Pith. sign in

REVIEW 2 major objections 6 minor 30 references

"Dialogue" vs "Dialog" in NLP and AI research: Statistics from a Confused Discourse

T0 review · 2 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A study of tens of thousands of NLP/AI papers claims the field is stuck on a stable, unexplained mix of 'dialogue' and 'dialog': 72% use 'dialogue', 24% use 'dialog', and 5% use both in the same title or abstract.

desk verdict Careful, honest description of a real orthographic split, probably right about the 72/24/5 numbers, but the hand-built corpus and missing robustness checks keep it conditional. read the letter →

arxiv 2501.00598 v1 pith:4RI5AMKI submitted 2024-12-31 cs.CL

David Gros This is my paper · ORCID
classification cs.CL
keywords dialoguevsdialogorthographicvariationspellingNLP/AIresearchcorpusstatisticsauthor-levelcontextualembeddingsscientificwriting
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

The paper sets out to establish that NLP/AI research is caught in a stable, large-scale spelling disagreement over 'dialogue' vs 'dialog', and that the disagreement is not explained by the usual suspects. Counting more than 87,000 bibliographic records and narrowing to high-impact computing venues, it reports that 72% of the sampled publications use 'dialogue', 24% use 'dialog', and 5% use both in the same title or abstract. Over roughly two decades the mix does not shift consistently, author nationality explains little, and surrounding text barely predicts which spelling appears. If the paper is right, the field is accepting two spellings at once for a word at the center of its subject matter.

What carries the argument

The load-bearing object is the 'Dialog(ue) Publication': any paper with 'dialog(s)' or 'dialogue(s)' in its title or abstract, classed into three mutually exclusive categories—dialogue-only, dialog-only, and both—after filtering to 25 high-impact NLP/AI venues chosen by mean citation count. The counts in these categories produce the 72/24/5 headline and the venue, year, and author breakdowns. Around that object the paper builds a supporting apparatus: noun-phrase extraction to test whether phrases like 'visual dialog' bias spelling, a logistic regression with false-discovery-rate correction to identify the few significant phrases, and a masked-language-model embedding classifier that asks whether surrounding context predicts the spelling. The negative result from that classifier is what supports the paper's reading that the spellings are largely interchangeable in context.

What would settle it

Recompute the dialogue/dialog/both percentages from the full text of every paper at the same venues, rather than from titles and abstracts alone; if the split changes by more than a few points or a clear time trend toward either spelling emerges, the stable-mix conclusion gives way.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the orthographic split in NLP/AI is real, large, and stable: 72% of the sampled publications use 'dialogue', 24% use 'dialog', and 5% use both in the same title and abstract, and two decades of data give no clear trend toward 'dialog'. The paper reports that this split is more common in computing than in other academic disciplines, that author nationality has only a weak association, that prolific authors often publish under both spellings, and that context predicts spelling only in narrow corners: plural forms favor 'dialogue', proper nouns and closed compounds favor 'dialog', and phrases like 'visual dialog' are outliers. The paper concludes that none of the three common explanations—American English, a computing-specific norm, or full interchangeability—completely accounts for the observed facts.

Load-bearing premise

The whole analysis rests on the assumption that the bibliographic database used to find papers correctly labels the venues and fields of NLP/AI research, so that the filtered sample really is the population of papers the conclusions are about.

Editorial extensions

If this is right

  • There is no empirical support in the title/abstract record for the claim that 'dialog' is taking over NLP/AI writing.
  • The 'dialog box' rule does not show up in the analyzed titles and abstracts, so the computing-specific explanation is at best incomplete for prose.
  • Prolific authors and coauthorship networks mix spellings, so individual-level accounts of the choice are unlikely to explain the aggregate pattern.
  • Context-sensitive spelling is rare: phrases such as 'visual dialog' are outliers, while plural forms and proper or compound forms carry small, systematic biases.
  • Source code already favors 'dialog' far more than paper prose, suggesting a code-to-prose carryover rather than the reverse.

