Elamur Research⌕
☰
home / papers / 2014 / ColLex.en: Automatically Generating and Evaluating a Full-form Lexicon for English

ColLex.en: Automatically Generating and Evaluating a Full-form Lexicon for English

Paper · 2014 · Tim vor der Brück, Alexander Mehler and Zahurul Islam

Summary

English has many lexical resources, but few are free, machine-readable full-form lexica that combine broad inflectional coverage with a large inventory of named entities. The authors build ColLex.EN fully automatically, so it can be regenerated whenever its sources change. They merge several lexica: ClearNLP with its word-formation rules, Wiktionary, Wikipedia, Hunspell, UBY and an inflected UBY+, WordNet as the reference, German National Library name lists, and the Preprocessor2010 lexicon. Known source errors are corrected, such as the proper/common noun split and capitalization. Open-class entries pass through a validation cascade that ends in letter-tetragram HMMs guessing PoS. Unknown Wikipedia tokens are then mapped to existing lemmas through an error-tolerant trie lookup, which captures spelling variants and misspellings. The result holds roughly 11.9 million word forms. On lemma and PoS coverage over SUSANNE and Open ANC, it beats every source lexicon. In SVM text categorization it comes a near second to WordNet. The paper shows that careful automatic merging yields a large yet controlled lexicon.

Key points

  • ColLex.EN contains 11,907,675 word forms, of which 10,306,741 are proper nouns, 1,129,183 nouns, 285,528 adjectives, 150,172 verbs and 15,382 adverbs.
  • Coverage on SUSANNE reaches lemma accuracy 0.976 and PoS accuracy 0.678. On the Open ANC slate subcorpus it reaches 0.972 and 0.871, with UBY+ second on lemma coverage in both corpora.
  • In text categorization (SVM, three samples of 500 OAI abstracts), ColLex.EN scores F = 0.735, just under WordNet's 0.737, while having the best recall (0.770).
  • The PoS guesser uses four separate letter-tetragram HMMs (adjective, adverb, noun, verb) trained on WordNet lemmas.
    • Scores are length-normalized via the geometric mean.
    • A word is rejected if its best score falls below the 2% quantile or if the HMM's PoS disagrees with the source's PoS.
  • Proper vs. common noun decisions follow a two-step rule.
    • If the word is in WordNet, it is a proper noun when it has an instance sense.
    • Otherwise, it is a proper noun when it appears capitalized more often than not in Wikipedia.

Interesting ideas

  • I1. The validation cascade runs from cheap to costly: exact WordNet lookup, then the last part of a hyphenated or spaced compound, then a PoS-indicating suffix, and only then a statistical model, so most decisions stay rule-based and inspectable.
  • I2. Character n-gram models trained per word class, with scores normalized by word length, can both guess a PoS and flag strings that look like no English class at all (below a low quantile).
  • I3. An entry is accepted only when an independent guesser agrees with the source's label, which favours precision over coverage when merging noisy resources.
  • I4. Spelling variants are found with a trie-based Levenshtein search that also allows transpositions but demands an exact match on the first characters, since typos rarely occur word-initially; this both speeds up the search and cuts false matches.
  • I5. The categorization result was read as a warning: a huge proper-noun list (containing entries like "Seventeen" or "All") can inject misleading features, so a bigger lexicon needs stricter selection downstream.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (tokens): The en crate's handling of unknown words could follow the same ordered fallbacks, logging which step fired so every lemma/PoS guess remains explainable.
  • I2 (tokens): A per-class letter model would give the tagger a calibrated "this is junk or foreign" signal for OOV tokens instead of forcing a tag.
  • I3 (learn): This mirrors witness-agreement for silver data, and lexicon entries pulled from Wiktionary/WordNet/Wikidata could require agreement between source tag and morphological model before entering training.
  • I4 (tokens): Mapping misspelled or variant forms to known lemmas this way would let the lemmatizer normalize noisy school-test or fable text before parsing.
  • I5 (world): When populating entities from Wikidata-scale name lists, Mova should require contextual evidence before treating a capitalized common word as a named entity in the text's world.

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

 Institutions

🎓Goethe University FrankfurtFrankfurt am Main · Germany18 works · 373 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.English · 2,263 works

Cite

Cite as Tim vor der Brück, Alexander Mehler and Zahurul Islam. 2014. ColLex.en: Automatically Generating and Evaluating a Full-form Lexicon for English. Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14).

