Pith. sign in

REVIEW 3 major objections 4 minor 41 references

The lexical and grammatical sources of neg-raising inferences

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Neg-raising inferences come from both the verb and the clause it embeds, a new model of 925 English verbs suggests.

desk verdict A valuable new dataset and a model that is honestly reported, but the structural-source conclusion is not supported by the paper's own model comparison. read the letter →

arxiv 1908.05253 v3 pith:TW44QWVJ submitted 2019-08-14 cs.CL

Hannah Youngeun An , Aaron Steven White This is my paper · ORCID
classification cs.CL
keywords neg-raisingclause-embeddingverbssemanticselectionbooleanmatrixfactorizationlexicalsemanticssubordinateclausestructureprobabilisticmodelentailmentjudgments
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

This paper investigates neg-raising, the phenomenon where negation on a verb like 'think' is interpreted inside its embedded clause: 'Jo doesn't think Bo left' implies 'Jo thinks Bo didn't leave'. The authors collect a large dataset of neg-raising judgments for 925 English clause-embedding verbs across six syntactic frames, two tenses, and two subject persons, and fit a statistical model that jointly learns which properties of verbs and which properties of subordinate clauses license the inference. The best-fitting model has exactly one lexical property and one structural property, and it attributes some neg-raising inferences to the predicate itself and others to the structure of the clause it embeds. This is a new empirical constraint on theories of neg-raising, which have traditionally tied the inference to individual predicates and not to complement structure.

What carries the argument

The key object is a probabilistic boolean matrix factorization model that extends a prior s-selection model from the literature. It factorizes a tensor of neg-raising judgments $N$ into a verb–lexical-property matrix $\Psi$, a lexical-property–structure–subject–tense tensor $\Phi$, a verb–structure-type matrix $\Lambda$, a structure-type–frame matrix $\Pi$, and a structure-type–subject–tense tensor $\Omega$, via the approximation $n_{vfjk} \approx \bigvee_{t,i} \lambda_{vt} \wedge \psi_{vi} \wedge \varphi_{ijk} \wedge \pi_{tf} \wedge \omega_{tjk}$. The number of lexical properties $|I|$ and structural properties $|T|$ are hyperparameters; cross-validation selects $|I|=|T|=1$ as the best trade-off between fit and generalization.

What would settle it

A permutation test that shuffles verbs across frames while preserving each verb's estimated lexical property should leave the frame-level structural parameters $\pi_{tf}$ unchanged if those parameters reflect real structure; if the $\pi_{tf}$ values collapse toward equality under such shuffling, the structural effect is an artifact of which verbs happen to appear in each frame.

Watch

Extended reading notes

Core claim

The central discovery is that neg-raising is not purely a product of lexical knowledge. In a five-fold cross-validation, the model that generalizes best posits one lexical property and one structural property; models with only lexical properties or only structural properties perform comparably, but the joint model is the best fit. The fitted parameters show that the lexical property and the structural property each contribute near-deterministically to neg-raising across subjects and tenses, while variability across verbs is captured by the probability that a verb has the lexical property and selects the structural property. Differences across syntactic frames are captured by the probability that the structural property maps onto each frame. The authors interpret this as evidence that properties of the subordinate clause—such as finiteness, presence of an overt subject, or eventivity—influence whether neg-raising is triggered, a role that prior proposals do not assign to complement structure.

Load-bearing premise

The paper's central claim depends on interpreting the model's single 'structural property' as a real property of subordinate clause structure, such as finiteness or eventivity, rather than as a free latent intercept that merely absorbs frame-level differences in the data.

Editorial extensions

If this is right

  • Theories of neg-raising must explain why subordinate clause structure modulates the inference, not just which predicates trigger it.
  • The same verb can differ in neg-raising behavior across complement types; the model predicts this is systematic, not noise.
  • Subject and tense variability in neg-raising is captured as predicate-specific idiosyncrasy, supporting pragmatic rather than purely syntactic accounts of that variability.
  • The dataset and model provide a reusable tool for measuring neg-raising at scale and for jointly modeling related lexically triggered inferences such as veridicality.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the structural property is a real syntactic feature, neg-raising diagnostics such as negative-polarity-item licensing and Horn-clause behavior should differ between finite and non-finite complements of the same verb; this is a testable consequence the paper does not draw.
  • Because the single lexical and structural properties are almost perfectly correlated, the model may be reducible to a one-dimensional 'neg-raising propensity' per verb modulated by frame-specific weights; a simpler model should be tested before accepting two separate sources.
  • The authors' treatment of subject and tense variation as noise suggests that first-person present-tense effects such as 'I don't know' are pragmatic speaker-commitment effects rather than lexical semantic facts, which could be tested with controlled decontextualized prompts.
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.