Reading between the lines

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

  • A natural next check is full-text analysis: if 'dialog box' appears in the bodies of papers, the computing-specific explanation may be stronger than the title/abstract data suggest.
  • If the author-level mixing reflects indifference, then a controlled survey of authors could test whether the choice is effectively random given identical contexts.
  • The 5% of papers that use both spellings in the same title or abstract is itself a signal that editors and authors do not enforce consistency; a direct extension would be to trace whether the two spellings occupy different sections or refer to different senses within those papers.
  • For downstream systems, the persistent mix implies practical tooling: search, evaluation, and code repositories that treat 'dialog' and 'dialogue' as the same token would become more robust.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper presents a descriptive corpus study of the spelling variation "dialog" vs "dialogue" in NLP/AI research, based on Semantic Scholar metadata for papers whose title or abstract contains the term. The headline finding is that among papers in a hand-selected list of top venues, 72% use "dialogue", 24% use "dialog", and 5% use both in the same title and abstract. The paper also reports no clear time trend toward "dialog" over roughly two decades, a weak association between author nationality and spelling, substantial within-author mixing of the two spellings, and limited influence of linguistic context on spelling choice. The study is explicitly framed as descriptive, and the authors are appropriately cautious in their negative conclusions.

Significance. If the headline statistics are robust, this is a useful, systematic quantitative contribution to a long-standing and practically relevant orthographic question in the NLP/AI community. The paper is honest about its limitations, reports confidence intervals for many comparisons, and uses multiple complementary probes (syntactic parses, RoBERTa embeddings, source-code corpora). It makes no claim to a parameter-free derivation and does not overstate the explanatory power of nationality or context. The main deficit is that the core corpus construction is not accompanied by a robustness analysis or a data/code release, which limits the verifiability of the central claims.

major comments (2)
  1. [Appendix A, Table 2] The definition of "High Impact Dialog(ue) Venues" is a hand-built list of the top 25 venues by mean citation count of Dialog(ue) Papers, and it includes non-NLP/AI venues such as CHI, KDD, CVPR, LAK, and CHIIR while excluding venues with lower average citations. Because the headline 72/24/5 split (§3) and the time-trend analysis (Fig. 4) are computed only over papers in this list, the central descriptive claims are conditional on this particular cutoff and on Semantic Scholar field-label accuracy, which the paper itself notes is imperfect (footnote in §5.1; the "Computational Logic" row in Table 2 is a clear metadata error). The paper should provide a sensitivity analysis over alternative venue sets—for example, a fixed canonical NLP/AI venue list, a different citation threshold, or the full set of CS venues—to show that the split and the null trend are not artifacts of this selection.
  2. [General (reproducibility)] The manuscript does not release the raw data, the exact Semantic Scholar query, the venue list, or the analysis code. Since all claims are built on a corpus that cannot be fully reconstructed from the description alone (e.g., the handling of "Both" across title and abstract, the search pattern details, and the sub-noun-phrase filtering in Appendix E), the reader cannot independently verify the headline percentages or the null shift. I would ask that the authors release a data/code package, or at minimum a table of paper IDs with venue, year, and spelling category, so that the central statistics can be checked.
minor comments (6)
  1. [Table 2] Row 9 lists "International Conference on Computational Logic"; the usual name of the venue is the International Conference on Computational Linguistics (COLING). This should be corrected or explicitly flagged as a Semantic Scholar metadata label.
  2. [Appendix H] The sentence "The accuracy showed no improvement over the baseline of always predicting dialogue (0.725 vs 0.739 baseline)" is confusing: if the baseline is always predicting the majority class, it should equal the dialogue prevalence (about 72%), not 73.9%. Please clarify these numbers, since the current wording suggests the model was actually slightly worse than the majority-class baseline.
  3. [Figure 2 caption] The caption reads "CS Dialogu(ue) Publications" — a typo for "Dialog(ue)".
  4. [Section 5.2] The text says "23.0 percent-points less than authors at British intuitions"; this should be "British institutions".
  5. [Abstract and Section 4] The abstract says "over ~20 years of NLP/AI research," but Figure 4 shows that before 2017 there were fewer than 100 papers per year, so the effective power to detect a shift is concentrated in the recent period. The paper already acknowledges this in §4, but the abstract should be more explicit that the null time-trend conclusion is primarily based on the post-2017 data.
  6. [Section 6.1] The noun-phrase analysis identifies "visual dialog" as the most significant phrase, but this is likely driven by the proper-noun task "Visual Dialog" from a highly cited paper. The paper discusses proper nouns in §6.3, but it would be helpful to explicitly separate named tasks/datasets from generic noun phrases in the logistic regression.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the headline statistics are direct corpus counts and the explanatory probes use external models and data; no fitted parameter is relabeled as a prediction.

