A Hierarchical Mining-Based Approach to Storyline Generation

Events serve as the fundamental units of human society. With the rapid proliferation of information networks and social media, the dissemination of event-related information has become more pervasive. However, this has also introduced challenges such as information fragmentation, redundancy, and noise, which hinder timely and comprehensive understanding of complex events. Recent research has focused on constructing coherent event narratives by organizing discrete pieces of event information into structured and logically connected developmental chains—a process commonly referred to as storyline generation. Despite notable progress, most existing approaches rely on single-dimensional associations between events and often require extensive multifaceted knowledge, limiting their scalability and effectiveness in handling complex, multifaceted event scenarios. To address this issue, we propose a hierarchical mining-based approach to storyline generation (HMSG). HMSG comprises three core modules. The data preprocessing algorithm is designed to extract and represent textual event features. The coarse-grained clustering algorithm identifies events in the form of story clusters, where events within the same story cluster belong to the same storyline. The fine-grained clustering algorithm divides each story cluster into smaller and semantically cohesive event clusters by capturing stronger local correlations among events. Subsequently, story clusters and event clusters are analyzed to construct the storyline structure, where the main narrative is represented as a trunk and subnarratives as branches. This process culminates in a tree-like representation that effectively illustrates the storyline. To validate the effectiveness of the proposed approach, we conduct extensive experiments on multiple real-world datasets from different domains. Comparative analysis against several state-of-the-art methods for storyline generation demonstrates that HMSG consistently achieves superior performance and robustness across diverse scenarios. The results further confirm that HMSG outperforms baseline approaches in both performance and robustness.