---
title: "Syllable weight encodes mostly the same information for English"
description: "Earlier models of English word segmentation improved by 2–10% when given stress marks. 24 references."
authors: ["John K Pate", "Mark Johnson"]
year: 2014
venue: "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)"
source: "https://aclanthology.org/D14-1091/"
doi: "10.3115/v1/d14-1091"
licence: "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)"
licence_url: "https://creativecommons.org/licenses/by-nc-sa/3.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/pate-2014-syllable-weight-encodes-mostly-same/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/pate-2014-syllable-weight-encodes-mostly-same/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2014](https://md.research.elamur.ai/free/en/papers/2014/page/4/) / Syllable weight encodes mostly the same information for English word segmentation as dictionary stress

# Syllable weight encodes mostly the same information for English word segmentation as dictionary stress

**Authors:** John K Pate and [Mark Johnson](https://md.research.elamur.ai/free/en/authors/mark-johnson/) · **Year:** 2014 · **Venue:** [Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)](https://md.research.elamur.ai/free/en/venues/emnlp/) ([EMNLP 2014](https://md.research.elamur.ai/free/en/venues/emnlp/2014/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY-NC-SA 3.0 (ACL Anthology, before 2016)](https://creativecommons.org/licenses/by-nc-sa/3.0/)

**Links:** [Publisher's page](https://aclanthology.org/D14-1091/) · [PDF](https://aclanthology.org/D14-1091.pdf) · [DOI 10.3115/v1/d14-1091](https://doi.org/10.3115/v1/d14-1091) · [OpenAlex](https://openalex.org/W2250613158)

**Citations:** 2 citations (OpenAlex, 2026-10-14)

## Summary
Earlier models of English word segmentation improved by 2–10% when given stress marks. Those marks were copied from a pronouncing dictionary, which hides how hard it is to detect stress in real speech. Pate and Johnson ask whether syllable weight can stand in for dictionary stress. Weight is a phonological property that is easy to read off the phones: a syllable is heavy if it has a long (tense) vowel or a coda, and heavy syllables tend to attract stress. They extend the Adaptor Grammar segmentation models of Börschinger and Johnson (2014) to use stress, weight, both, or neither. They train on child-directed speech, mainly the Alex corpus, and plot learning curves up to 10,000 utterances. Adaptor Grammars can only attach parameters to whole PCFG rules, so the authors run two experiments to work around this. In the first, weight beats stress once the input is large. The second experiment adds word-final coda modelling, and that earlier advantage goes away. The authors conclude that weight carries largely the same segmentation information as dictionary stress. This suggests a learner could get the benefit of stress without first solving stress perception.

## Key points
- The models are Bayesian Adaptor Grammars, PCFGs that cache whole subtrees rooted at adapted non-terminals.
- Training setup:
  - Training sizes run from 100 to 10,000 utterances.
  - The test set is always the last 1,000 utterances of each corpus.
  - Each model runs as eight chains of 1,000 iterations, followed by minimum Bayes risk decoding.
  - Hyperparameters are set to favour onset maximization.
- Experiment I has parallel weight and stress structures and only an initial/non-initial distinction. Token F at 10,000 utterances:
  - weight-only and weight+stress: 0.861;
  - stress-only: 0.831;
  - baseline: 0.830.
- Experiment II adds word-final coda distinctions. Token F at 10,000 utterances:
  - baseline: 0.846;
  - stress: 0.880;
  - combined: 0.875;
  - weight: 0.865.
- The two grammars without weight improve sharply in Experiment II, while the weight grammars barely change. This suggests the weight advantage in Experiment I came from modelling codas, not from weight itself.
- The learned weight grammar sends 98.1% of initial rhymes to HeavyRhyme, compared with 54.5% of non-initial rhymes. Diphthongs are seldom analysed as ShortVowel.

## Interesting ideas
- Replace a cue that needs costly annotation (dictionary stress) with a cheaply observable correlate (syllable weight), then measure whether the two carry redundant information.
- The paper explains a surprising win as a confound: the weight grammar was implicitly modelling final codas. A second experiment gives the competing model the same capacity, and the win disappears.
- Evaluate every training size on one fixed held-out test set, so the learning curves are truly comparable across data amounts.
- The paper notes that parameters tied to whole rules force symbol splitting. It proposes scoring a rule through several features of its right-hand side instead.
- Inspect the learned grammar itself, such as rhyme-type rewrite rates and vowel heatmaps, to confirm the model found the intended generalization.

## For Mova

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

- **I1** (learn): We could test whether an expensive silver-annotation feature can be swapped for a cheap surface correlate by comparing learning curves with and without each.
- **I2** (learn): When one parser variant beats another, add a control that gives the rival the suspected hidden capacity before crediting the new feature.
- **I3** (learn): Keep a frozen evaluation slice when growing silver training data, so that gains reflect data size and not test drift.
- **I4** (tree): Scoring dependency attachments with decomposed features, rather than one-hot labels for whole configurations, could share statistics across rare UD constructions.
- **I5** (tokens): Report what learned morphology rules actually map, for example which suffix classes they attach to, as an explainable sanity check.

