GeoTPE: A neural network model for geographical topic phrases extraction from literature based on BERT enhanced with relative position embedding

Geographical Topic Phrases (GTPs) are specialized terms for describing geographical objects, phenomena, or events and are frequently used to organize, navigate, and index geographical resources (e.g., geographical data hosted in geoportals). Typically, GTPs are stored in knowledge bases (e.g., a thesaurus). However, most existing knowledge bases are manually constructed and often updated on an annual or decennial cycle, leading to the exclusion of many newly emerging GTPs. These emerging GTPs are often discussed in geographical literature. Therefore, there is an urgent need for a method to automatically extract out-of-vocabulary GTPs from geographical literature to either create a new knowledge base or automatically update existing ones. The state-ofthe-art GTPs extraction approaches are deep learning-based models. The existing ones, however, did not consider the relative distance between vocabularies, leading to their limited capability of capturing and learning relationships between words in a sequence. In this work, we present GeoTPE, a neural network model fusing BiLSTM-CRF and BERT enhanced with relative position embedding for extracting GTPs from literature, and evaluate this model by applying it to two datasets, including a dataset harvested from geographical literature of high-ranked journals. The experimental results show that our model can not only achieve the best performance in comparison with baseline models, but also can discover novel GTPs, thus enriching existing knowledge bases.