Referee Report

3 major / 4 minor

Summary. The paper collects a large-scale dataset of neg-raising judgments for 925 English clause-embedding verbs across six syntactic frames, two matrix tenses, and two matrix subjects. It extends White and Rawlins' boolean matrix factorization model of s-selection with latent lexical properties and latent structural properties, and fits variants with different numbers of lexical and structural properties using held-out weighted KL loss. On the basis of the fitted model with one lexical and one structural property, the paper claims that some neg-raising inferences are attributable to properties of particular predicates while others are attributable to subordinate clause structure (abstract, §5, §7).

Significance. If the central claim were supported, the paper would provide a novel large-scale empirical constraint on theories of neg-raising, and the dataset of 7,936 verb-tense-frame-subject items would be a useful resource. The validation experiments in Appendix C and the care taken with stimulus construction are strengths. However, the two load-bearing premises — that the model comparison favors a joint lexical/structural model, and that the fitted latent 'structural property' corresponds to an identifiable property of subordinate clause structure — are not established by the paper's own analyses. Since the abstract's main finding depends on both premises, the contribution as stated is not supported.

major comments (3)
  1. [§5, Results] The model comparison reported in §5 does not support the claim that subordinate clause structure contributes to neg-raising. The paper lists the three best held-out models (Figure 2, starred) 'in order': |I|=1, |T|=0 (pure lexical), |I|=0, |T|=1 (pure structural), and |I|=1, |T|=1 (joint), and states that none is reliably different from the others. The subsequent sentence, 'Among these three, the model with the best fit to the dataset has |I|=1 and |T|=1,' is contradicted by that ordering; and even if the joint model were numerically best, the reported 95% confidence intervals for pairwise differences imply the difference is within noise. The abstract's attribution of some inferences to subordinate clause structure is therefore not licensed by the held-out loss comparison.
  2. [§4, Eq. (21); §6, Table 3] The structural property is a latent factor fitted to the neg-raising judgments themselves: P(πtf) is a free parameter in Eq. (21), optimized against the same responses it is later used to explain. Table 3 therefore reports fitted frame-level values of this latent factor, not evidence that any independently identifiable syntactic property (finiteness, overt subject, eventivity) drives neg-raising. With |I|=|T|=1, the model has only one structural factor, and the six nonzero P(πtf) values in Table 3 are, in effect, a per-frame latent intercept. Nothing outside the model confirms the mapping from this factor to 'subordinate clause structure' asserted in the abstract, so the central claim rests on an interpretation that the modeling framework does not provide.
  3. [§6, Table 2 and Footnotes 9–10] The analysis of the selected model does not provide independent evidence for two distinct sources of neg-raising. Table 2 reports P(φijk) and P(ωtjk) values that are all near 1 and nearly identical; Footnote 9 explains that the optimizer sets these to the square root of the largest expected neg-raising value. Footnote 10 further states that P(ψvi) and P(λvt) are effectively a single parameter p. As a result, the decomposition into 'lexical' and 'structural' sources is not identified: the same predictions could be obtained by a single per-verb propensity combined with a per-frame effect, which is exactly what the pure-lexical boundary model provides. The paper does not supply an identifiability or model-comparison test that would distinguish these alternatives beyond the reported, non-significant held-out loss differences.
minor comments (4)
  1. [§4, boundary models] The description of the |T|=0 boundary model, which fixes Λ=1|V| and Π=1|F|, is formally unclear if the tensor dimensions are |V|×0 and 0×|F|; the boundary model should be defined directly, e.g., by omitting the λ and π factors from Eq. (21).
  2. [§5, Method] The fold-construction constraint that every verb-frame combination appears at least once in the training folds makes the held-out loss closer to interpolation over frames than to prediction of unseen verb-frame combinations; this should be stated when interpreting the model comparison.
  3. [Figure 2] The y-axis label reads 'Held-out weighted KL divergence (nats)' while the text and caption describe the sum of the weighted KL loss across all five folds; the figure should make explicit whether one fold or the summed value is plotted.
  4. [Reference list] The reference 'Pollack 1976' may be a misspelling of 'Pollock'; please verify the cited author's name.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the structural property is a free latent parameter tested by held-out model comparison, not defined by the target neg-raising inference.

