Elamur Research⌕
☰
home / papers / 2026 / A Large-Scale Dataset for Linking-Based Geocoding

A Large-Scale Dataset for Linking-Based Geocoding

Summary

Linking-based geocoding maps a place name in text to a specific entry in a geographic database, which yields both coordinates and attributes. Japanese lacks training data for this task: the existing resource, ATD-MCL, has only about 6,000 manually annotated mentions. Nakatani et al. address this by building a training corpus automatically. They take hyperlinks in the opening paragraphs of Japanese Wikipedia articles and keep those that point to Wikidata items with geographic coordinates. Each linked span becomes a labelled mention, so no human annotation is needed. The resulting dataset has more than 2.4 million mentions, covering places from countries down to schools and stations. The authors trained lexical (BM25) and dense (E5) retrievers on it. The models performed well on held-out Wikipedia text and also transferred to Nikkei newspaper articles. Performance grew with training data size, and hard negative mining clearly helped separate confusable candidates. The pipeline is language-independent, so it can be reproduced for any language with good Wikipedia and Wikidata coverage, including Ukrainian.

Key points

  • Scale: 2,423,134 mentions in total, split into 1,938,507 train, 242,313 dev and 242,314 test mentions. Earlier linking datasets (GeoWebNews, TR-News, LGL, ATD-MCL) each contain between about 1,300 and 6,100 mentions.
  • Geo-DB: It was filtered from the Wikidata dump of September 2, 2024, which has 112,413,055 entities. Of these, 3,496,520 have a Japanese label and 435,867 have coordinates (property P625). An entity is kept only if it has both a ja label and P625.
  • Tooling: The pipeline uses the build-dump-db command of Wikipedia2Vec for text and links. It uses kwnlp-sql-parser on the page and page_props SQL dumps to map page titles to Wikidata QIDs.
  • Evaluation: Baselines are BM25 and multilingual E5. The out-of-domain test is small (28 Nikkei articles), and the authors list this as a limitation. Hard negative mining improved top-1 accuracy and cross-domain robustness.
  • Availability: The dataset is released at github.com/naist-nlp/wiki-geocoding-dataset.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/2pv6oidqzqs9

@inproceedings{nakatani2026large,
  title = {{A Large-Scale Dataset for Linking-Based Geocoding}},
  author = {Nakatani, Hibiki and Yasui, Yuichiro and Wakamoto, Ryosuke and Ishii, Masayuki and Suizu, Tetsuhisa and Ouchi, Hiroki and Watanabe, Taro},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/2pv6oidqzqs9},
  url = {https://doi.org/10.63317/2pv6oidqzqs9}
}

