Abstractive Event Summarization on Twitter

Quanzhi Li, Qiong Zhang

Abstract: This paper presents a new approach for automatically summarizing a social media event. It utilizes the BERT model as the encoder and a Transformer architecture as the decoder. The framework also includes an event topic prediction component, and the predicted event topic will help the decoder focus more on the specific aspects of the topic category when generating summary. To make the summary more succinct and coherent, the most important messages from an event cluster are selected by a message selection model and encoded by the BERT model. Our preliminary experiment shows that our approach outperforms the baseline methods.