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

@inproceedings{bruck2014collex,
  title = {{ColLex.en: Automatically Generating and Evaluating a Full-form Lexicon for English}},
  author = {Tim vor {der Brück}, Alexander Mehler and Zahurul Islam},
  booktitle = {Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14)},
  year = {2014},
  doi = {10.63317/4ekt9s567q99},
  url = {https://doi.org/10.63317/4ekt9s567q99}
}

References

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

  1. Núria Bel, Muntsa Padró, and Silvia Necsulescu. 2011. A method towards the fully automatic merging of lexical resources. In Proceedings of the ACL Workshop on Language Resources, Technology and Services in the Sharing Paradigm, pages 8–15. find
  2. Brijesh Bhatt, Lahari Poddar, and Pushpak Bhattacharyya. 2013. IndoNet: A multilingual lexical knowledge network for indian languages. In 51st Annual Meeting of the Association for Computational Linguistics. find
  3. Steven Bird. 2006. NLTK: The natural language toolkit. In Proceedings of the COLING/ACL on Interactive presentation sessions, pages 69–72.
  4. Dick Crouch and Tracy Holloway King. 2005. Unifying lexical resources. In Proceedings of the Interdisciplinary Workshop on the Identification and Representation of Verb Features and Verb Classes. find
  5. Andreas Eisele and Tim vor der Brück. 2004. Error-tolerant finite-state lookup for trademark search. In 27th Annual German Conference on AI, KI 2004, number 3238 in LNCS, pages 112–126. find
  6. Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database. Cambridge. MIT Press. find
  7. Iryna Gurevych, Judith Eckle-Kohler, Silvana Hartmann, Michael Matuschek, Christian M. Meyer, and Christian Wirth. 2012a. UBY – a large-scale unified lexical-semantic resource based on LMF. In Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2012). find
  8. Iryna Gurevych, Michael Matuschek, Tri Duc Nghiem, Judith Eckle-Kohler, Silvana Hartmann, and Christian M. Meyer. 2012b. Navigating sense-aligned lexical-semantic resources: The web interface to UBY. In Proceedings of the 11th Konferenz zur Verarbeitung natürlicher Sprache (KONVENS), pages 194–198. find
  9. N. Ide and K. Suderman. 2004. The American National Corpus first release. In Proceedings of the Fourth Language Resources and Evaluation Conference (LREC), pages 1681–1684, Lisbon, Portugal. find
  10. Thorsten Joachims. 2002. Learning to classify text using support vector machines. Kluwer, Boston. ↗
  11. Karin Kipper, Anna Korhonen, Neville Ryant, and Martha Palmer. 2008. A large-scale classification of english verbs. Language Resources and Evaluation, 42(1):21– 40. ↗
  12. Claudia Kunze and Lothat Lemnitzer. 2002. GermaNet – representation, visualization, application. In M. Rodriguez González and C. Paz Suárez Araujo, editors, Proceedings of LREC 2002, pages 1485–1491, Paris. find
  13. Vladimir I. Levenshtein. 1966. Binary codes capable of correcting deletions, insertions, and reversals. Soviet Physics Doklady, 10(8):707–710. find
  14. Mathias Lösch, Ulli Waltinger, Wolfram Horstmann, and Alexander Mehler. 2011. Building a DDC-annotated corpus from OAI metadata. Journal of Digital Information, 12(2). find
  15. Alexander Mehler and Ulli Waltinger. 2009. Enhancing document modeling by means of open topic models: Crossing the frontier of classification schemes in digital libraries by example of the DDC. Library Hi Tech, 27(4):520–539. find
  16. R. Navigli and S. Ponzetto. 2012. BabelNet: The automatic construction, evaluation and application of a wide-coverage multilingual semantic network. Artificial Intelligence, 193:217–250. find
  17. Roberto Navigli. 2013. A quick tour of BabelNet 1.1. In Computational Linguistics and Intelligent Text Processing, pages 25–37. Springer. find
  18. Ian Niles and Adam Pease. 2001. Towards a standard upper ontology. In Proceedings of the international conference on Formal Ontology in Information Systems-Volume 2001, pages 2–9. ACM. find
  19. G.R. Sampson. 1995. English for the Computer: The SU-SANNE Corpus and Analytic Scheme. Oxford University Press. find
  20. Hiroshi Uchida, Meiying Zhu, and Tarcisio Della Senta. 2000. UNL: A gift for a millennium. Technical report, The United Nations University. find
  21. Ulli Waltinger, Alexander Mehler, Mathias Lösch, and Wolfram Horstmann. 2011. Hierarchical classification of OAI metadata using the DDC taxonomy. In Advanced Language Technologies for Digital Libraries, number 6699 in LNCS, pages 29–40. Springer, Heidelberg. ↗
  22. Ulli Waltinger. 2010. On Social Semantics in Information Retrieval: From Knowledge Discovery to Collective Web Intelligence in the Social Semantic Web. Ph.D. thesis, University of Bielefeld. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bruck-2014-collex-en-automatically-generating-evaluating/