Non-binary bottom-up constituency parsing without arity actions

Eunkyul Leah Jo; Jungyeul Park; Zihao Huang

REVIEW 2 major objections 5 minor 14 references

Non-binary bottom-up constituency parsing without arity actions

T0 review · 2 major / 5 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Arity is not a necessary transition primitive for non-binary bottom-up constituency parsing.

desk verdict Clean factorization showing arity need not be an action label in non-binary bottom-up parsing; controlled reimpl holds up, but the printed D-SHIFT oracle rule is incomplete for common nested-then-sibling cases. read the letter →

arxiv 2607.10591 v1 pith:H2WLPI3V submitted 2026-07-12 cs.CL

Jungyeul Park , Eunkyul Leah Jo , Zihao Huang This is my paper · ORCID

classification cs.CL

keywords constituencyparsingtransition-basednon-binarytreesbottom-uparitystackdelimitersactioninventory

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Bottom-up constituency parsers that build original non-binary trees usually pack the mother label and the number of children into one action, such as Reduce-X#k. This paper argues that the arity number k is not required as an action at all. Constituent labels are introduced separately, while stack delimiters mark the left edge of each reduction domain so that the child span is read off the parser state. In a well-formed configuration the arity is uniquely fixed by the active delimiter and the label marker, so it becomes a derived property rather than a classifier target. On standard English and Chinese treebanks the resulting system stays competitive with an arity-specific baseline under the same implementation, uses a much smaller action inventory, and does not collapse high-arity constituents.

What carries the argument

Delimiter-guided reduction (arity recoverability): dummy delimiters placed by D-SHIFT bound the stack interval that REDUCE or D-REDUCE will compose under a separately introduced label, so k is an invariant of state rather than an action label.

What would settle it

On the same PTB and CTB splits and framework, show that the delimiter-guided parser’s high-arity (k≥4 or k≥5) constituent F-scores fall sharply relative to the arity-specific baseline, or that its predicted arity histogram diverges strongly from gold while overall F1 remains high only by collapsing to low-arity trees.

Watch

Extended reading notes

Core claim

The arity parameter of non-binary bottom-up reduction is not a necessary transition primitive. Once labels are introduced by a separate NT(X) action and reduction domains are bounded by stack delimiters, arity is uniquely recovered from the completed items between the active delimiter and the label marker, while still constructing original non-binary trees directly.

Load-bearing premise

The training oracle’s rule for when to insert and when to keep or consume delimiters must produce sequences a neural policy can learn well enough that recovered spans stay accurate; if boundary decisions systematically fail, arities go wrong even though recoverability holds by construction.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The punchline is simple: you can drop REDUCE-X#k entirely from non-binary bottom-up constituency parsing. Introduce the label with NT(X), mark left edges with a dummy delimiter via D-SHIFT, and let REDUCE/D-REDUCE recover the child span from the stack interval. Arity becomes a derived property of a well-formed configuration rather than a classifier target. That factorization is new relative to Fernández-González & Gómez-Rodríguez 2019 and keeps direct construction of the original trees.

They run the comparison cleanly. Same Stanza encoder, composition, optimizer, and eval script for the binary baseline, the arity-specific reimplementation, and their delimiter system. Inventory shrinks from ~200 to ~80 actions, development and test F1 stay competitive with the arity baseline, the predicted arity histogram tracks gold, and high-arity F-scores do not collapse. They also document the real costs—longer transition sequences and the new error modes around delimiter insertion and persistence—without overselling. That is useful, proportionate work.

The soft spot is the static oracle in §3.3. The printed rule (D-SHIFT on sentence start or grandparent change with non-negative depth delta) correctly handles flat same-depth siblings and entry into nests, but it suppresses the required delimiter when a new sibling phrase begins after nested material: depth decreases, so the condition fails even though the grandparent changes. Complex subjects followed by VP are common in PTB/CTB; without that delimiter the subsequent reduce would pull in the wrong material. Yet the reported F1 and arity tables look solid. The recoverability claim itself is definitional and fine; the concrete oracle description is incomplete or imprecise. Probably the implementation marks left edges correctly and the text is just underspecified, but it needs fixing.

