Entity-biased Multi-Source Pointer Generator For Article Event Summarization

With the information explosion on the web and the internet,search engines and news systems are requested to grasp the main events from redundant articles and generate more refined event description for guideline and reading.Summarizing the main event from long articles is of great value to various real-world application and crucial to reading experience.Therefore,we propose to address a challenging task called Article Event Summarization(AES),which aims to generate concise event summaries for articles from multiple sources in open domain.Comparing to conventional summarization tasks,as all key event elements need to be correctly detected and properly organized,it is in multiple,unbalanced source setting,and open-domain online articles are usually of diverse topics,various styles and ambiguous descriptions.To address these issues,we propose a Multi-Source Entity-biased Pointer Generator to summarize key event elements and integrate multiple source knowledge via an attentive fusion mechanism,and further leverage user search log data in a pretrain manner to determine high salience elements in open domain.We also create two large-scale datasets with manual-written summaries to facilitate future research.Extensive evaluations,including automatic evaluation and manual evaluation,demonstrate the significant superiority of the proposed method over a wide range of related baselines.Lastly,the proposed AES scheme has been deployed in the mobile social application for generating hotspot recommendation and query suggestion,which has over one billion of active daily users.This work helps a lot in alleviating the laboriousness in generating event summaries and improving the user clicks.