Elamur Research
☰
home / papers / 2010 / TAGME: On-the-fly Annotation of Short Text Fragments (by Wikipedia Entities)

TAGME: On-the-fly Annotation of Short Text Fragments (by Wikipedia Entities)

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "author's copy (Università di Pisa) — read only" does not allow redistributing the text itself. Original — pages.di.unipi.it/ferragina/cikm2010.pdf

Summary

Earlier entity-linking systems, such as those of Milne&Witten and Chakrabarti et al., were designed for long documents. They perform poorly on very short, noisy inputs like search-engine snippets, tweets or news items. In such texts, unambiguous "context" anchors are often missing, or inference is too slow for query-time use. Ferragina and Scaiella present Tagme, which links spots in these fragments to Wikipedia articles on the fly.

Candidate spots are drawn from Wikipedia anchor texts, and their candidate senses are the pages those anchors point to. Every other anchor in the text casts a vote for each possible sense. A vote averages the Milne–Witten link-based relatedness between that anchor's senses and the candidate sense, weighted by commonness. A threshold-based disambiguator then picks the most common sense among the best-voted ones. A pruning step removes anchors with both low link probability and weak coherence.

On 30-word fragments, Tagme outperforms adapted competitors, reaching about 78% F-measure with tunable precision and recall. It needs under 2 ms per anchor, which makes semantic annotation practical in real time.

Key points

  • Resources: Built from the Wikipedia snapshot of 6 Nov 2009. It has roughly 3M anchors and about 2.7M sense pages, both indexed with Lucene, plus a link graph of about 147M edges held in memory with Webgraph.
  • Evaluation data: The iitb dataset (about 100 documents, 19K anchors) was split into 30-word fragments. Two Wikipedia-derived sets were also used: Wiki-Disamb30, with 1.4M short texts, for disambiguation, and Wiki-Annot30 for end-to-end annotation.
  • Final configuration:
    • Disambiguation by Threshold (dt), with τ = 2% and ε = 30%.
    • Pruning with ρ_AVG, a simple average of link probability and coherence. It was preferred over the regression-trained ρ_LR because it scored slightly better and needs no training.
    • The ρ_na threshold trades precision (up to ~90%) against recall (up to ~80%).
  • Comparison: On Wiki-Annot30, Tagme beats Milne&Witten across the whole ρ_na range. On long texts, processed with a sliding window of about 10 anchors, it reaches about 78% F, compared with 74.8% reported by Milne&Witten.
  • Speed: Tagme takes under 2 ms per anchor on a commodity PC. Chakrabarti et al.'s system takes more than 2 s for 15 anchors. Demo: tagme.di.unipi.it.

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: Computing & GPUs

Cited by

106 papers of our library cite it (of the 140673 whose bibliographies we read); 1 of them have a page here, 105 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.1145/1871437.1871689

@inproceedings{ferragina2010tagme,
  title = {{TAGME: On-the-fly Annotation of Short Text Fragments (by Wikipedia Entities)}},
  author = {Ferragina, Paolo and Scaiella, Ugo},
  booktitle = {CIKM 2010},
  year = {2010},
  doi = {10.1145/1871437.1871689},
  url = {https://doi.org/10.1145/1871437.1871689}
}

References

8 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. C. Carpineto, S. Osiński, G. Romano, and D. Weiss. A survey of web clustering engines. ACM Comput. Surv., 41(3):1–38, 2009. find
  2. S. Cucerzan. Large-scale named entity disambiguation based on Wikipedia data. Proc. of Empirical Methods in NLP, 2007. find
  3. P. Ferragina and A. Gulli. A personalized search engine based on web-snippet hierarchical clustering. Softw. Pract. & Exper., 38(2): 189-225, 2008. find
  4. P. Ferragina and U. Scaiella. TAGME: On-the-fly annotation of short text fragents (by Wikipedia entities). Available on http://arxiv.org/abs/1006.3498. — [arXiv:1006.3498](https://arxiv.org/abs/1006.3498) ↗
  5. S. Kulkarni, A. Singh, G. Ramakrishnan, and S. Chakrabarti. Collective annotation of Wikipedia entities in web text. In Proc. ACM KDD, 457–466, 2009. find
  6. R. Mihalcea and A. Csomai. Wikify!: linking documents to encyclopedic knowledge. In Proc. ACM CIKM, 233–242, 2007. find
  7. D. Milne and I. H. Witten. An effective, low-cost measure of semantic relatedness obtained from Wikipedia links. In Proc. AAAI Workshop on Wikipedia and Artificial Intelligence, 2008. find
  8. D. Milne and I. H. Witten. Learning to link with Wikipedia. In Proc. ACM CIKM, 509–518, 2008. find