Elamur Research⌕
☰
home / papers / 2009 / Glen, Glenda or Glendale: Unsupervised and Semi-supervised Learning of English Noun Gender

Glen, Glenda or Glendale: Unsupervised and Semi-supervised Learning of English Noun Gender

Paper · 2009 · Shane Bergsma, Dekang Lin and Randy Goebel

Summary

English third-person pronouns signal gender and number, so a pronoun resolver can discard candidate nouns that do not match. Earlier gender models counted how often each noun type co-occurred with pronouns of each class. Rare and unseen nouns therefore got no reliable estimate. Ambiguous types such as lawyer or Ford were also given a single label regardless of context.

The authors do not use these counts as the final answer. Instead they filter the Bergsma and Lin (2006) gender database down to a set of very reliable "pseudo-seed" nouns. They find occurrences of these nouns in about a million documents and use the seed genders to auto-label training data for a 4-way discriminative classifier (masculine, feminine, neutral, plural). All mentions of a noun within one document are classified together, using context features and type-level features (morphological and categorical cues).

A semi-supervised variant combines the classifier's output with the original counts, tuned on a small labeled set. It reaches 95.5% accuracy, roughly halving the error of the prior state of the art. The method needs no manual labels to train the base model and can handle unseen words.

Key points

  • Base resource: the path-based gender database of Bergsma and Lin (2006). It lists masculine, feminine, neutral and plural pronoun counts for 3.1 million lower-cased nouns. A log-count classifier over this database serves as the baseline, PATHGENDER.
  • Pseudo-seed filtering: nouns are kept only if they have at least 50 counts, their most-frequent pronoun gender (MFPG) covers at least 85% of those counts, and their database-to-web count ratio clears a threshold. About 45 thousand seed nouns remain.
  • Training examples:
    • Noun mentions are grouped by lower-case string within each document, with ±5 tokens of context per mention.
    • At most 200 groups are kept per noun string, so frequent nouns do not dominate.
    • Only features seen more than 5 times are retained.
  • Main result: FULL-SEMI achieves 95.5% accuracy on 2596 test nouns, a 50% error reduction over prior work.
  • Post-processing: a surname-matching step raises FULL-SEMI to 96.7%, PATHGENDER to 93.3% and PATHGENDER+ to 94.3%. Remaining errors include collective nouns such as United States, group and public, which were labeled plural but predicted neutral.

Interesting ideas

  • I1. An imperfect, sparse statistical resource is used as a labeler rather than as the final model: only its most confident entries become training labels for a richer classifier that generalizes to rare and unseen items.
  • I2. Seed reliability is checked by comparing how often a word appears in the coreference database with how often it appears on the web. Words that are frequent overall but rarely act as antecedents (like importance) get flagged as untrustworthy, which hints at how "referential" a noun is.
  • I3. All mentions of the same noun string in a document are pooled into one composite vector, relying on the one-sense-per-discourse tendency. This yields one consistent decision per entity and more context evidence.
  • I4. Gender is treated as a property of the referent in context, not of the word type. The same noun (lawyer, Ford) can therefore get different classes in different documents.
  • I5. A cheap name-matching post-processor lets a later short mention (a surname) inherit the gender of an earlier full name. It helped every system tested.

For Mova

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

  • I1 (learn): Mova's silver-data pipeline could keep only entries where witness parsers or count statistics agree strongly, then train a feature-rich model that also covers words the statistics never saw.
  • I2 (frames): the coreference component could use a similar ratio as a prior for whether a noun is a plausible pronoun antecedent at all, pruning candidates before agreement checks.
  • I3 (world): when building the entity list of a text, gender, number and type evidence from every mention could be pooled into one entity record, so a single clear cue (e.g. he says) settles the entity throughout the story.
  • I4 (world): entity attributes like gender or animacy should live on the text's entities with their evidence chain, with lexicon defaults used only as overridable priors, which keeps decisions explainable.
  • I5 (frames): coreference could add an explicit rule linking surname mentions to earlier full names and copying agreement features, a transparent step that is easy to explain in the answer trace.

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

🏢GoogleMountain View · United States492 works · 440,462 citations🎓University of AlbertaEdmonton · Canada51 works · 5,017 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 1Research subjects and languages the work is about.English · 2,379 works

Cite

Cite as Shane Bergsma, Dekang Lin and Randy Goebel. 2009. Glen, Glenda or Glendale: Unsupervised and Semi-supervised Learning of English Noun Gender. Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL-2009).

