Elamur Research⌕
☰
home / papers / 2018 / Modeling the Decline in English Passivization

Modeling the Decline in English Passivization

Paper · 2018 · Liwen Hou and David Smith

Summary

The paper asks why the English passive has become rarer over two centuries. Using the Hansard record of British parliamentary debates, the authors show that by-phrase passives fell from about 8% of two-argument clauses in the 1830s to under 6% in the 1990s. They parsed the debates automatically with the Stanford dependency parser, extracted transitive clauses, and trained logistic regression classifiers to predict active or passive voice. The features came from earlier linguistic work: person of agent and patient, pronominality, givenness, constituent length, relative and wh-clauses, recently used passives, and frequent lemmas. They also tracked how the average value and the weight of each feature changed decade by decade, both with separate per-decade models and with a hierarchical Bayesian model in Stan. The main finding is that first- and second-person agents became more common and were also passivized less and less. This shift alone accounts for at least half of the decline. The study shows that diachronic syntax can be studied at scale with automatic parsing instead of hand annotation.

Key points

  • Data: about 26 million two-argument verb phrases from Hansard, roughly 1.5 million of them passive, grouped by decade (1830s–1990s). The WSJ portion of the Penn Treebank was re-parsed the same way so the two corpora could be compared.
  • Passives were detected with Stanford dependency labels (auxpass, nsubjpass, nmod:agent), and active objects with dobj. Short passives without a by-phrase were excluded, so that only clauses where active and passive truly compete were counted.
  • Manual check of 300 parsed clauses: voice plus valency was correct in 95% (actives), 94% (passives) and 92% (one-argument clauses). Full argument identification was correct in 84–89%. Wrong voice was rare; valency mistakes were the main error.
  • Performance:
    • WSJ: accuracy 0.964 against a 0.957 always-active baseline, with passive F1 only 0.395.
    • Hansard, per decade: F1 between about 0.67 and 0.75.
  • Diachronic models:
    • Models trained on balanced per-decade sets of 64,000 examples predicted the 1990s better the closer their training decade was.
    • In the Stan hierarchical model, the local-agent coefficient has a negative slope over time (a = −0.05).

Interesting ideas

  • I1. The authors manually check a few hundred parser outputs on old text, and score both strict argument accuracy and looser voice/valency accuracy, before trusting the parser for a corpus study.
  • I2. Detecting voice from a small set of dependency labels (passive subject, passive auxiliary, agent obliques) is a cheap and dependable signal. In their checks, voice errors were much rarer than argument-attachment errors.
  • I3. Counting only passives that have a by-phrase ensures that active and passive compete for exactly the same meaning, which avoids comparing unlike constructions.
  • I4. The authors add pronoun features alongside person features so that a person effect is not just a hidden givenness effect, and they test each feature class by ablation.
  • I5. A hierarchical model lets each feature weight drift linearly over time while sharing structure across periods. It also shows that training data closer in time to the test data transfers better.

For Mova

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

  • I1 (learn): Before silver data from a new genre (fables, archaic tales) enters training, sample a few hundred outputs and score separate strictness levels so we know which parts are reliable.
  • I2 (frames): Frame building can treat the voice decision as high-confidence and focus explanation and repair effort on choosing agent and patient, especially when a by-phrase is instrumental rather than agentive.
  • I3 (back): When generation chooses between active and passive for a known agent and patient, use local-person agents, givenness and pronominality as explainable features. Prefer the active voice when the agent is "I" or "you".
  • I4 (learn): Evaluations of our learned rules should pair correlated features with explicit controls and use ablations as negative controls before claiming that a linguistic cue matters.
  • I5 (tree): When mixing treebanks from different eras or dialects, weight training data by closeness to the target text, or model a shift in feature weights, instead of pooling everything uniformly.