Everything else is minor: numbers are not SOTA, only two treebanks, longer derivations, no public code mentioned. Citations are standard and fair. This is for people who design transition systems and care about action inventories. The core methodological result is real and well-supported under the matched setup. Send it to referees; it deserves the discussion.

Referee Report

2 major / 5 minor

Summary. The paper proposes a non-binary bottom-up transition system for constituency parsing in which reduction arity is not an action parameter. Constituent labels are introduced by NT(X), while reduction spans are recovered from dummy delimiters placed by D-SHIFT and managed by REDUCE vs. D-REDUCE. In a well-formed configuration σ|D|s1|…|sk|X, arity k is uniquely determined by the stack interval between the active delimiter and the label marker. The system is evaluated against a reimplemented arity-specific baseline (REDUCE-X#k) and a binary bottom-up baseline in a shared Stanza framework on PTB and CTB, reporting competitive F1, a substantially smaller action inventory, near-gold arity distributions, and no high-arity collapse.

Significance. If the transition system is correctly specified and the empirical results hold under that specification, the paper makes a clean and useful design contribution: it shows that direct non-binary bottom-up construction need not encode the cross-product of labels and arities in the action inventory. The controlled same-framework comparison, inventory compression (Table 3), arity-profile analysis (Tables 4–5), and explicit boundary-error audit (Table 7) are strengths. The contribution is incremental rather than transformative, but it reframes bottom-up reduction as a state-recoverable domain problem and is of clear interest to transition-based parsing.

major comments (2)

Section 3.3 (D-SHIFT oracle): the stated rule D-SHIFT(wi)=1 only when i=0 or when g(wi)≠g(wi-1) and d(wi)-d(wi-1)≥0 fails for a common gold configuration—post-nested siblings. After a nested child (e.g., a deep PP inside a subject NP), the next sibling phrase (e.g., VP) has a changed grandparent but strictly smaller terminal depth, so the rule suppresses the required left-edge delimiter. Without that delimiter, the subsequent phrase-level REDUCE over the outer active D incorrectly consumes the preceding sibling(s). PTB/CTB contain many such cases; the written static oracle therefore cannot generate correct gold derivations for a large fraction of trees, even though recoverability is definitional for well-formed configurations (Section 3.5). This is load-bearing for the claim that the published system constructs original non-binary trees. Please give a corrected, complete oracle (with a n
Appendix A / relation of oracle to association rule a(D): Section 3.3 associates each delimiter with the highest phrasal ancestor whose left edge coincides with the shifted token, and chooses REDUCE vs. D-REDUCE by whether the completed constituent equals a(D). The incomplete D-SHIFT rule above breaks this association for right siblings after nesting. Even if the depth condition is only a sketch, the paper currently provides no alternative decision procedure that is both (i) local enough for a static oracle and (ii) complete for gold trees. A short completeness argument or oracle pseudocode covering nested and flat branching is needed before the factorization claim can be assessed as realized by the concrete system.

minor comments (5)

Table 2: the upper block mixes published numbers from heterogeneous setups with the controlled lower block; a clearer visual separation or caption note would avoid over-reading cross-paper F1 gaps.
Table 1 vs. Table 2: development binary F1 exceeds both non-binary systems by a large margin on CTB; a brief discussion of why binarization helps under this encoder would help readers interpret the non-binary comparison.
Section 5.5 / Table 7: error counts are absolute; rates per sentence or per gold delimiter would make PTB vs. CTB severity easier to compare.
Figure 1 and Section 3.2: FINISH is listed but the completion flag f is otherwise unused in the transition semantics; a one-line note on why f is retained (framework compatibility) would avoid confusion.
Notation: D is introduced as a dummy delimiter implemented as empty NT(); stating once whether NT() is a distinct action or only an internal stack token would clarify the inventory in Table 3.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: arity recoverability is an intentional design invariant of the new transitions, and empirical claims are tested against held-out treebanks with independent baselines.

