Introduction
Geocoding, or toponym resolution, is the critical task of mapping location mentions within text to specific entries in a geospatial ontology, such as GeoNames. This process is complicated by linguistic ambiguities: polysemy, where a single name like "Paris" refers to multiple distinct locations (e.g., France vs. Texas), and synonymy, where one location has multiple names. While neural models have improved performance, they often struggle to capture the implicit geographic context provided by surrounding mentions. This paper introduces GeoPLACE (GEOgraphical normalization by Predicting Location Attributes to Constrain ontology Entries), a "predict-then-constrain" framework that leverages transformer-based attribute prediction to resolve these ambiguities.
Methodology
GeoPLACE decomposes the toponym resolution task into three modular stages: attribute prediction, candidate generation, and constrained re-ranking.
1. Attribute Predictor
Rather than directly selecting a database entry, GeoPLACE uses a BERT-base transformer to predict latent geographic attributes. Using a novel prompt—[CLS] This document discusses these location mentions: [M1, M2...] in which [Target] is [MASK] located in [MASK] of [MASK] [SEP]—the model predicts the Feature Class (e.g., populated place), State, and Country. Crucially, this uses a Masked Language Modeling (MLM) objective, treating attributes as single tokens to avoid the hallucinations common in sequence-to-sequence models.
2. Candidate Generator
A Lucene-based search engine retrieves potential matches from the GeoNames ontology (containing ~7 million entries) based on string similarity to the mention.
3. The Constrainer
The final stage is a deterministic algorithm that re-ranks candidates based on their alignment with the predicted attributes. If the predictor identifies "Texas" and "USA" for a mention of "Paris," the Constrainer de-prioritizes the French capital in favor of the Texan town.
Experimental Setup
The model was evaluated on three benchmark datasets: LGL (Local Global Lexicon), GeoWebNews, and TR-News. To enhance performance, the authors employed synthetic pre-training, using GeoNames data to teach the model geographic hierarchies before fine-tuning on news corpora.
Results and Discussion
GeoPLACE established a new state-of-the-art across all tested benchmarks. As shown in the performance table below, the model significantly outperformed previous leaders like GeoNorm and ReFinED.
On the LGL dataset, accuracy improved from 0.807 to 0.863, while TR-News accuracy reached 0.947. Most notably, GeoPLACE achieved a 30% reduction in mean error distance on LGL and TR-News. These gains are attributed to the model's ability to infer a document's "geographic center of gravity" through implicit context. For example, the presence of "Dallas" in a text helps the model correctly predict "Texas" for a nearby "Paris" mention, even if the state is not explicitly named.
Conclusion
GeoPLACE demonstrates that predicting structured geographic attributes is a more robust strategy for toponym resolution than end-to-end neural ranking. By transforming a fuzzy search problem into a constrained filtering task, the system achieves superior accuracy and interpretability. Future research will explore the sensitivity of the model to different prompt structures and its applicability to non-English texts and informal domains like social media.
Summary updated