\%$$ average accuracy improvement comparing baselines, $$25\%$$ BA-BC. BA-BC is consists two models aimed capture dis-jointly formal language articles informal also evaluate human labeling performance understand difficulty task. agreement comment-article “moderate” per Krippendorff’s alpha score, which suggests task difficult.