Despite their recent success, the geospatial reasoning capabilities of large language models (LLMs)—which require understanding spatial relationships among real-world geo-entities—remain underexplored.We propose an automatic method for constructing compositional geographic question answering datasets that jointly consider spatial and entity constraints.The generated dataset serves as a principled benchmark for evaluating how LLMs coordinate spatial computation with entity-level understanding under diverse compositional settings.We evaluate two state-of-the-art LLMs, GPT-5.2 and Gemini 3 Flash, on our dataset. Experimental results show that while the models perform relatively well on questions involving rich entity grounding, their accuracy drops substantially on questions requiring precise quantitative spatial reasoning, such as distance estimation and containment judgment.Our dataset is publicly available for research and reproduction.

Linking-based geocoding is the task of linking location mentions in text to their corresponding entries in a geographic database (Geo-DB) and assigning precise coordinates. Although the task and its technology are essential for spatial information extraction, existing datasets are manually curated and lack sufficient data for training accurate models. To address this limitation, we automatically construct a large-scale dataset for linking-based geocoding by leveraging publicly available resources to generate data efficiently at scale. Specifically, we align location mentions in the first paragraphs of Japanese Wikipedia articles with their associated Wikidata entries containing geographic attributes. Wikipedia provides natural textual contexts, while Wikidata offers structured data such as coordinates, place types, and administrative divisions, which can serve as rich metadata for future extensions. Our experiments show that models trained on our dataset achieve strong performance not only on in-domain data, i.e., Wikipedia, but also on out-of-domain newspaper articles, and further confirm that hard negative mining substantially improves disambiguation among confusable candidates. Although the dataset focuses on Japanese, the construction method is language-agnostic and can be extended to other languages with sufficient Wikipedia and Wikidata coverage.