Storyline Generation from News Articles Based on Approximate Personalized Propagation of Neural Predictions

Junli Wang, Xujian Zhao, Peiquan Jin, Chunming Yang, Bo Li, Hui Zhang

Abstract: Generating a storyline is aiming to discover the evolution of events from news websites. Some existing approaches aim to automatically cluster news articles into events and connect related events in growing trees to generate storylines. Unfortunately, these methods did not perform well in learning the implicit associations of events. More recently, Graph Convolutional Network (GCN) based methods are proposed to learn the implicit associations between events. However, since the event representation in GCN model tends to be consistent after multi-layer propagation, the events cannot be correctly distinguished, which is not conducive to comprehensively learning the implicit associations between different events. In this paper, we propose an effective storyline generation method for news articles. Firstly, a novel model is presented based on Approximate Personalized Propagation of Neural Predictions for Story Branch Construction Model, called SBCM, which preserves local features and can better learn the implicit association between different events. Then, we utilize a statistical method to identify transition events in news articles, and connect story branches with transition events through temporal relationships to finally generate storylines. The experimental results on two real-world Chinese news datasets show that our proposal outperforms several state-of-the-art methods.