TempReasoner: neural temporal graph networks for event timeline construction

Constructing event timelines from unstructured temporal data is a fundamental challenge for knowledge extraction and reasoning systems. Existing temporal reasoning methods face challenges in jointly modelling fine-grained temporal dependencies, sparse event interactions, and maintaining coherent causal orderings across multi-domain datasets. In this paper, the author proposes TempReasoner, a new neural temporal graph network, based on dynamic spatio-temporal attention and reinforced temporal reasoning, an architecture that builds automated event timelines. We combine temporal knowledge graphs with adaptive graph neural networks and a multi-scale temporal attention model that jointly represent local event dependencies and global temporal patterns. The proposed system uses a hierarchical temporal encoder with gated recurrent units and introduces a new temporal consistency loss to maintain temporal coherence. Comprehensive testing on five benchmark datasets shows that TempReasoner achieves 94.3% accuracy in ordering event timelines and operates in real time with an average latency of 127 ms per event sequence. The system performs well across various areas, such as legal investigations, news analysis, and tracking biomedical events, and can therefore be easily integrated into enterprise applications.