Better heads do not guarantee better binarized constituency parsing

Dependency-induced learned heads improve head accuracy but not punctuation-aware binarized constituency parsing

arXiv 2026

Zeyao Qi, Yige Chen, Eitan Klinger, Vivaan Wadhwa, Jungyeul Park*

The Chinese University of Hong Kong, Ma Liu Shui, Hong Kong
The University of British Columbia, Vancouver, Canada
Korea Advanced Institute of Science & Technology, Daejeon, South Korea

*Corresponding author. Affiliations are listed as printed in the paper, which does not map them to individual authors.

Paper (arXiv) PDF
Three trees for 'The little boy likes red tomatoes .'. (a) Original tree: S with children NP (The little boy), VP (likes, NP red tomatoes), and the final period. (b) Standard binarization: S splits into NP and @S, and @S groups VP with the period, so the punctuation attaches near the root. (c) Punctuation-aware binarization: S splits into @S and the period, and @S groups NP and VP, keeping the period as a right boundary of the clause.
Figure 1. Standard binarization may construct binary structure after punctuation has been excluded, whereas punctuation-aware binarization preserves the boundary relation before binary branching is imposed. (a) Original tree with punctuation; (b) standard binarization, where the period ends up grouped with the VP under @S; (c) punctuation-aware binarization, where @S first groups NP and VP and the period closes the sentence.

Abstract

We revisit punctuation-aware tree binarization for constituency parsing and ask whether dependency-induced headedness improves binary parser supervision. Although learned heads substantially outperform rule-based heads in intrinsic head prediction, they do not yield consistent parsing gains after debinarization. In particular, punctuation-conditioned evaluation shows that learned headedness underperforms rule-based binarization in macro-average punctuation-sensitive F1, despite a small overall gain on CTB. Similar instability appears under cross-treebank transfer. These results suggest that linguistically grounded headedness is not necessarily parser-optimal when used as a binarization control signal. The paper presents a negative result: better head prediction does not imply better punctuation-sensitive constituency parsing.

Keywords Tree binarization Constituency parsing Head finding / head percolation Dependency-induced heads Punctuation-aware parsing JP-EVALB Penn Treebank (PTB) Chinese Treebank (CTB) Negative result

Method

Binarized trees are parser training targets, not just a reversible encoding: different binary trees can debinarize to the same original tree while exposing different composition orders and intermediate categories to the learner. The paper tests whether more accurate heads give better binary targets under a fixed parser and evaluation:

  1. Punctuation-aware binarization. Punctuation stays in the local child sequence as boundary structure before branching is imposed: right-boundary punctuation (commas, periods) groups with material to its left, (X α β p) ⇒ (X (@X α β) p), and left-boundary punctuation (opening quotes, parentheses) is treated dually. Debinarization removes only the @-prefixed nodes, so the transformation is exactly reversible.
  2. Head-driven binary conversion. @X intermediate nodes are introduced until each node is binary, keeping the designated head child as the projection anchor of X.
  3. Two head sources. Rule-based Collins-style head percolation (as implemented in Stanford CoreNLP) versus a learned, unlexicalized Transformer head selector trained on dependency-induced gold heads (the child dominating the dependency head of the span).
  4. Same parser, same evaluation. The Berkeley Neural Parser is trained on each set of binary targets (PTB 3 and CTB 5.1), and predicted trees are debinarized before scoring with standard evalb and with JP-EVALB, which keeps punctuation-bearing constituents. Three conditions are compared: R∅ (rule-based, no punctuation awareness), Rp (rule-based, punctuation-aware), and Lp (learned heads, punctuation-aware).

Main Results

  • Learned heads are far more accurate. Intrinsic head accuracy against dependency-induced gold heads rises from 67.89 to 98.54 on PTB and from 92.42 to 100.00 on CTB (rule-based vs. learned; Table 1).
  • But parsing does not consistently improve. With JP-EVALB, learned heads give 95.35 vs. 95.43 for rule-based punctuation-aware binarization on PTB, and only 92.56 vs. 92.40 on CTB (Table 2).
  • Punctuation-sensitive evaluation reverses the CTB gain. Macro-average punctuation-conditioned F1 is lowest for learned heads on both treebanks: PTB 92.63 / 92.43 / 92.27 and CTB 94.22 / 93.99 / 93.50 for R∅ / Rp / Lp (Table 3).
  • Cross-treebank transfer is unstable. Heads transferred from a CTB-trained selector reach 79.03 on Sinica; on FTB they give 86.82, below rule-based 88.65 and learned 88.92 (Table 4). Intrinsic transfer accuracy is 69.41 (CTB→Sinica), 54.95 (PTB→FTB), and 64.70 (CTB→FTB).
  • Conclusion. Head accuracy is not binarization utility: linguistically grounded heads can yield binary spines that are less regular or less aligned with punctuation-bearing constituency structure.
Table 2. Parsing accuracy (labeled bracket F1) after deterministic debinarization. R∅: rule-based binarization without punctuation awareness; Rp: rule-based punctuation-aware; Lp: learned-head punctuation-aware. evalb excludes punctuation; JP-EVALB retains punctuation-bearing constituents. (Table 2 in the paper.)
TreebankBinarizationEVALBJP-EVALB
PTBR∅95.5094.92
Rp95.9795.43
Lp95.8895.35
CTBR∅92.1790.29
Rp94.2692.40
Lp94.4692.56

BibTeX

@misc{qi2026betterheadsguaranteebetter,
      title={Better heads do not guarantee better binarized constituency parsing}, 
      author={Zeyao Qi and Yige Chen and Eitan Klinger and Vivaan Wadhwa and Jungyeul Park},
      year={2026},
      eprint={2605.28131},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.28131}, 
}
More projects from KAIST MLP Lab →