full rationale

The paper's central claim is that neg-raising inferences have both lexical and structural sources. The evidence is the model comparison in Section 5: boundary models with |I|=0 or |T|=0 are fit and evaluated on held-out data, so the inclusion of a structural property is not forced by construction. The latent parameter P(pi_tf) in Table 3 is fitted to the neg-raising judgments, but it is presented descriptively as the fitted relationship, not as an out-of-sample prediction, and the paper does not define the structural property in terms of the neg-raising response; it is a free factor in the factorization in Eq. (21). The use of White and Rawlins' s-selection model and MegaAcceptability is a modeling framework and data resource, not a load-bearing self-citation: the neg-raising measurement itself is externally validated in Appendix C against canonical neg-raising predicates from the literature. The fact that the pure lexical model has the lowest held-out loss and is not reliably different from the joint model is a serious evidential weakness in the paper's conclusion, but that is a correctness and statistical-inference concern, not circularity. No equation reduces to its own input, and no fitted parameter is relabeled as an independent prediction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The model's explanatory power comes from latent factors that are fit to the same neg-raising judgments they are said to explain. The per-frame structural factor is the most consequential: it is the only evidence for the paper's central structural claim, and it is a free parameter rather than a measured property. The dataset and task design are more constrained and are the paper's more reliable contribution.

free parameters (6)
  • P(ψvi), probability verb v has the lexical property = estimated per verb from neg-raising responses
    Latent per-verb factor capturing lexical neg-raising tendency; no independent measurement.
  • P(λvt), probability verb v selects the structural property = estimated per verb; paper reports near-perfect rank correlation with P(ψvi)
    Together with P(ψvi) it forms a single per-verb scalar; not separately identifiable in the |I|=|T|=1 model.
  • P(φijk), probability lexical property licenses neg-raising for subject j, tense k = 0.93, 0.98, 0.95, 0.98 (Table 2)
    Fitted to the neg-raising judgments; near 1 for all subject-tense combinations.
  • P(ωtjk), probability structural property licenses neg-raising for subject j, tense k = same values as P(φijk), Table 2
    Fitted to the same data; the paper notes the values coincide because the model sets both to roughly the square root of the largest expected response.
  • P(πtf), probability structural property maps to frame f = 0.84 to 0.98 (Table 3)
    Per-frame latent parameter; this is the main evidence for structural effects and is entirely data-fitted.
  • Acceptability means αvfjk and response scale terms σ0, β0, participant random effects = not reported in detail
    Used for normalization and loss weighting; adds further fitted degrees of freedom.
assumptions (4)
  • domain assumption Strong pairwise independence of all latent parameters in the factorization (Eqs. 20 and 22)
    The model computes P(nvfjk) as a product over independent latent properties; this is a statistical simplification, not a linguistically motivated constraint.
  • domain assumption The 0-1 slider question measures neg-raising inferences rather than some other judgment
    The task asks how likely the speaker means the embedded negated proposition; validation on 32 verbs supports the measure, but generalization to 925 verbs relies on the same task.
  • domain assumption Acceptability-weighted loss appropriately downweights unreliable items
    The paper weights the neg-raising KL loss by normalized acceptability responses (Eq. 26); this assumes acceptability is a valid proxy for reliability.
  • ad hoc to paper The single structural latent factor corresponds to a property of subordinate clause structure
    Section 6 interprets the fitted P(πtf) as a relationship between structure and neg-raising, but the latent factor has no independent definition outside the model.
invented entities (2)
  • lexical property (model latent ψ)
    purpose: to capture predicate-specific neg-raising tendencies
    No independent measurement; only the fitted probability per verb supports it.
  • structural property (model latent λ/π/ω)
    purpose: to capture complement-structure-specific neg-raising tendencies
    A fitted per-frame latent factor; the paper offers no falsifiable handle outside the fitted model that would distinguish it from a frame-level random effect.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The lexical and grammatical sources of neg-raising inferences." pith.science (2026). https://pith.science/paper/TW44QWVJ

@misc{pith2026190805253,
  author       = {Pith},
  title        = {Pith review of: The lexical and grammatical sources of neg-raising inferences},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TW44QWVJ}},
  note         = {Machine review of arXiv:1908.05253}
}
read the original abstract

We investigate neg(ation)-raising inferences, wherein negation on a predicate can be interpreted as though in that predicate's subordinate clause. To do this, we collect a large-scale dataset of neg-raising judgments for effectively all English clause-embedding verbs and develop a model to jointly induce the semantic types of verbs and their subordinate clauses and the relationship of these types to neg-raising inferences. We find that some neg-raising inferences are attributable to properties of particular predicates, while others are attributable to subordinate clause structure.