full rationale

The paper's central formal claim (arity is uniquely determined by the active delimiter and label marker in a well-formed configuration) is true by construction of the transition definitions it introduces (REDUCE/D-REDUCE compose exactly the completed items between D and X; see §3.2–3.5 and Figure 1). This is not a circular derivation that smuggles the conclusion into the premises and then presents it as an independent prediction; it is the explicit design of a new transition system whose purpose is to make arity a derived state property rather than an action label. The empirical claims (competitive F1, preserved arity profile, no high-arity collapse) are evaluated on standard held-out PTB/CTB splits against reimplemented baselines under a shared framework, with no fitted constants later re-labeled as predictions, no load-bearing self-citation uniqueness theorems, and no ansatz imported via self-citation. Self-citations (JP-EVALB) are only for an evaluation script and do not force the result. The system is therefore self-contained against external benchmarks; any potential oracle incompleteness is a correctness/specification issue outside the circularity criteria.

Assumptions & free parameters 1 free parameters · 3 assumptions · 1 invented entities

The central claim rests on standard transition-system machinery plus one new stack marker (the dummy delimiter) whose insertion is governed by a gold-tree oracle. No free parameters are fitted to produce the recoverability property itself; neural hyper-parameters are shared across systems and therefore cancel in the comparison. The only invented entity is the delimiter, which is purely instrumental and never appears in output trees.

free parameters (1)

Stanza neural hyper-parameters (hidden size 512, dropout 0.2, 100 epochs, etc.)
Shared across all three systems; they affect absolute F1 but cancel in the relative comparison that supports the claim that arity actions are unnecessary.

assumptions (3)

ad hoc to paper A well-formed stack configuration of the form σ|D|s1|…|sk|X uniquely determines reduction arity k as the number of completed items between the active delimiter and the label marker.
Stated as the recoverability invariant in Section 3.5; true by construction once the transition rules are fixed.
ad hoc to paper Static oracle for D-SHIFT inserts a delimiter at sentence start or when the gold grandparent changes without a strictly shallower depth move.
Section 3.3; the training signal for boundary placement depends on this gold-derived rule.
domain assumption Standard shift-reduce configuration triple (stack, buffer index, finish flag) and EVALB/JP-EVALB evaluation conventions.
Inherited from the transition-based parsing literature cited in Section 2.

invented entities (1)

Dummy delimiter D (implemented as empty NT())
purpose: Marks the left edge of an active reduction domain so that arity can be recovered from stack geometry rather than predicted as an action parameter.
Introduced in Section 3; has no tree-theoretic content and is excluded from output trees. Independent evidence is limited to the empirical success of the resulting parser.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Non-binary bottom-up constituency parsing without arity actions." pith.science (2026). https://pith.science/paper/H2WLPI3V

@misc{pith2026260710591,
  author       = {Pith},
  title        = {Pith review of: Non-binary bottom-up constituency parsing without arity actions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H2WLPI3V}},
  note         = {Machine review of arXiv:2607.10591}
}

abstract

Non-binary bottom-up constituency parsing is usually taken to require arity actions: reductions such as \(\textsc{Reduce-}X\#k\) specify both the mother label and the number of children to be composed. We show that this arity parameter is not a necessary transition primitive. Our parser introduces constituent labels separately and recovers reduction spans from delimiter-bounded stack configurations. In a well-formed reduction configuration, arity is uniquely determined by the active delimiter and the label marker, making it a derived property of parser state rather than an action label. This factorization removes label--arity-specific reduce actions while preserving direct construction of original non-binary trees. Experiments on PTB and CTB show that the delimiter-guided parser remains competitive with an arity-specific bottom-up baseline under the same implementation framework, with substantially smaller action inventories. Analyses further show that its predicted arity profile remains close to the gold treebanks and that high-arity constituents do not collapse when arity actions are removed.

