UniTopRank: a scalable and language-independent method for toponym resolution

Kurzfassung

Toponym resolution is essential for extracting geographic informa tion from unstructured texts. Although recent deep learning and large language model (LLM)–based methods achieve strong per formance, they typically require substantial computational resour ces and have slow inference speeds, limiting their practicality for resource-constrained, large-scale or time-sensitive applications. Moreover, most existing approaches are developed for English or a limited set of languages, restricting their cross-lingual applicability. To fill the gaps, we propose UniTopRank, a universal, scalable, and language-independent method based on rule-driven ranking. It operates in two steps: first, candidate locations for all toponyms in a text are retrieved from gazetteers; second, the candidates are ranked using name similarity, population, administrative level, and spatial relationships among toponyms through beam search. We evaluate UniTopRank on 25 datasets in 14 languages and 7 text types, covering 186,100 toponyms worldwide. For 11 English data sets, we compare it with 21 existing methods, including deep learn ing– and LLM-based models, and for non-English datasets, with 12 representative approaches. Results show that UniTopRank attains accuracy comparable to state-of-the-art methods—though not the best—and generalizes well across different languages, while being significantly faster and much more resource-efficient. Running entirely on standard CPUs, it is well-suited for large-scale, multilin gual, or resource-constrained applications.