Elamur Research⌕
☰
home / papers / 2007 / Exploiting Wikipedia as External Knowledge for Named Entity Recognition

Exploiting Wikipedia as External Knowledge for Named Entity Recognition

Paper · 2007 · Jun'ichi Kazama and Kentaro Torisawa

Summary

Kazama and Torisawa address a long-standing NER bottleneck: gazetteers help a lot, but building them by hand is slow, and inducing them from raw text requires complex pattern learning. They note that most Wikipedia articles describe named entities and usually open with a defining sentence. For each capitalized word sequence in the input text, their method looks up a matching Wikipedia article, follows redirects, cleans the markup, and takes the head noun of the first noun phrase after a copula as a fine-grained category label (for example, "politician" or "musician"). These labels, together with information about whether a match was found, become features in a CRF-based NE tagger. On the English CoNLL 2003 data, the Wikipedia features add 1.58 F-points over a baseline. They add 1.21 F-points over a model that uses only the official CoNLL gazetteers. The full system reaches 88.02 F. The work shows that a nearly unsupervised, easily refreshed encyclopedia can stand in for hand-built entity dictionaries.

Key points

  • Resource: English Wikipedia dump from February 2007 (4,030,604 pages), indexed with the HyperEstraier search engine. Only namespace 0 (regular articles) is searched.
  • Lookup candidates are word sequences of up to eight words whose first word contains a capital letter, with a small stop list excluding words like "It" and "He". Candidates are turned into entity names by joining words with underscores.
  • Label extraction runs in three steps:
    • Markup and templates are stripped, and indented hatnote lines starting with ":" are removed.
    • TagChunk performs POS tagging and chunking.
    • The last word of the first NP after is/was/are/were becomes the label. The method moves to the second NP if the first ends in "one", "kind", "sort", "type", or "name" followed by "of".
  • Results on CoNLL 2003 English:
    • +1.58 F over the baseline.
    • +1.21 F over the model using only the CoNLL gazetteers.
    • Final model with all features: 88.02 F, which is 3.03 points above the baseline without any gazetteer-type feature.
  • No entity disambiguation is done, and disambiguation pages are ignored. More than 1,200 distinct category labels are used, yet they still help a 4-class task (PER/LOC/ORG/MISC).

Interesting ideas

  • I1. Treat the encyclopedia's opening sentence as a definition. A simple copula-plus-head-noun rule turns that sentence into a semantic type. This is a cheap way to build typed lexicons for any language with a reasonably sized Wikipedia, Ukrainian included.
  • I2. Keep fine-grained labels raw and let the model learn the mapping. Instead of hand-mapping categories like "guitarist" to PER via WordNet, the CRF learns conjunction features such as Wikipedia tag × NE tag. This avoids brittle heuristic mappings when external label sets do not match the target schema.
  • I3. Deliberately skip disambiguation. The method relies on redirects pointing to an entity's dominant meaning and gets no label from disambiguation pages. Abstaining like this limits noise, which is a useful baseline policy before investing in a full entity-linking component.
  • I4. Small linguistic patches have outsized value. Skipping "kind of"/"type of" style phrases to reach the real head noun is a tiny rule that fixes a systematic extraction error. It is worth replicating in any definition-mining pipeline.
  • I5. Re-tag the benchmark with your own tagger. The authors replaced the CoNLL-provided POS/chunk tags with TagChunk output and saw better accuracy, which suggests checking the quality of supplied preprocessing layers rather than accepting them as given.

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, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

🧭 Topics 1Research subjects and languages the work is about.Named entity recognition · 200 works

Cite

Cite as Jun'ichi Kazama and Kentaro Torisawa. 2007. Exploiting Wikipedia as External Knowledge for Named Entity Recognition. Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL).

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{kazama2007exploiting,
  title = {{Exploiting Wikipedia as External Knowledge for Named Entity Recognition}},
  author = {Jun'ichi Kazama and Kentaro Torisawa},
  booktitle = {Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL)},
  year = {2007},
  url = {https://aclanthology.org/D07-1073/}
}

References

16 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. O. Etzioni, M. Cafarella, D. Downey, A. M. Popescu, T. Shaked, S. Soderland, D. S. Weld, and A. Yates. 2005. Unsupervised named-entity extraction from the web – an experimental study. Artificial Intelligence Journal. find
  2. J. Kazama and K. Torisawa. 2007. A new perceptron algorithm for sequence labeling with non-local features. In EMNLP-CoNLL 2007. find
  3. J. Lafferty, A. McCallum, and F. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In ICML 2001, pages 282–289. find
  4. D. Nadeau, Peter D. Turney, and Stan Matwin. 2006. Unsupervised named-entity recognition: Generating gazetteers and resolving ambiguity. In 19th Canadian Conference on Artificial Intelligence. find
  5. S. P. Ponzetto and M. Strube. 2006. Exploiting semantic role lebeling, WordNet and Wikipedia for coreference resolution. In NAACL 2006. find
  6. L. R. Rabiner. 1989. A tutorial on hidden Markov models and selected applications in speech recognition. Proceedings of the IEEE, 77(2):257–286. find
  7. E. Riloff and R. Jones. 1999. Learning dictionaries for information extraction by multi-level bootstrapping. In 16th National Conference on Artificial Intelligence (AAAI-99). find
  8. M. Ruiz-Casado, E. Alfonseca, and P. Castells. 2006. From Wikipedia to semantic relationships: a semi-automated annotation approach. In Third European Semantic Web Conference (ESWC 2006). find
  9. S. Sekine, K. Sudo, and C. Nobata. 2002. Extended named entity hierarchy. In LREC ’02. find
  10. K. Shinzato, S. Sekine, N. Yoshinaga, and K. Torisawa. 2006. Constructing dictionaries for named entity recognition on specific domains from the Web. In Web Content Mining with Human Language Technologies Workshop on the 5th International Semantic Web. find
  11. M. Strube and S. P. Ponzetto. 2006. WikiRelate! computing semantic relatedness using Wikipedia. In AAAI 2006. find
  12. P. P. Talukdar, T. Brants, M. Liberman, and F. Pereira. 2006. A context pattern induction method for named entity extraction. In CoNLL 2006. find
  13. M. Thelen and E. Riloff. 2002. A bootstrapping method for learning semantic lexicons using extraction pattern context. In EMNLP 2002. find
  14. E. F. Tjong, K. Sang, and F. De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In CoNLL 2003.
  15. A. Toral and R. Muñoz. 2006. A proposal to automatically build and maintain gazetteers for named entity recognition by using Wikipedia. In EACL 2006. find
  16. T. Zesch, I. Gurevych, and M. Möhlhäuser. 2007. Analyzing and accessing Wikipedia as a lexical semantic resource. In Biannual Conference of the Society for Computational Linguistics and Language Technology. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/kazama-2007-exploiting-wikipedia-as-external-knowledge/