Learning Constituent Headedness

Predicting the head child of every constituent from aligned constituency and dependency treebanks instead of hand-written head-percolation rules

arXiv 2026

Zeyao Qi1,†, Yige Chen1,†, KyungTae Lim2, Haihua Pan1,*, Jungyeul Park2,*

1The Chinese University of Hong Kong 2Korea Advanced Institute of Science & Technology (KAIST)

†Equally contributed authors   *Corresponding authors

Paper (arXiv) PDF

Abstract

Headedness is widely used as an organizing device in syntactic analysis, yet constituency treebanks rarely encode it explicitly and most processing pipelines recover it procedurally via percolation rules. We treat this notion of constituent headedness as an explicit representational layer and learn it as a supervised prediction task over aligned constituency and dependency annotations, inducing supervision by defining each constituent head as the dependency span head. On aligned English and Chinese data, the resulting models achieve near-ceiling intrinsic accuracy and substantially outperform Collins-style rule-based percolation. Predicted heads yield comparable parsing accuracy under head-driven binarization, consistent with the induced binary training targets being largely equivalent across head choices, while increasing the fidelity of deterministic constituency-to-dependency conversion and transferring across resources and languages under simple label-mapping interfaces.

Keywords Constituent headedness Head finding Head-percolation rules (Collins) Head-driven binarization Constituency-to-dependency conversion Penn Treebank / Penn Chinese Treebank Cross-lingual transfer Treebank processing

Method

Headedness is treated as an explicit layer over a constituency tree: a function that picks exactly one head child for every nonterminal. Instead of hand-written percolation tables, gold heads are induced from treebanks where constituency and dependency annotations exist for the same sentences:

  1. Dependency span head. For each constituent, find the unique token in its yield whose dependency governor lies outside the constituent (its attachment point). Sentences where this token is not unique are excluded.
  2. Gold head child. The head child is the child that dominates that token. This criterion is independent of phrase labels, word order, and language-specific rules.
  3. Head prediction. A BERT-base encoder with a two-layer MLP reads only the parent label and the ordered child labels (no words) and selects one child, a constrained k-way classification over gold trees.
  4. Downstream tests. The predicted heads drive (i) head-driven binarization for a fixed Stanza constituency parser and (ii) a deterministic constituency-to-dependency conversion; the only thing that changes between conditions is the head source (Collins-style rules in Stanford CoreNLP versus learned heads).

Experiments use aligned subsets of the Penn Treebank (English) and the Penn Chinese Treebank 5.1; transfer is tested on the Sinica Treebank and the French Treebank through deterministic label mappings, without retraining.

Main Results

  • Near-ceiling head accuracy. Learned heads reach 98.54% on English and 100.00% on Chinese, compared with 67.89% and 92.42% for Collins-style percolation rules.
  • Parsing stays stable. Under head-driven binarization, labeled bracket F1 is 96.32 vs. 96.24 (English) and 94.20 vs. 93.84 (Chinese); the differences are not significant (p = 0.636 and 0.497). The two binarized training sets share most brackets (bracketing F1 95.31 and 97.63) even though only 38.75% and 68.71% of trees have identical spines.
  • Much more faithful conversion. Constituency-to-dependency conversion UAS rises from 51.39 to 94.99 on English and from 77.73 to 92.22 on Chinese.
  • Transfer depends on head conventions. Head accuracy is 69.41 for CTB → Sinica, 64.70 for CTB → FTB, and 54.95 for PTB → FTB. PPs explain the gap: PTB PPs are 97.98% head-final, while FTB and CTB PPs are almost always head-initial, so PTB → FTB PP accuracy is 0.09% versus 89.81% for CTB → FTB.
Tables 1–3 (combined). Rule-based (Collins-style percolation) versus learned constituent headedness on aligned English (Penn Treebank) and Chinese (Penn Chinese Treebank) data: head prediction accuracy (Table 1), parsing labeled bracket F1 under head-driven binarization (Table 2), and UAS of deterministic constituency-to-dependency conversion (Table 3).
LanguageHeadsHead accuracy (%)Parsing F1Conversion UAS
EnglishRule-based67.8996.2451.39
Learned98.5496.3294.99
ChineseRule-based92.4293.8477.73
Learned100.0094.2092.22

BibTeX

@misc{qi2026learningconstituentheadedness,
      title={Learning Constituent Headedness},
      author={Zeyao Qi and Yige Chen and KyungTae Lim and Haihua Pan and Jungyeul Park},
      year={2026},
      eprint={2603.14755},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.14755},
}
More projects from KAIST MLP Lab →