Improving Toponym Resolution with Better Candidate Generation, Transformer-based Reranking, and Two-Stage Resolution

Proceedings of the The 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 48–60 July 13-14, 2023 ©2023 Association for Computational Linguistics Improving Toponym Resolution with Better Candidate Generation, Transformer-based Reranking, and Two-Stage Resolution Zeyu Zhang and Steven Bethard School of Information, The University of Arizona, Tucson, AZ, USA {zeyuzhang, bethard}@arizona.edu Abstract Geocoding is the task of converting location mentions in text into structured data that en- codes the geospatial semantics. We propose a new architecture for geocoding, GeoNorm. GeoNorm first uses information retrieval tech- niques to generate a list of candidate entries from the geospatial ontology. Then it reranks the candidate entries using a transformer-based neural network that incorporates information from the ontology such as the entry’s popula- tion. This generate-and-rerank process is ap- plied twice: first to resolve the less ambigu- ous countries, states, and counties, and sec- ond to resolve the remaining location men- tions, using the identified countries, states, and counties as context. Our proposed toponym resolution framework achieves state-of-the-art performance on multiple datasets. Code and models are available at https://github. com/clulab/geonorm. 1 Introduction Geospatial information extraction is a type of se- mantic extraction that plays a critical role in tasks such as geographical document classification and retrieval (Bhargava et al., 2017), historical event analysis based on location data (Tateosian et al., 2017), tracking the evolution and emergence of infectious diseases (Hay et al., 2013), and disas- ter response mechanisms (Ashktorab et al., 2014; de Bruijn et al., 2018). Such information extraction can be challenging because different geographical locations can be referred to by the same place name (e.g., San Jose in Costa Rica vs. San Jose in Cali- fornia, USA), and different place names can refer to the same geographical location (e.g., Leeuwar- den and Ljouwert are two names for the same city in the Netherlands). It is thus critical to resolve these place names by linking them with their cor- responding coordinates from a geospatial ontology or knowledge base. Geocoding, also called toponym resolution or toponym disambiguation, is the subtask of geop- arsing that disambiguates place names (known as toponyms) in text. Given a textual mention of a location, a geocoder chooses the corresponding geospatial coordinates, geospatial polygon, or en- try in a geospatial database. Approaches to geocod- ing include generate-and-rank systems that first use information retrieval systems to generate can- didate entries and then rerank them with hand- engineered heuristics and/or supervised classifiers (e.g., Grover et al., 2010; Speriosu and Baldridge, 2013; Wang et al., 2019), vector-space systems that use deep neural networks to encode place names and database entries as vectors and measure their similarity (e.g., Hosseini et al., 2020; Ardanuy et al., 2020), and tile-classification systems that use deep neural networks to directly predict small tiles of the map rather than ontology entries (e.g., Gritta et al., 2018a; Cardoso et al., 2019; Kulkarni et al., 2021). The deep neural network tile-classification approaches have been the most successful, but they do not naturally produce an ontology entry, which contains semantic …

Content truncated to 500 words