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 3The field of science this work belongs to (OpenAlex fields).Arts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksLiterature and Literary Theory · 519 works

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

Cite

Cite as Liwen Hou and David Smith. 2018. Modeling the Decline in English Passivization. Proceedings of the Society for Computation in Linguistics (SCiL) 2018.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.7275/r5zc812c

@inproceedings{hou2018modeling,
  title = {{Modeling the Decline in English Passivization}},
  author = {Liwen Hou and David Smith},
  booktitle = {Proceedings of the Society for Computation in Linguistics (SCiL) 2018},
  year = {2018},
  doi = {10.7275/r5zc812c},
  url = {https://doi.org/10.7275/r5zc812c}
}

References

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

  1. Christian Mair and Geoffrey Leech. 2006. Current changes in English syntax. The handbook of English linguistics, 36. find
  2. Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David McClosky. 2014. The Stanford CoreNLP natural language processing toolkit. In Association for Computational Linguistics (ACL) System Demonstrations, pages 55–60. ↗
  3. Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn treebank. Computational Linguistics, 19(2):313–330. ↗
  4. Mitchell Marcus, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, Ann Bies, Mark Ferguson, Karen Katz, and Britta Schasberger. 1994. The Penn treebank: annotating predicate argument structure. In Proceedings of the workshop on Human Language Technology, pages 114–119. Association for Computational Linguistics.
  5. Douglas Roland, Frederic Dick, and Jeffrey L Elman. 2007. Frequency of basic English grammatical structures: A corpus analysis. Journal of memory and language, 57(3):348–379. ↗
  6. Kan Sasaki and Akie Yamazaki. 2006. Two types of detransitive constructions in the Hokkaido dialect of Japanese. Passivization and typology: Form and function, 68. find
  7. Stan Development Team. 2017. Pystan: the python interface to stan, version 2.16.0.0. http://mc-stan.org. ↗
  8. E Judith Weiner and William Labov. 1983. Constraints on the agentless passive. Journal of Linguistics, 19(1):29–58. ↗
  9. Christoph Wolk, Joan Bresnan, Anette Rosenbach, and Benedikt Szmrecsnyi. 2013. Dative and genitive variability in Late Modern English: Exploring cross-constructional variation and change. Diachronica, 30(3):382–419. ↗
  10. Judith Aissen. 1999. Agent focus and inverse in Tzotzil. Language, 75:451–485. ↗
  11. Marc Alexander and Mark Davies. 2015. Hansard corpus 1803-2005. http://www.hansard-corpus.org. Accessed: 2016-03-12. ↗
  12. Lieselotte Anderwald. 2012. Clumsy, awkward or having a peculiar propriety? Prescriptive judgements and language change in the 19th century. Language Sciences, 34(1):28–53. find
  13. Alexandre Bouchard-Côté, David Hall, Thomas L Griffiths, and Dan Klein. 2013. Automated reconstruction of ancient languages using probabilistic models of sound change. Proceedings of the National Academy of Sciences, 110(11):4224–4229. ↗
  14. Joan Bresnan, Shipra Dingare, and Christopher D. Manning. 2001. Soft constraints mirror hard constraints: Voice and person in English and Lummi. In Proceedings of LFG 01, pages 13–32. find
  15. Bob Carpenter, Daniel Lee, Marcus A Brubaker, Allen Riddell, Andrew Gelman, Ben Goodrich, Jiqiang Guo, Matt Hoffman, Michael Betancourt, and Peter Li. 2017. Stan: A probabilistic programming language. Journal of Statistical Software, 76(1). ↗
  16. Ludovic De Cuypere, Kristof Baten, and Gudrun Ra-woens. 2014. A corpus-based analysis of the Swedish passive alternation. Nordic Journal of Linguistics, 37(2):199–223. ↗
  17. Marie-Catherine De Marneffe, Bill MacCartney, Christopher D Manning, et al. 2006. Generating typed dependency parses from phrase structure parses. In Proceedings of LREC, volume 6, pages 449–454. Genoa Italy.
  18. T Mark Ellison and Simon Kirby. 2006. Measuring language divergence by intra-lexical comparison. In Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the Association for Computational Linguistics, pages 273–280. Association for Computational Linguistics. find
  19. Michael AK Halliday. 1967. Notes on transitivity and theme in English: Part 2. Journal of linguistics, 3(2):199–244. find
  20. Geraldine Legendre, Yoshiro Miyata, and Paul Smolensky. 1990. Harmonic grammar – a formal multi-level connectionist theory of linguistic well-formedness: Theoretical foundations. In Proceedings of the Twelfth Annual Conference of the Cognitive Science Society, pages 388–395. find
  21. Erez Lieberman, Jean-Baptiste Michel, Joe Jackson, Tina Tang, and Martin A Nowak. 2007. Quantifying the evolutionary dynamics of language. Nature, 449(7163):713–716. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/hou-2018-modeling-decline-english-passivization/