nlpir2025 Entity Linking for Geographical Mentions Using Address Hierarchy

Transcript

1. Entity Linking for Geographical Mentions Using Address Hierarchy University of Tsukuba Name: Takeru Mitsumori, Takashi Inui December 14, 2025 Session 7 KS011 1 NLPIR 2025
2. Entity Linking (EL) 2
3. Research Background and Purpose ⚫ Document geolocation ⚫ Estimate the posting location of social media ⚫ Applying EL to mentions ⚫ Utilizing entity information ⚫ EL for geographical mention is limited ⚫ Geographical mentions ⚫ Mentions with geographical attributes ⚫ e.g. Japan, Tokyo, Kyushu University ⚫ Propose an entity disambiguation method specialized for geographical mentions in Japanese text 3
4. Related Work ⚫ [Leidner 2004] ⚫ Toponym Resolution (Place names in text → real world locations) ⚫ Disambiguation using geographical distance ⚫ [Yamada+ 2022] ⚫ Deep learning-based entity disambiguation model ⚫ Model corresponding for general tasks 4 [Leidner 2004] Leidner, Jochen L. "Toponym resolution in text:“Which Sheffield is it?”." Proceedings of the the 27th annual international ACM SIGIR conference (SIGIR 2004). 2004. [Yamada 2022] Ikuya Yamada, Koki Washio, Hiroyuki Shindo, Yuji Matsumoto. Global Entity Disambiguation with BERT. Association for Computational Language. 2022
5. Outline ⚫ Investigation of [Yamada+2022] ⚫ Model Input and Output ⚫ Model Details ⚫ Model Accuracy, Error Analysis on disambiguation ⚫Proposed Method ⚫ Evaluation Experiment ⚫ Experimental Setup ⚫ Experimental Results ⚫ Analysis of Results 5
6. Input/Output of [Yamada+2022] 6
7. Disambiguation Model ⚫ LUKE (Language Understanding with Knowledge- based Embeddings) ⚫ Geographical mention → Acquire embedding representation ⚫ Entity Prediction Head ⚫ Predict entity from candidates based on embedding ⚫ Classification problem for candidates 7
8. Language Model LUKE ⚫Handling words and entities(mentions) ⚫ Input text X (token sequence {𝑡1 , 𝑡2 … }) ⚫ 𝑒 ∈ 1, 0 Whether token is a word or mention E𝑚𝑏𝑒𝑑𝑑𝑖𝑛𝑔 = 𝐿𝑈𝐾𝐸(𝑋, 𝑡𝑖 , 𝑒) 8
9. Language Model LUKE(2) ⚫Distinguishes between words and mentions to entities ⚫Uses Japanese pre-trained LUKE* ⚫Model pre-trained on Wikipedia 9 ・Anchor Link → Mention to Entity ・Others → Word * https://huggingface.co/studio-ousia/luke-japanese-base
10. Evaluation of [Yamada2022+] ⚫ Dataset: Japanese Wikification Corpus* ⚫ Corpus based on newspaper articles ⚫ Mention → Mention category, Corresponding Wikipedia article ⚫ Target: 5,525 geographical mentions where the correct answer is included in candidates ⚫ Average number of candidates per mention: 32.9 10 [Yamada+2022] Accuracy(%) 89.8 (4961 / 5525) [Jargalsaikhan+] Jargalsaikhan et at. Building a Corpus for Japanese Wikification with Fine-Grained Entity Classes, Proceedings of the ACL 2016 Student Research
11. Error Analysis ⚫ Confusion with same-named places 【Correct: Hiyoshi-cho(Kyoto) Output: Hiyoshi-cho(Tokorozawa)】 Example: … Center in Hiyoshi-cho, Kyoto Prefecture ⚫ Confusion with entities of different categories 【Correct: Noboribetsu City, Output: Noboribetsu Station】 Example: The restaurant guide … is the first in Noboribetsu 11
12. Outline ⚫ Investigation of [Yamada+2022] ⚫ Model Input and Output ⚫ Model Details ⚫ Model Accuracy, Error Analysis on disambiguation ⚫Proposed Method ⚫ Evaluation Experiment ⚫ Experimental Setup ⚫ Experimental Results ⚫ Analysis of Results 12