full rationale

The paper's derivation chain is descriptive rather than self-referential. RQ1's 72/24/5 split is computed directly from Semantic Scholar title/abstract matches and the Appendix A venue list; the venue list is chosen by mean citation counts, not by the spelling distribution, so the percentage split is not equal by construction to the selection criterion. RQ2's 'no clear evidence of a shift' is a cautious reading of year-by-year counts with bootstrap confidence intervals, not a fitted trend that is then presented as a finding. RQ3 reports a weak nationality correlation with McFadden R2=0.037, which is a description of a fitted model's explanatory power, not a claim that the model's output was predicted from its own fitted constant. RQ4 uses RoBERTa, an external pretrained model, as a probe; the finding that the classifier gives no improvement over the base rate is a negative empirical result, and the base rate itself is just the corpus frequency, not a target that the model was fit to reproduce. The Appendix A venue filter and the abstract-only scope (Appendix D.1) are acknowledged limitations that condition the estimates, but conditioning a statistic on a corpus definition is not circular unless the statistic is logically entailed by that definition, and it is not here. No author self-citations are used as load-bearing evidence, and no uniqueness theorem or imported ansatz forces the conclusions. The honest non-finding is therefore appropriate: no specific reduction can be exhibited, and the paper's claims stand or fall on external data quality rather than on circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper is a descriptive empirical study. Its central numbers depend on hand-chosen corpus filters (venue set, citation percentile, author thresholds and weights) rather than on fitted parameters or new entities. The main unexamined premise is the reliability of the Semantic Scholar metadata and the sensitivity of the chosen thresholds. No invented entities (particles, mediators, forces, dimensions) appear.

free parameters (5)
  • Author nationality weighting (first author 4x, last author 2x)
    Hand-chosen weighting used in §5.2 to aggregate institution countries per paper; affects the reported US/CN/GB dialogue differences and their confidence intervals.
  • Venue inclusion cutoff (top 25 venues by mean citations)
    Defines the 'High Impact Dialog(ue) Venues' population for all headline statistics in §3 and the abstract; changing the cutoff would shift the 72/24/5 split.
  • Cross-discipline citation percentile (75th)
    Used in §3 to build comparable discipline samples; different cutoffs would change the cross-discipline dialogue percentages.
  • Sub-noun-phrase frequency filter (at least 50 unique authors)
    Filter in §6.1 and Appendix E.2 restricts which noun phrases are analyzed; affects which phrases reach significance in the logistic model.
  • Both-counts-as-0.5 rule = 0.5
    Author consistency fractions in §5.1 treat 'both' papers as half dialog; this analytic choice influences the reported mean author dialog fraction of 0.248.
assumptions (4)
  • domain assumption Semantic Scholar's field labels, title/abstract text, and citation counts accurately identify the population of NLP/AI 'Dialog(ue) Papers'.
    Invoked in §2 and Appendix A for all headline statistics; the authors note S2 misidentification splits some author records (§5.1 footnote) but assume the corpus itself is reliable.
  • domain assumption Title and abstract usage of dialog(ue) is a valid proxy for the phenomenon of interest.
    Body text is not parsed (acknowledged in Appendix D.1); the paper restricts conclusions to titles and abstracts, so the proxy is explicit but still an assumption about relevance.
  • domain assumption RoBERTa masked-token embeddings are a sufficient representation of context to reveal spelling influences if they exist.
    In §6.2 and Appendix H, the null classifier result is interpreted as 'limited influence of context'; this holds only if the embedding probe is sensitive to the relevant contextual signals.
  • standard math Standard statistical inference tools (bootstrap CIs, Fisher exact tests, Benjamini-Hochberg correction) are applied correctly.
    Standard practice used throughout; the study contains no derivations, so the main correctness burden is the data rather than the mathematics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of "Dialogue" vs "Dialog" in NLP and AI research: Statistics from a Confused Discourse." pith.science (2026). https://pith.science/paper/4RI5AMKI

@misc{pith2026250100598,
  author       = {Pith},
  title        = {Pith review of: "Dialogue" vs "Dialog" in NLP and AI research: Statistics from a Confused Discourse},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4RI5AMKI}},
  note         = {Machine review of arXiv:2501.00598}
}
read the original abstract

Within computing research, there are two spellings for an increasingly important term - dialogue and dialog. We analyze thousands of research papers to understand this "dialog(ue) debacle". Among publications in top venues that use "dialog(ue)" in the title or abstract, 72% use "dialogue", 24% use "dialog", and 5% use both in the same title and abstract. This split distribution is more common in Computing than any other academic discipline. We investigate trends over ~20 years of NLP/AI research, not finding clear evidence of a shift over time. Author nationality is weakly correlated with spelling choice, but far from explains the mixed use. Many prolific authors publish papers with both spellings. We use several methods (such as syntactic parses and LM embeddings) to study how dialog(ue) context influences spelling, finding limited influence. Combining these results together, we discuss different theories that might explain the dialog(ue) divergence.