References

21 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Diego Gomes, Ross S Purves, and Michele Volpi. 2024. Fine-tuning Transformers for toponym resolution: A contextual embedding approach to candidate ranking. In Proceedings of The GeoExT 2024: Geographic Information Extraction from Texts Workshop, pages 43–51. find
  2. Milan Gritta, Mohammad Taher Pilehvar, and Nigel Collier. 2018. Which Melbourne? Augmenting geocoding with maps. In Proceedings of ACL, pages 1285–1296. find
  3. Milan Gritta, Mohammad Taher Pilehvar, and Nigel Collier. 2020. A pragmatic guide to geoparsing evaluation. Language Resources and Evaluation, 54(3):683–712. find
  4. Andrew Halterman. 2023. neural geoparser and arXiv:2303.13675. — [arXiv:2303.13675](https://arxiv.org/abs/2303.13675) find
  5. Shohei Higashiyama, Masao Ideuchi, and Masao Utiyama. 2024a. Construction of the administrative agency web document corpus for Japanese entity linking [in Japanese]. IPSJ SIG Technical Report, 2024-NL-260(10):1–15. find
  6. Shohei Higashiyama, Hiroki Ouchi, Hiroki Teranishi, Hiroyuki Otomo, Yusuke Ide, Aitaro Yamamoto, Hiroyuki Shindo, Yuki Matsuda, Shoko Wakamiya, Naoya Inoue, Ikuya Yamada, and Taro Watanabe. 2024b. Arukikata travelogue dataset with geographic entity mention, coreference, and link annotation. In Findi… find
  7. Xuke Hu, Zhiyong Zhou, Hao Li, Yingjie Hu, Fuqiang Gu, Jens Kersten, Hongchao Fan, and Friederike Klan. 2022. Location reference recognition from texts: A survey and comparison. arXiv:2207.01683. — [arXiv:2207.01683](https://arxiv.org/abs/2207.01683) find
  8. Jizhou Huang, Haifeng Wang, Yibo Sun, Yunsheng Shi, Zhengjie Huang, An Zhuo, and Shikun Feng. 2022. ERNIE-GeoL: A geography-and-language pre-trained model and its applications in Baidu maps. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’22, pages 3029–… find
  9. Ehsan Kamalloo and Davood Rafiei. 2018. A coherent unsupervised model for toponym resolution. In Proceedings of the 2018 World Wide Web Conference, WWW ’18, page 1287–1296, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee. find
  10. Sayali Kulkarni, Shailee Jain, Mohammad Javad Hosseini, Jason Baldridge, Eugene Ie, and Li Zhang. 2021. Multi-level gazetteer-free geocoding. In Proceedings of Second International Combined Workshop on Spatial Language Understanding and Grounded Communication for Robotics, pages 79–88, Online. Assoc… find
  11. Zekun Li, Wenxuan Zhou, Yao-Yi Chiang, and Muhao Chen. 2023. GeoLM: Empowering language models for geospatially grounded language understanding. In Proceedings of EMNLP, pages 5227–5240. find
  12. Michael D. Lieberman, Hanan Samet, and Jagan Sankaranarayanan. 2010. Geotagging with local lexicons to build indexes for textually-specified spatial data. In 2010 IEEE 26th International Conference on Data Engineering, pages 201– 212. IEEE. find
  13. Lajanugen Logeswaran, Ming-Wei Chang, Kenton Lee, Kristina Toutanova, Jacob Devlin, and Honglak Lee. 2019. Zero-shot entity linking by reading entity descriptions. In Proceedings of ACL, pages 3449–3460, Florence, Italy. Association for Computational Linguistics. find
  14. Xing Han Lù. 2024. BM25S: Orders of magnitude faster lexical search via eager sparse scoring. arXiv:2407.03618. — [arXiv:2407.03618](https://arxiv.org/abs/2407.03618) find
  15. Hibiki Nakatani, Hiroki Teranishi, Shohei Higashiyama, Yuya Sawada, Hiroki Ouchi, and Taro Watanabe. 2025. A Text Embedding Model with Contrastive Example Mining for Point-of-Interest Geocoding. In Proceedings of COLING. find
  16. Keyaki Ohno, Hirotaka Kameko, Keisuke Shirai, Taichi Nishimura, and Shinsuke Mori. 2024. Automatic construction of a large-scale corpus for geoparsing using Wikipedia hyperlinks. In Proceedings of LREC-COLING, pages 1883–1888, Torino, Italia. ELRA and ICCL. find
  17. Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. 1995. Okapi at TREC-3. Nist Special Publication Sp, 109:109. find our card
  18. Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2024a. Text embeddings by weakly-supervised contrastive pre-training. arXiv:2212.03533. find our card
  19. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024b. Multilingual E5 text embeddings: A technical report. arXiv:2402.05672. find our card
  20. Ikuya Yamada, Akari Asai, Jin Sakuma, Hiroyuki Shindo, Hideaki Takeda, Yoshiyasu Takefuji, and Yuji Matsumoto. 2020. Wikipedia2Vec: An efficient toolkit for learning and visualizing the embeddings of words and entities from Wikipedia. In Proceedings of EMNLP, pages 23–30. Association for Computation… find
  21. Zeyu Zhang, Egoitz Laparra, and Steven Bethard. 2024. Improving toponym resolution by predicting attributes to constrain geographical ontology entries. In Proceedings of NAACL, pages 35–44, Mexico City, Mexico. Association for Computational Linguistics. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nakatani-2026-large-scale-dataset-linking-based/