Geoparser: A transformer-based bi-encoder approach for efficient toponym disambiguation

Information retrieval systems face significant challenges when interpreting toponyms in unstructured text due to their inherent ambiguity and context dependency. Traditional methods often struggle with the linguistic complexity involved, making it difficult to resolve these ambiguities effectively. While transformer models offer advanced linguistic capabilities, they are computationally expensive, and machine learning models, in general, often face difficulties in generalising across different domains. This thesis addresses these challenges by exploring how transformer models can be efficiently and effectively applied to toponym resolution and how their capabilities transfer across domains. A new method is proposed that uses a bi-encoder architecture within the SentenceTransformers framework to efficiently compare contextualised toponyms with potential location candidates from a gazetteer. This approach reduces computational demands by encoding toponyms and candidates separately, allowing for scalable similarity comparisons. The method is integrated into an end-to-end geoparsing pipeline through the development of the Geoparser Python library, which leverages spaCy for toponym recognition and provides functionalities for customisation and adaptation to specific text corpora. Experiments were conducted replicating a standardised evaluation framework to assess the performance of the proposed method. The results demonstrate that Geoparser achieves competitive performance compared to state-of-the-art systems, particularly excelling in computational efficiency. Further experiments show that the transformer model’s performance can be enhanced for different domains with minimal additional training data. This work highlights the potential of transformer models for efficient and effective toponym resolution, offering promising directions for future research in Geographic Information Retrieval.