A Hybrid Model for Spatio-Temporal Information Recognition in COVID-19 Trajectory Text

Haoyu Yu, Xuan Pan, Dongming Zhao, Yanlong Wen, Xiaojie Yuan

Abstract: Since the outbreak of the COVID-19 epidemic at the end of 2019, the normalization of epidemic prevention and control has become one of the core tasks of the entire country. Health self-examination by checking the trajectory of diagnosed patients has gradually become everyone’s basic necessity and essential to epidemic prevention. The COVID-19 patient’s spatio-temporal information helps to facilitate the self-inspection of the masses of whether their trajectory overlaps with the confirmed cases, which promotes the epidemic prevention work. This paper, proposes a named entity recognition model to automatically identify the time and place information in the COVID-19 patient trajectory text. The model consists of an ALBERT layer, a Bi-GRU layer, and a GlobalPointer layer. The previous two layers jointly focus on extracting the context’s characteristics and the semantic dependencies. And the GlobalPointer layer extracts the corresponding named entities from a global perspective, which improves the recognition ability for the long-nested place and time entities. Compared to the conventional name entity recognition models, our proposed model has high effectiveness because it has a smaller parameter scale and faster training speed. We evaluate the proposed model using a dataset crawled from the official COVID-19 trajectory text. The F1-score of the model has reached 92.86%, which outperforms four traditional named entity recognition models.