13. Proposed Method: Idea ⚫The error on Hiyoshi-cho in previous slide ⚫The address-like relationship between mentions is not captured? ⚫Relationships between mentions ⚫Propose Hierarchical Insertion ⚫Edit input text ⚫Explicitly reflect address hierarchy relationships 13
14. Hierarchical Insertion (HI) ⚫Each geographical mention has a category ⚫Country, City, Station, etc. ⚫Categories with address-like hierarchy ⚫Country > Province > County > City・・・ ⚫Ignore hierarchy other than the 4 categories 14
15. Execution of Hierarchical Insertion ⚫ Condition: A mention is followed by a lower-category mention ⚫ Example: “Ibaraki-ken no Tsukuba-shi” (Tsukuba City in Ibaraki Prefecture) ⚫ Ibaraki-ken → Province, Tsukuba-shi → City ⚫ Operation: Insert the upper-level mention before the lower-level one ⚫ Result: “Ibaraki-ken no Ibaraki-ken Tsukuba-shi” ⚫ This edited text is used as the model input ⚫ The generated entity candidates remain unchanged 15
16. Why Hierarchical Insertion Works ⚫ Target Geographical Mention→ Treated as “Mention” ⚫Entity identity is [Unknown] ⚫Inserted Place name → Treated as “Word” ⚫Enriched context 16
17. Evaluation Experiment ⚫With HI vs Without HI ⚫Both adopt [Yamada+2022] model ⚫Difference is whether to edit input text with HI ⚫Train and infer with edited data ⚫Category information for HI ⚫Category Estimation Model (Details in Appendix) ⚫Infer category from embedding of geographical mention 17
18. Evaluation Results 18 ⚫Improvement in accuracy observed ⚫Mentions where insertion occurred: 345 ⚫Significant difference at 1% significance level Without HI With HI Accuracy (Correct mentions/total mentions) 89.8 (4,961 / 5,525) 94.0 (5,195 / 5,525)
19. Evaluation Results(2) ⚫Improvement seen in mentions other than place names ⚫Improved mentions: 416 mentions ⚫Editing input texts ⚫Influenced the model itself 19
20. Improvement Example(1) ⚫ Hierarchical information provided directly 【Correct: Miyata Town(Fukuoka), Baseline: Mityata Town(Aichi)】 ⚫ Original: “Fukuoka-ken Miyata-cho no…” ⚫ (Miyata Town in Fukuoka Prefecture) ⚫ After Insertion ⚫ “Fukuoka-ken Fukuoka-ken Miyata-cho no…” ⚫ Appears redundant? ⚫ The inserted “Fukuoka-ken” is treated as a word ⚫ →Reinforces geographical context 20 Province City
21. Improvement Example(2) ⚫ Improvement on “Different Category” errors 【Correct: Mediterranean Sea, Baseline: Mediterranean diet】 ・Original: “Chichuu-kai no Shichiria-tou…” (Grew up on Sicily Island in the Mediterranean sea…) ⚫ Improvement without insertion ⚫ Input text remains the same ⚫ The model learned from other cases where insertion occurred 21 Sea Island
22. Error Case(Limitation) ⚫ Both models failed to disambiguate 【Correct: South Island(New Zealand) Output: Yuzhny Island*】 ・“Nyuujiirando no Minami-jima ni aru…” (…Located in New Zealand’s South Island) ⚫ Address-line hierarchy exists (Country→Island) ⚫We restricted the target categories for this experiment ⚫Future Work→Expanding applicable categories 22 Island Country *The southern island of the Novaya Zemlya archipelago, Russia
23. Summary ⚫Entity Disambiguation for Geographical Mentions ⚫Hierarchical Insertion functions significantly ⚫Reduces confusion with same-named places ⚫Reduces confusion with entities of different categories ⚫Future Work ⚫Expansion of Hierarchical Insertion rules ⚫Mentions with insertion were about 6% in total 23