Uniform Information Density and Syntactic Reduction: Revisiting that-Mentioning in English Complement Clauses

What it is about

The optional complementizer that ("the boss complained (that) they were crazy") is omitted more often when the complement clause is predictable. The authors re-test this on about 51,000 complement clauses from 50 matrix verbs in the CANDOR corpus of spontaneous Zoom conversation (about 8 million words, parsed with spaCy). The old predictor, the verb's subcategorization probability, is compared with a predictor learned by a small feed-forward network from GPT-2 embeddings of the verb, trained on 236,504 examples with a 33.01% rate of following clauses. The embedding predictor reaches F1 0.691 against 0.660 for subcategorization, and log loss 0.442 against 0.491. Controls include subject form, frequency, priming, locality, speaker commitment and disfluency. The main finding replicates: higher information density gives more overt that. But once a random intercept per verb is added, the subcategorization effect drops out (beta 0.14, p = 0.18), while the embedding effect stays (beta 0.12, p < 0.001). The AIC of the embedding model is 25 points lower.

Ideas

• I1. Per-verb complementizer table in 23. Complementation terms: count, for each matrix verb, how often a clause follows and how often that appears, in a BTreeMap. Use it in en/src/lin.rs to choose between "that" and zero when generating, and measure the choice on the school texts.
• I2. Lemma control as a rule for trust: any rate that looks like a general effect must survive a per-verb split, as in the paper's random-intercept check. Build this check into the measurement script, so a new count-based rule cannot ship unless it holds across verbs.
• I3. Priming as dialogue state: keep the last complement choice of each speaker in the world state (world/src/events.rs or prag dialogue state) and let that be more likely after a that. Measure on the dialogue lessons, with a negative control in which the previous choice is shuffled.
• I4. That-doubling avoidance as a deterministic rule in en/src/lin.rs: do not produce "that that" when the complement begins with that. The paper reports it as a clear effect, so write the rule and count the cases it changes.
• I5. Keep the paper's method, not its network: the feature selection by AIC/BIC over hand-chosen features is a rule that Mova can run with counts. Use it to pick which features (subject form, frequency, verb) enter a count model for that.

For Mova

The main chapter is 23. Complementation. Priming and dialogue: 38. Dialogue, discourse markers and reported speech. Register and speaker commitment: 32. Politeness and hedging and 37. Register and variation. Code: en/src/lin.rs (the choice in generation) and world/src/events.rs (priming state). Measure: accuracy of that-choice on held-out school dialogues split by matrix verb, run twice for determinism, with a shuffled-verb negative control.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. arXiv 2509.05254

@misc{hao2025uniform,
  title = {{Uniform Information Density and Syntactic Reduction: Revisiting that-Mentioning in English Complement Clauses}},
  author = {Hao, Hailin and Kaiser, Elsi},
  year = {2025},
  eprint = {2509.05254},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.05254}
}

References

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

1. Hirotugu Akaike. 1974. A new look at the statistical model identification. In Springer Series in Statistics, pages 215–222. Springer. ↗
2. Suhas Arehalli and Tal Linzen. 2020. Neural language models capture some, but not all, agreement attraction effects. In Proceedings of the 42nd Annual Meeting of the Cognitive Science Society, pages 370–376. Cognitive Science Society. ↗
3. Fatemeh Torabi Asr and Vera Demberg. 2015. Uniform information density at the level of discourse relations: Negation markers and discourse connective omission. In Proceedings of the 11th International Conference on Computational Semantics (IWCS). find
4. Matthew Aylett and Alice Turk. 2004. The smooth signal redundancy hypothesis: A functional explanation for relationships between redundancy, prosodic prominence, and duration in spontaneous speech. Language and Speech, 47(1):31–56. ↗
5. Douglas Bates, Martin Mächler, Ben Bolker, and Steven Walker. 2015. Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1):1– 48. ↗

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/hao-2025-uniform-information-density-syntactic-reduction-revisiting/