Figures

Figures reproduced from arXiv: 2607.10591 by the authors.

**Figure 2.** Constituent-building action schemas across the three bottom-up systems. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 4.** gives the corresponding oracle sequence. Transition 1 D-SHIFT(Ms.) 2 NT(NNP) 3 U-REDUCE 4 SHIFT(Haag) 5 NT(NNP) 6 U-REDUCE 7 NT(NP) 8 D-REDUCE 9 D-SHIFT(plays) 10 NT(VBZ) 11 U-REDUCE 12 D-SHIFT(Elianti) 13 NT(NNP) 14 U-REDUCE 15 NT(NP) 16 REDUCE 17 NT(VP) 18 REDUCE 19 SHIFT(.) 20 NT(.) 21 U-REDUCE 22 NT(S) 23 D-REDUCE 24 NT(TOP) 25 REDUCE 26 FINISH [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

Figures from the paper (5 more)

**Figure 3.** Figure 3: shows the PTB-style tree used in the example. TOP S NP NNP Ms. NNP Haag VP VBZ plays NP NNP Elianti . [PITH_FULL_IMAGE:figures/full_fig_p010_3.png]

**Figure 5.** Figure 5: Transition inventory of the binary bottom-up baseline. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

**Figure 6.** Figure 6: Transition inventory of the arity-specific non-binary bottom-up baseline. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

**Figure 7.** Figure 7: Example PTB-style, punctuation-aware binary [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

**Figure 8.** Figure 8: Oracle transition sequences for the binary and [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 2 canonical work pages

[1]

Ezra Black, Steven P. Abney, Dan Flickinger, Claudia Gdaniec, Ralph Grishman, Phil Harrison, Don Hindle, Robert Ingria, Frederick Jelinek, Judith Klavans, Mark Liberman, Mitch Marcus, Salim Roukos, Beatrice Santorini, and Tomek Strzalkowski. 1991. https://aclanthology.org/H91-1060/ A Procedure for Quantitatively Comparing the Syntactic Coverage of English...

1991
[2]

Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. 2016. https://doi.org/10.18653/v1/N16-1024 Recurrent Neural Network Grammars . In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 199--209, San Diego, California. Association for Computa...

doi:10.18653/v1/n16-1024 2016
[3]

Daniel Fern \' a ndez-Gonz \' a lez and Carlos G \' o mez-Rodr \' i guez. 2019. https://doi.org/10.1016/j.artint.2019.07.006 Faster shift-reduce constituent parsing with a non-binary, bottom-up strategy . Artificial Intelligence, 275:559--574

doi:10.1016/j.artint.2019.07.006 2019
[4]

Eunkyul Leah Jo, Angela Yoonseo Park, and Jungyeul Park. 2024. https://aclanthology.org/2024.cl-3.10 A Novel Alignment-based Approach for PARSEVAL Measures . Computational Linguistics, 50(3):1181--1190

2024
[5]

Jiangming Liu and Yue Zhang. 2017 a . https://doi.org/10.1162/tacl \_ a \_ 00070 In-Order Transition-based Constituent Parsing . Transactions of the Association for Computational Linguistics, 5:413--424

doi:10.1162/tacl 2017
[6]

Jiangming Liu and Yue Zhang. 2017 b . https://transacl.org/ojs/index.php/tacl/article/view/927 Shift-Reduce Constituent Parsing with Neural Lookahead Features . Transactions of the Association for Computational Linguistics, 5:45--58

2017
[7]

Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini

Mitchell P. Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. 1993. https://aclanthology.org/J93-2004 Building a Large Annotated Corpus of English: The Penn Treebank . Computational linguistics, 19(2):313--330

1993
[8]

Jungyeul Park, Junrui Wang, Eunkyul Jo, and Angela Park. 2024. https://aclanthology.org/2024.naacl-demo.7 jp-evalb: Robust Alignment-based PARSEVAL Measures . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: System Demonstrations), pages 70--77, Mex...

