Elamur Research⌕
☰
home / papers / 2023 / Lexicalised and non-lexicalized multi-word expressions in WordNet: a cross-encoder approach

Lexicalised and non-lexicalized multi-word expressions in WordNet: a cross-encoder approach

Summary

Princeton WordNet contains many multi-word entries, but not all of them are genuine lexical units. Some are taxonomy labels, quantifier phrases or ordinary collocations. The authors want to separate lexicalised multi-word lexical units (MWLUs) from ad hoc combinations, and they improve on their earlier rule-based and logistic-regression approach.

They collected about 39k multi-word entries from Princeton WordNet and enWordNet, excluding proper names and chemical and biological taxonomy terms. A sample of 387 items was labelled by checking whether six major English dictionaries list each one as a headword. They then fine-tuned a cross-encoder (setu4993/smaller-LaBSE, via sentence-transformers) on pairs of a lemma and its gloss, with the gloss prefixed by hypernym lemmas.

Under .632 bootstrap validation, the model reached roughly 0.78 F1 on the MWLU class, a clear gain over the earlier feature-based model, though it was not significantly better than the majority-class baseline on that measure.

The work offers a way to clean WordNet's MWE inventory, which NLP research often uses as a gold standard. Because the model is language-agnostic, it could be transferred to other wordnets, such as Polish.

Key points

  • Dataset and labels: 39,406 MWEs were drawn from Princeton WordNet and enWordNet, and 86% of them are nouns. The annotated sample of 387 items contains 243 MWLUs and 144 non-MWLUs, labelled against NODE, M-W, Collins, New World, COBUILD and Longman.
  • Model setup: the model input is the lemma paired with hypernym lemmas plus the definition, truncated to 48 tokens. Training ran for 100 bootstrap iterations of 4 epochs each.
  • Results: for the MWLU class, precision, recall and F1 were each about 0.78; for non-MWLU they were about 0.62–0.63.
  • Comparisons:
    • MWLU F1 rose from 58% with the earlier logistic regression to 78% (p<0.01), while non-MWLU F1 was essentially unchanged.
    • Against the majority baseline (F1 0.77), the MWLU F1 difference was not significant (p=.32).
  • Application and release: applied to the full WordNet set, the model tagged 25.5k entries as lexicalised. The data are released on GitHub (MarekMaziarz/MWE-recognition-in-WN) under CC BY 4.0.

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: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2023.gwc-1.28

@inproceedings{maziarz2023lexicalised,
  title = {{Lexicalised and non-lexicalized multi-word expressions in WordNet: a cross-encoder approach}},
  author = {Maziarz, Marek and Grabowski, Łukasz and Piotrowski, Tadeusz and Rudnicka, Ewa and Piasecki, Maciej},
  booktitle = {Proceedings of the 12th Global Wordnet Conference},
  year = {2023},
  doi = {10.18653/v1/2023.gwc-1.28},
  url = {https://doi.org/10.18653/v1/2023.gwc-1.28}
}

References

22 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Meghdad Farahmand and Ronaldo Teixeira Martins. 2014. A supervised model for extraction of multiword expressions, based on statistical context features. In Proceedings of the 10th workshop on multiword expressions (MWE), pages 10–16. find
  2. Christiane Fellbaum, editor. 1998. WordNet – An Electronic Lexical Database. The MIT Press. find
  3. Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. 2020. Language-agnostic bert sentence embedding. find our card
  4. Polona Gantar, Lut Colman, Carla Parra Escartín, and Héctor Martínez Alonso. 2018. Multiword expressions: Between lexicography and NLP. 32(2):138– 162. _eprint: https://academic.oup.com/ijl/article-pdf/32/2/138/29012810/ecy012.pdf. ↗
  5. Howard Jackson. 2022. The Bloomsbury Handbook of Lexicography. Second ed. Bloomsbury Academic, London. find
  6. Elisabetta Jezek. 2016. The Lexicon: An Introduction (Oxford Textbooks in Linguistics. Oxford University Press, Oxford. find
  7. Wenyu Jiang and Richard Simon. 2007. A comparison of bootstrap methods and an adjusted bootstrap approach for estimating the prediction error in microarray classification. Statistics in medicine, 26(29):5320–5334. find
  8. Leonhard Lipka. 1990. An Outline of English Lexicology; Lexical Structure, Word Semantics, and Word-formation. Max Niemeyer, Tuebingen. find
  9. Laurie Bauer. 2019. Complex lexical units. Compounds and multi-word expressions, chapter Compounds and multi-word expressions in English. de Gruyter. find
  10. Marek Maziarz, Łukasz Grabowski, Tadeusz Piotrowski, Ewa Rudnicka, and Maciej Piasecki. 2023. Lexicalisation of polish and english word combinations: an empirical study. Poznan Studies in Contemporary Linguistics. In print. find
  11. Mathieu Constant, Gülşen Eryiǧit, Johanna Monti, Lonneke van der Plas, Carlos Ramisch, Michael Rosner, and Amalia Todirascu. 2017. Survey: Multiword expression processing: A Survey. Computational Linguistics, 43(4):837–892. find our card
  12. Marek Maziarz, Ewa Rudnicka, and Łukasz Grabowski. 2022. Multi-word lexical units recognition in wordnet. In Proceedings of the 18th Workshop on Multiword Expressions@ LREC2022, pages 49–54. find
  13. Anthony Cowie, editor. 2009. The Oxford History of English Lexicography. Clarendon Press. Clarendon Press, London. find
  14. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. find
  15. Bradley Efron. 1983. Estimating the error rate of a prediction rule: improvement on cross-validation. Journal of the American statistical association, 78(382):316–331. find
  16. Bradley Efron and Robert Tibshirani. 1997. Improvements on cross-validation: the 632+ bootstrap method. Journal of the American Statistical Association, 92(438):548–560. find
  17. Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084. find
  18. Martin Riedl and Chris Biemann. 2016. Impact of mwe resources on multiword recognition. In Proceedings of the 12th Workshop on Multiword Expressions, pages 107–111. find
  19. Ivan A. Sag, Timothy Baldwin, Francis Bond, Ann Copestake, and Dan Flickinger. 2002. Multiword expressions: A pain in the neck for NLP. In Computational Linguistics and Intelligent Text Processing, pages 1–15. Springer Berlin Heidelberg. find our card
  20. Nathan Schneider, Emily Danchik, Chris Dyer, and Noah A Smith. 2014. Discriminative lexical semantic segmentation with gaps: running the mwe gamut. Transactions of the Association for Computational Linguistics, 2:193–206. find our card
  21. Tinne Van Rompaey, Kristin Davidse, and Peter Petré. 2015. Lexicalization and grammaticalization: The case of the verbo-nominal expressions be on the/one’s way/road. Functions of Language, 22(2):232–263. find
  22. Michał Woźniak. 2017. Jak znaleźć igl˛e w stogu siana? Automatyczna ekstrakcja wielosegmentowych jed-nostek leksykalnych z tekstu polskiego. IJP PAN, Kraków. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/maziarz-2023-lexicalised-non-lexicalized-multi-word/