DePNR: A DeBERTa‐based deep learning model with complete position embedding for place name recognition from geographical literature

Abstract

AbstractPlace names play an important role in linking physical places to human perception and are highly frequently used in the daily lives of people to refer to places in natural language. However, many place names may not be recorded in typical gazetteers due to their new establishment, colloquial nature, and different concerns. These unrecorded toponyms are often discussed in geographical literature; thus, it is necessary to automatically identify them from geographical literature and update existing gazetteers using computational approaches. Currently, the most advanced approaches are deep learning‐based models. However, existing models used only partial position information rather than complete position information of words in a sentence, which limits their performance in recognizing toponyms. To this end, we develop DePNR, a DeBERTa‐based deep learning model with complete position embedding for place name recognition from geographical literature. We train DePNR on two datasets and test it on a real dataset from geographical literature to evaluate its performance. The results show that DePNR achieves an F‐score of 0.8282, outperforming previous approaches, and can recognize new toponyms from literature text, potentially enriching existing gazetteers.