*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

- [Macquarie University](https://md.research.elamur.ai/free/en/universities/macquarie-university/) — University, Sydney, Australia · 44 works · 5,778 citations

📚 Shelves (6): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works · [💭 Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/) · 1,649 works · [💭 Experimental and Cognitive Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/experimental-and-cognitive-psychology/) · 930 works

🧭 Topics (2): [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,379 works · [✂️ Tokenization & segmentation](https://md.research.elamur.ai/free/en/topics/tokenization/) · 124 works

## Cite

Cite as: John K Pate and Mark Johnson. 2014. Syllable weight encodes mostly the same information for English word segmentation as dictionary stress. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP).

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

```bibtex
@inproceedings{pate2014syllable,
  title = {{Syllable weight encodes mostly the same information for English word segmentation as dictionary stress}},
  author = {John K Pate and Mark Johnson},
  booktitle = {Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year = {2014},
  doi = {10.3115/v1/d14-1091},
  url = {https://doi.org/10.3115/v1/d14-1091}
}
```

- [← previous Some Aspects of Verbal Politeness in Maghrebi Arabic Dialects](https://md.research.elamur.ai/free/en/papers/d-anna-2014-some-aspects-verbal-politeness-maghrebi/)
- [next → The Morphological and Semantic Types of Old English Lost Adjectives](https://md.research.elamur.ai/free/en/papers/allo-2014-morphological-semantic-types-old-english/)
- [↑ back to papers of 2014](https://md.research.elamur.ai/free/en/papers/2014/page/4/)

