Neural Abstractive Unsupervised Summarization of Online News Discussions

Summarization has usually relied on gold standard summaries to train extractive or abstractive models. Social media brings a hurdle summarization techniques since it requires addressing multi-document multi-author approach. We address this challenging task by introducing novel method that generates of online news discussions. Our extends BERT-based architecture, including an attention encoding fed comments’ likes during the training stage. To our model, we define which consists reconstructing high impact comments based popularity (likes). Accordingly, model learns summarize discussions their most relevant comments. approach provides summary represents aspects item users comment on, incorporating social context as source information texts in networks. is evaluated using ROUGE scores between generated and each thread. encoding, significantly outperforms both methods such evaluation.