Figures

Figures reproduced from arXiv: 1908.05253 by the authors.

Figure 1
Figure 1. Normalized neg-raising scores for different subject, tense, and frame pairs. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Sum of the weighted KL divergence loss across all five folds of the cross-validation for each set￾ting of |I| (# of lexical properties) and |T | (# of struc￾tural properties). |I| = |T | = 0 was not run. properties relevant to neg-raising that it assumes; and (ii) the number |T | of structural properties rel￾evant to neg-raising that it assumes. As with other models based on matrix factorization, higher val￾ues for … view at source ↗
Figure 3
Figure 3. Distribution of P(ψvi) × P(λvt) across predicates, along with selected neg-raising (toward right) and non-neg-raising (toward left) predicates in |I| = |T | = 1 model. (Label height is jittered to avoid overplotting.) described as non-neg-raising (know, notice, real￾ize, love, etc.) fall to left. Thus, in some sense, a predicate’s probability of having the model’s sin￾gle lexical property (plus its probability of se… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 39 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Murray, Benoit Steiner, Paul Tucker, Vijay Vasudevan, Pete Warden, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng

    Martin Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, Manjunath Kudlur, Josh Levenberg, Rajat Monga, Sherry Moore, Derek G. Murray, Benoit Steiner, Paul Tucker, Vijay Vasudevan, Pete Warden, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. 2016. https://www.usenix.org/s...

    work page 2016

  4. [4]

    Dorit Abusch. 2002. Lexical alternatives as a source of pragmatic presuppositions. Semantics and Linguistic Theory, 12:1--19

    work page 2002

  5. [5]

    Renate Bartsch. 1973. `` N egative transportation'' gibt es nicht. Linguistische Berichte, 27(7)

    work page 1973

  6. [6]

    Marc Brysbaert and Boris New. 2009. Moving beyond kucera and francis: A critical evaluation of current word frequency norms and the introduction of a new and improved word frequency measure for american english. Behavior Research Methods, 41:977--990

    work page 2009

  7. [7]

    Gennaro Chierchia. 2013. Logic in grammar: polarity, free choice, and intervention, first edition. Oxford University Press, Oxford

    work page 2013

  8. [8]

    Chris Collins and Paul Martin Postal. 2014. Classical NEG Raising: An Essay on the Syntax of Negation. MIT Press

    work page 2014

Show all 41 references
  1. [9]

    Chris Collins and Paul Martin Postal. 2017. Interclausal neg raising and the scope of negation. Glossa: A Journal of General Linguistics, 2:1--29

    2017

  2. [10]

    Chris Collins and Paul Martin Postal. 2018. Disentangling two distinct notions of neg raising. Semantics and Pragmatics, 11(5)

    2018

  3. [11]

    Paul Crowley. 2019. Neg-raising and neg movement. Natural Language Semantics, 27(1):1--17

    2019

  4. [12]

    Mark Davies. 2017. https://www.english-corpora.org/coca/ Corpus of Contemporary American English (COCA)

    2017

  5. [13]

    Fillmore

    Charles J. Fillmore. 1963. The position of embedding transformations in a grammar. WORD, 19(2):208--231

    1963

  6. [14]

    Gajewski

    Jon R. Gajewski. 2007. Neg-raising and polarity. Linguistics and Philosophy, 30(3):289--328

    2007

  7. [15]

    Gajewski

    Jon R. Gajewski. 2011. Licensing strong npis. Natural Language Semantics, 19(2):109--148

    2011

  8. [16]

    Vincent Homer. 2015. Neg-raising and positive polarity: The view from modals. Semantics and Pragmatics, 8(4):1--88

    2015

  9. [17]

    Laurence Robert Horn. 1971. Negative transportation: unsafe at any speed? Papers from the seventh regional meeting, Chicago Linguistic Society, pages 120--133

    1971

  10. [18]

    Laurence Robert Horn. 1978. Remarks on neg-raising. Syntax and Semantics, 9:129--220