2024

Show all 14 references

[9]

Kenji Sagae and Alon Lavie. 2005. http://www.aclweb.org/anthology/W/W05/W05-1513 A Classifier-Based Parser with Linear Run-Time Complexity . In Proceedings of the Ninth International Workshop on Parsing Technology (IWPT2005), pages 125--132, Vancouver, British Columbia. Associ...

2005
[10]

Mengqiu Wang, Kenji Sagae, and Teruko Mitamura. 2006. https://doi.org/10.3115/1220175.1220229 A Fast, Accurate Deterministic Parser for Chinese . In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Com...

2006 doi
[11]

Taro Watanabe and Eiichiro Sumita. 2015. https://doi.org/10.3115/v1/P15-1113 Transition-based Neural Constituent Parsing . In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language ...

2015 doi
[12]

Nianwen Xue, Fei Xia, Fu-dong Chiou, and Marta Palmer. 2005. https://doi.org/10.1017/S135132490400364X The Penn Chinese TreeBank: Phrase Structure Annotation of a Large Corpus . Natural Language Engineering, 11(2):207--238

2005 doi
[13]

Yue Zhang and Stephen Clark. 2009. https://aclanthology.org/W09-3825 Transition-Based Parsing of the Chinese Treebank using a Global Discriminative Model . In Proceedings of the 11th International Conference on Parsing Technologies (IWPT'09), pages 162--171, Paris, France. Ass...

2009
[14]

Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. 2013. https://aclanthology.org/P13-1043 Fast and Accurate Shift-Reduce Constituent Parsing . In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), page...

2013

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.

[1] [1]

Ezra Black, Steven P. Abney, Dan Flickinger, Claudia Gdaniec, Ralph Grishman, Phil Harrison, Don Hindle, Robert Ingria, Frederick Jelinek, Judith Klavans, Mark Liberman, Mitch Marcus, Salim Roukos, Beatrice Santorini, and Tomek Strzalkowski. 1991. https://aclanthology.org/H91-1060/ A Procedure for Quantitatively Comparing the Syntactic Coverage of English...

1991

[2] [2]

Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. 2016. https://doi.org/10.18653/v1/N16-1024 Recurrent Neural Network Grammars . In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 199--209, San Diego, California. Association for Computa...

doi:10.18653/v1/n16-1024 2016

[3] [3]

Daniel Fern \' a ndez-Gonz \' a lez and Carlos G \' o mez-Rodr \' i guez. 2019. https://doi.org/10.1016/j.artint.2019.07.006 Faster shift-reduce constituent parsing with a non-binary, bottom-up strategy . Artificial Intelligence, 275:559--574

doi:10.1016/j.artint.2019.07.006 2019

[4] [4]

Eunkyul Leah Jo, Angela Yoonseo Park, and Jungyeul Park. 2024. https://aclanthology.org/2024.cl-3.10 A Novel Alignment-based Approach for PARSEVAL Measures . Computational Linguistics, 50(3):1181--1190

2024

[5] [5]

Jiangming Liu and Yue Zhang. 2017 a . https://doi.org/10.1162/tacl \_ a \_ 00070 In-Order Transition-based Constituent Parsing . Transactions of the Association for Computational Linguistics, 5:413--424

doi:10.1162/tacl 2017

[6] [6]

Jiangming Liu and Yue Zhang. 2017 b . https://transacl.org/ojs/index.php/tacl/article/view/927 Shift-Reduce Constituent Parsing with Neural Lookahead Features . Transactions of the Association for Computational Linguistics, 5:45--58

2017

[7] [7]

Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini

Mitchell P. Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. 1993. https://aclanthology.org/J93-2004 Building a Large Annotated Corpus of English: The Penn Treebank . Computational linguistics, 19(2):313--330

1993

[8] [8]

Jungyeul Park, Junrui Wang, Eunkyul Jo, and Angela Park. 2024. https://aclanthology.org/2024.naacl-demo.7 jp-evalb: Robust Alignment-based PARSEVAL Measures . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: System Demonstrations), pages 70--77, Mex...

2024