## References

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

1. Benjamin Börschinger and Mark Johnson. 2014. Exploring the role of stress in Bayesian word segmentation using Adaptor Grammars. Transactions of the ACL, 2:93–104. [↗](https://doi.org/10.1162/tacl_a_00168)
2. Michael R Brent and Timothy A Cartwright. 1996. Distributional regularity and phonotactic constraints are useful for segmentation. Cognition, 61:93–125. [↗](https://doi.org/10.1016/s0010-0277(96)00719-6)
3. Michael Brent. 1999. An efficient, probabilistically sound algorithm for segmentation and word discovery. Machine Learning, 34:71–105. [↗](https://doi.org/10.1023/a:1007541817488)
4. Nick Campbell and Mary Beckman. 1997. Stress, prominence, and spectral tilt. In Proceedings of an ESCA workshop, pages 67–70, Athens, Greece. [find](https://scholar.google.com/scholar?q=Nick+Campbell+and+Mary+Beckman%2E+1997%2E+Stress%2C+prominence%2C+and+spectral+tilt%2E+In+Proceedings+of+an+ESCA+workshop%2C+pages+67%E2%80%9370%2C+Athens%2C+Greece%2E)
5. Morten H. Christiansen and Suzanne L Curtin. 1999. The power of statistical learning: No need for algebraic rules. In Proceedings of the 21st annual conference of the Cognitive Science Society. [find](https://scholar.google.com/scholar?q=Morten+H%2E+Christiansen+and+Suzanne+L+Curtin%2E+1999%2E+The+power+of+statistical+learning%3A+No+need+for+algebraic+rules%2E+In+Proceedings+of+the+21st+annual+conference+of+the+Cognitive+Science+Society%2E)
6. Morten H. Christiansen, Joseph Allen, and Mark S. Seidenberg. 1998. Learning to segment speech using multiple cues: A connectionist model. Language and Cognitive Processes, 13:221—268. 2008. The CMU pronouncing dictionary. http://www.speech.cs.cmu.edu/cgi-bin/cmudict. [↗](https://doi.org/10.1080/016909698386528)
7. Anne Cutler and David M Carter. 1987. The predominance of strong initial syllables in the English vocabulary. Computer Speech & Language, 2(3):133– 142. [↗](https://doi.org/10.1016/0885-2308(87)90004-0)
8. Katherine Demuth, Jennifer Culbertson, and Jennifer Alter. 2006. Word-minimality, epenthesis, and coda licensing in the acquisition of English. Language and Speech, 49:137–174. [find](https://scholar.google.com/scholar?q=Katherine+Demuth%2C+Jennifer+Culbertson%2C+and+Jennifer+Alter%2E+2006%2E+Word%2Dminimality%2C+epenthesis%2C+and+coda+licensing+in+the+acquisition+of+English%2E+Language+and+Speech%2C+49%3A137%E2%80%93174%2E)
9. Gabriel Doyle and Roger Levy. 2013. Combining multiple information types in Bayesian word segmentation. In Proceedings of NAACL 2013, pages 117– 126. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Gabriel+Doyle+and+Roger+Levy%2E+2013%2E+Combining+multiple+information+types+in+Bayesian+word+segmentation%2E+In+Proceedings+of+NAACL+2013%2C+pages+117%E2%80%93+126%2E+Association+for+Computational+Linguistics%2E)
10. D B Fry. 1955. Duration and intensity as physical correlates of linguistic stress. J. Acoust. Soc. of Am., 27:765–768. [↗](https://doi.org/10.1121/1.1908022)
11. D B Fry. 1958. Experiments in the perception of stress. Language and Speech, 1:126–152. [↗](https://doi.org/10.1177/002383095800100207)
12. Sharon Goldwater, Thomas L. Griffiths, and Mark Johnson. 2009. A Bayesian framework for word segmentation: Exploring the effects of context. Cognition, 112(1):21–54. [↗](https://doi.org/10.1016/j.cognition.2009.03.008)
13. Mark Johnson and Sharon Goldwater. 2009. Improving nonparametric Bayesian inference: experiments on unsupervised word segmentation with adaptor grammars. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 317–325. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Mark+Johnson+and+Sharon+Goldwater%2E+2009%2E+Improving+nonparametric+Bayesian+inference%3A+experiments+on+unsupervised+word+segmentation+with+adaptor+grammars%2E+In+Proceedings+of+Human+Language+Technologies%3A)
14. Mark Johnson, Thomas L Griffiths, and Sharon Goldwater. 2007. Adaptor grammars: A framework for specifying compositional nonparametric Bayesian models. In B Schoelkopf, J Platt, and T Hoffmann, editors, Advances in Neural Information Processing Systems, volume 19. The MIT Press. [↗](https://doi.org/10.7551/mitpress/7503.003.0085)
15. Peter W Jusczyk, Anne Cutler, and Nancy J Redanz. 1993. Infants’ preference for the predominant stress patterns of English words. Child Development, 64(3):675–687. [↗](https://doi.org/10.1111/j.1467-8624.1993.tb02935.x)
16. Peter W Juszcyk, Derek M Houston, and Mary Newsome. 1999. The beginnings of word segmentation in English-learning infants. Cognitive Psychology, 39(3–4):159–207. [find](https://scholar.google.com/scholar?q=Peter+W+Juszcyk%2C+Derek+M+Houston%2C+and+Mary+Newsome%2E+1999%2E+The+beginnings+of+word+segmentation+in+English%2Dlearning+infants%2E+Cognitive+Psychology%2C+39%283%E2%80%934%29%3A159%E2%80%93207%2E)
17. Mark Liberman and Alan Prince. 1977. On stress and linguistic rhythm. Linguistic Inquiry, 8(2):249–336, Spring. [find](https://scholar.google.com/scholar?q=Mark+Liberman+and+Alan+Prince%2E+1977%2E+On+stress+and+linguistic+rhythm%2E+Linguistic+Inquiry%2C+8%282%29%3A249%E2%80%93336%2C+Spring%2E)
18. Constantine Lignos and Charles Yang. 2010. Recession segmentation: simpler online word segmentation using limited resources. In Proceedings of ACL 2010, pages 88–97. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Constantine+Lignos+and+Charles+Yang%2E+2010%2E+Recession+segmentation%3A+simpler+online+word+segmentation+using+limited+resources%2E+In+Proceedings+of+ACL+2010%2C+pages+88%E2%80%9397%2E+Association+for+Computational+Ling)
19. Constantine Lignos. 2011. Modeling infant word segmentation. In Proceedings of the fifteenth conference on computational natural language learning, pages 29–38. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Constantine+Lignos%2E+2011%2E+Modeling+infant+word+segmentation%2E+In+Proceedings+of+the+fifteenth+conference+on+computational+natural+language+learning%2C+pages+29%E2%80%9338%2E+Association+for+Computational+Linguisti)
20. Constantine Lignos. 2012. Infant word segmentation: An incremental, integrated model. In Proceedings of the West Coast Conference on Formal Linguistics 30. [find](https://scholar.google.com/scholar?q=Constantine+Lignos%2E+2012%2E+Infant+word+segmentation%3A+An+incremental%2C+integrated+model%2E+In+Proceedings+of+the+West+Coast+Conference+on+Formal+Linguistics+30%2E)
21. C Anton Rytting, Chris Brew, and Eric Fosler-Lussier. 2010. Segmenting words from natural speech: subsegmental variation in segmental cues. Journal of Child Language, 37(3):513–543. [↗](https://doi.org/10.1017/s0305000910000085)
22. Erik D Thiessen and Jenny R Saffran. 2003. When cues collide: use of stress and statistical cues to word boundaries by 7-to-9-month-old infants. Developmental Psychology, 39(4):706–716. [↗](https://doi.org/10.1037/0012-1649.39.4.706)
23. Alice Turk, Peter W Jusczyk, and Louann Gerken. 1995. Do English-learning infants use syllable weight to determine stress? Language and Speech, 38(2):143–158. [↗](https://doi.org/10.1177/002383099503800202)
24. Charles Yang. 2004. Universal grammar, statistics or both? Trends in Cognitive Science, 8(10):451–456. [↗](https://doi.org/10.1016/j.tics.2004.08.006)

*Updated 2 November 2026*