An ESTs detection research based on paper entity mapping: Combining scientific text modeling and neural prophet

Existing studies on the detection of emerging scientific topics (ESTs) overemphasize the newness and neglect content innovation of knowledge. Moreover, they also ignore the lag existing in knowledge diffusion. In this paper, we propose a four-stage detection framework for ESTs that maps emerging attributes from paper entities to scientific topics. Empirical studies based on two significantly different disciplinary datasets, IS-LS, and AI, which contain 73,601 and 255,620 publications, respectively, are employed to validate our approach. First, we generate 29 and 47 candidate scientific topics based on topic modeling, respectively. Second, we represent the novelty of paper entities based on pre-trained language models, which is mapped to scientific topic entities along with knowledge distributions to obtain topic emerging attributes: topic novelty, relative share and growth. Third, we propose to predict future trends of these attributes with Neural Prophet, which outperforms four baseline models in $R^2$, $MAE$ and $RMSE$. Finally, combining future values of candidate scientific topics, they are grouped into 8 clusters containing two ESTs types through strategic market theory and clustering model. From the correlation and feature distribution analysis of emerging attributes, we discover the existence of resilience and scale advantage in the diffusion of scientific knowledge. There also exists significant uncertainty in previous citation-based scientific topic evaluation patterns caused by the complexity of citation behavior. Overall, this research enriches theoretical knowledge and detection frameworks of ESTs, and provides detailed insights into comprehensive assessment and dissemination of scientific topics.