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

@inproceedings{bergsma2009glen,
  title = {{Glen, Glenda or Glendale: Unsupervised and Semi-supervised Learning of English Noun Gender}},
  author = {Shane Bergsma, Dekang Lin and Randy Goebel},
  booktitle = {Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL-2009)},
  year = {2009},
  doi = {10.3115/1596374.1596396},
  url = {https://doi.org/10.3115/1596374.1596396}
}

References

30 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Shane Bergsma and Dekang Lin. 2006. Bootstrapping path-based pronoun resolution. In COLING-ACL, pages 33–40.
  2. Shane Bergsma. 2005. Automatic acquisition of gender information for anaphora resolution. In Canadian Conference on Artificial Intelligence, pages 342–353. find
  3. Avrim Blum and Tom Mitchell. 1998. Combining labeled and unlabeled data with co-training. In COLT, pages 92–100. ↗
  4. Sergey Brin. 1998. Extracting patterns and relations from the world wide web. In WebDB Workshop at 6th International Conference on Extending Database Technology, pages 172–183. find
  5. Razvan Bunescu and Raymond J. Mooney. 2004. Collective information extraction with relational Markov networks. In ACL, pages 438–445. find
  6. Eugene Charniak and Micha Elsner. 2009. EM works for pronoun anaphora resolution. In EACL.
  7. Colin Cherry and Shane Bergsma. 2005. An expectation maximization approach to pronoun resolution. In CoNLL, pages 88–95.
  8. Michael Collins and Yoram Singer. 1999. Unsupervised models for named entity classification. In EMNLP-VLC, pages 100–110.
  9. Koby Crammer and Yoram Singer. 2001. On the algorithmic implementation of multiclass kernel-based vector machines. Journal of Machine Learning Research, 2:265–292. find
  10. Silviu Cucerzan and David Yarowsky. 1999. Language independent named entity recognition combining morphological and contextual evidence. In EMNLP-VLC, pages 90–99. find
  11. Silviu Cucerzan and David Yarowsky. 2003. Minimally supervised induction of grammatical gender. In NAACL, pages 40–47. find
  12. Hal Daumé III and Daniel Marcu. 2006. Domain adaptation for statistical classifiers. Journal of Artificial Intelligence Research, 26:101–126. ↗
  13. Pascal Denis and Jason Baldridge. 2007. Joint determination of anaphoricity and coreference using integer programming. In NAACL-HLT, pages 236–243. find
  14. Oren Etzioni, Michael Cafarella, Doug Downey, Ana-Maria Popescu, Tal Shaked, Stephen Soderland, Daniel S. Weld, and Alexander Yates. 2005. Unsupervised named-entity extraction from the web: an experimental study. Artif. Intell., 165(1):91–134. ↗
  15. Richard Evans and Constantin Orăsan. 2000. Improving anaphora resolution by identifying animate entities in texts. In DAARC, pages 154–162. find
  16. Michael Fleischman, Eduard Hovy, and Abdessamad Echihabi. 2003. Offline strategies for online question answering: answering questions before they are asked. In ACL, pages 1–7. find
  17. Niyu Ge, John Hale, and Eugene Charniak. 1998. A statistical approach to anaphora resolution. In Proceedings of the Sixth Workshop on Very Large Corpora, pages 161–171.
  18. Andrew R. Golding and Dan Roth. 1999. A Winnow-based approach to context-sensitive spelling correction. Machine Learning, 34(1-3):107–130. ↗
  19. Jan Hajič and Barbora Hladká. 1997. Probabilistic and rule-based tagger of an inflective language: a comparison. In ANLP, pages 111–118. find
  20. Sanda Harabagiu, Razvan Bunescu, and Steven Maiorano. 2001. Text and knowledge mining for coreference resolution. In NAACL, pages 55–62. find
  21. Nancy Ide and Keith Suderman. 2004. The American National Corpus first release. In LREC, pages 1681– 84. find
  22. Christopher Kennedy and Branimir Boguraev. 1996. Anaphora for everyone: Pronominal anaphora resolution without a parser. In COLING, pages 113–118.
  23. Shalom Lappin and Herbert J. Leass. 1994. An algorithm for pronominal anaphora resolution. Computational Linguistics, 20(4):535–561. ↗
  24. Dekang Lin. 1998. Dependency-based evaluation of MINIPAR. In LREC Workshop on the Evaluation of Parsing Systems. find
  25. Patrick Pantel and Deepak Ravichandran. 2004. Automatically labeling semantic classes. In HLT-NAACL, pages 321–328.
  26. Rion Snow, Daniel Jurafsky, and Andrew Y. Ng. 2005. Learning syntactic patterns for automatic hypernym discovery. In NIPS, pages 1297–1304. find
  27. Wee Meng Soon, Hwee Tou Ng, and Daniel Chung Yong Lim. 2001. A machine learning approach to coreference resolution of noun phrases. Computational Linguistics, 27(4):521–544. ↗
  28. Ioannis Tsochantaridis, Thomas Hofmann, Thorsten Joachims, and Yasemin Altun. 2004. Support vector machine learning for interdependent and structured output spaces. In ICML. find
  29. Ellen Vorhees. 2002. Overview of the TREC 2002 question answering track. In Proceedings of the Eleventh Text REtrieval Conference (TREC). find
  30. David Yarowsky. 1995. Unsupervised word sense disambiguation rivaling supervised methods. In ACL, pages 189–196.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bergsma-2009-glen-glenda-or-glendale-unsupervised/