    1978

  11. [19]

    Laurence Robert Horn. 1989. A Natural History of Negation. University of Chicago Press

    1989

  12. [20]

    Laurence Robert Horn. 2014. The cloud of unknowing. In Jack Hoeksema and Dicky Gilbers, editors, Black Book: A Festschrift for Frans Zwarts, pages 178--196. University of Groningen, Groningen, the Netherlands

    2014

  13. [21]

    Laurence Robert Horn and Samuel Bayer. 1984. Short-circuited implicature: A negative contribution. Linguistics and Philosophy, 7(4):397--414

    1984

  14. [22]

    Jackendoff

    Ray S. Jackendoff. 1971. On some questionable arguments about quantifiers and negation. Language, 47(2):282--297

    1971

  15. [23]

    Adam Kilgarriff, Vít Baisa, Jan Bušta, Miloš Jakubíček, Vojtěch Kovář, Jan Michelfeit, Pavel Rychlý, and Vít Suchomel. 2014. http://www.sketchengine.eu The sketch engine: ten years on . Lexicography, pages 7--36

    2014

  16. [24]

    Kingma and Jimmy Lei Ba

    Diederik P. Kingma and Jimmy Lei Ba. 2015. Adam: A method for stochastic optimization. Proceedings of the 3rd International Conference on Learning Representations

    2015

  17. [25]

    Paul Kiparsky. 1970. Semantic rules in grammar. In Hreinn Benediktsson, editor, The Nordic Languages and Modern Linguistics, pages 262--285. Visindafelag Islendinga, Reykjavik

    1970

  18. [26]

    Edward S. Klima. 1964. Negation in English. Englewood Cliffs, NJ: Prentice-Hall

    1964

  19. [27]

    Howard Lasnik. 1972. Analyses of Negation in English. Ph.D. thesis, Massachusetts Institute of Technology, Cambridge, MA

    1972

  20. [28]

    Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, and Edouard Duchesna...

    2011

  21. [29]

    Jay M. Pollack. 1976. A re-analysis of neg-raising in english. Working Papers in Linguistics, 21:189--239

    1976

  22. [30]

    Jacopo Romoli. 2013. A scalar implicature-based approach to neg-raising. Linguistics and Philosophy, 36(4):291--353

    2013

  23. [31]

    Jacopo Romoli and Matthew Mandelkern. 2019. What’s not to like. Linguistic Inquiry, 0(ja):1--21

    2019

  24. [32]

    Nadine Theiler, Floris Roelofsen, and Maria Aloni. 2017. What’s wrong with believing whether? Semantics and Linguistic Theory, 27:248--265

    2017

  25. [33]

    Lucia M. Tovena. 2001. Neg-raising: negation as failure. In Jack Hoeksema, Hotze Rullmann, Víctor Sánchez-Valencia, and Ton van der Wouden, editors, Perspectives on Negation and Polarity Items, pages 331--356. John Benjamins, Amsterdam

    2001

  26. [34]

    Aaron Steven White. 2019. Lexically triggered veridicality inferences. To appear in Handbook of Pragmatics

    2019

  27. [35]

    Aaron Steven White and Kyle Rawlins. 2016. A computational model of s-selection. Semantics and Linguistic Theory, 26:641--663

    2016

  28. [36]

    Aaron Steven White and Kyle Rawlins. 2018. The role of veridicality and factivity in clause selection. Proceedings of the 48th Meeting of the North East Linguistic Society

    2018

  29. [37]

    Aaron Steven White and Kyle Rawlins. 2019. Frequency, acceptability, and selection: A case study of clause-embedding. Under revision for Glossa

    2019

  30. [38]

    Aaron Steven White, Rachel Rudinger, Kyle Rawlins, and Benjamin Van Durme. 2018. Lexicosyntactic inference in neural models. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4717--4724

    2018

  31. [39]

    Yimei Xiang. 2013. Neg-raising: Focus and implicatures. Proceedings of Sinn und Bedeutung, 18:487--503

    2013

  32. [40]

    Hedde Zeijlstra. 2018. Does neg-raising involve neg-raising? Topoi, 37(3):417--433

    2018

  33. [41]

    Frans Zwarts. 1998. Three types of polarity. In Fritz Hamm and Erhard Hinrichs, editors, Plurality and Quantification, pages 177--238. Springer Netherland, Dordrecht

    1998

Pith tools

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