Generation of News Articles from Tweets : An Experiment

Abstract

With millions of users, Twitter is among the most popular social media platforms today, making it one of the largest sources of continuous real-time information on the Internet. Users frequently broadcast events occurring across the world on this platform, providing a scope for the production of news articles by aggregating event-specific tweets. This paper contributes to this task and evaluates the existing state-of-the-art abstractive summarization models to see how efficient or capable they are when it comes to the task of generating news articles from tweets. We compare summaries generated by five such models qualitatively in order to see how concise or coherent they are, whether they can capture the important facts of the entire tweet dataset relating to a topic and how similar they are to form news articles written by humans. PEGASUS (Pre-training with Extracted Gap-Sentences for Abstractive SUmmarization Sequence-to-sequence models) pretrained with CNN/Daily mail dataset outperformed over T5 (Text-to-Text Transfer Transformer). Manual evaluation shows that the generated summaries are concise and cohesive.

Tasks

Reproductions