Figures

Figures reproduced from arXiv: 2501.00598 by the authors.

Figure 1
Figure 1. Examples of varying uses of dialog(ue) in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Distribution for CS Dialogu(ue) Publications. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Dialog(ue) Papers across disciplines. Computing has the lowest dialogue use, and a much larger use of both. One might think this gives evidence towards Explanation-2, but the similar use in fields like Biology diminishes such conclusions. For the rest of the paper we focus on AI/NLP research [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: CS Dialog(ue) Publications by Year. To re [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]
Figure 6
Figure 6. Figure 6: Author nationality with a 95% CI on dialogue We find some differences from author institu￾tion. Authors at American institutions use dialogue 12.9 percent-points less than authors at Chinese institutions (diff C95 6.1–19.7), and 23.0 percent￾points less than authors at…
Figure 5
Figure 5. Figure 5: Distribution for 100 authors with most CS [PITH_FULL_IMAGE:figures/full_fig_p003_5.png]
Figure 7
Figure 7. Figure 7: Noun phrases with a 95% CI. Darker shaded values statistically significant. See appendix for more. We use the Spacy dependency parser (Honni￾bal et al., 2020) to extract noun phrases3 . A noun phrase is a noun with additional words describ￾ing the noun (for example, "c…
Figure 8
Figure 8. Figure 8: Data from (Google Ngram). This shows data from books indexed by Google. pending on the terms "institution computer science language". We then prompt GPT-4 to use these results to estimate the name’s nationality informed by the web search results. We find GPT-4 demonstr…
Figure 10
Figure 10. Figure 10: Data from estimated home nationality. Iden [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 9
Figure 9. Figure 9: Top 25 sub-noun-phrases. Indentation of counts on the left side of each bar is used to emphasize how more complex sub-noun-phrases can be narrowing versions others. The [DIAL] category is a special cate￾gory for uses of dialog(ue) which is independent and not part of a…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 21 canonical work pages

  1. [1]

    Dialogue

    2024. Dialogue. https://www.merriam-webster.com/dictionary/dialogue

    work page 2024

  2. [2]

    Barr, Premkumar T

    Miltiadis Allamanis, Earl T. Barr, Premkumar T. Devanbu, and Charles Sutton. 2017. https://api.semanticscholar.org/CorpusID:207591052 A survey of machine learning for big code and naturalness . ACM Computing Surveys (CSUR), 51:1 -- 37

    work page 2017

  3. [3]

    Asaf Amrami and Yoav Goldberg. 2019. https://api.semanticscholar.org/CorpusID:168169605 Towards better substitution-based word sense induction . ArXiv, abs/1905.12598

    arXiv 2019

  4. [4]

    Apple Computer

    Inc. Apple Computer. 1983. https://bitsavers.org/pdf/apple/lisa/owners_manual/Lisa_Owners_Guide_1983.pdf Lisa Owner's Guide

    work page 1983

  5. [5]

    John August. 2012. https://johnaugust.com/2012/dialogue-and-dialog Dialogue and dialog

    work page 2012

  6. [6]

    Yoav Benjamini and Yosef Hochberg. 1995. https://api.semanticscholar.org/CorpusID:45174121 Controlling the false discovery rate: a practical and powerful approach to multiple testing . Journal of the royal statistical society series b-methodological, 57:289--300

    work page 1995

  7. [7]

    Guy Deutscher. 2005. The unfolding of language: An evolutionary tour of mankind's greatest invention

    work page 2005

  8. [8]

    Google Books Ngram Viewer

    Google Ngram . Google Books Ngram Viewer . https://books.google.com/ngrams/graph?content=dialog

    work page

Show all 30 references
  1. [9]

    Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. 2020. https://doi.org/10.5281/zenodo.1212303 spacy: Industrial-strength natural language processing in python . Version 3.7.4 and the en\_core\_web\_trf model

    2020 doi

  2. [10]

    Graham, F.Q

    Rodney Michael Kinney, Chloe Anastasiades, Russell Authur, Iz Beltagy, Jonathan Bragg, Alexandra Buraczynski, Isabel Cachola, Stefan Candra, Yoganand Chandrasekhar, Arman Cohan, Miles Crawford, Doug Downey, Jason Dunkelberger, Oren Etzioni, Rob Evans, Sergey Feldman, Joseph Go...

    2023 arXiv

  3. [11]

    Denis Kocetkov, Raymond Li, Loubna Ben Allal, Jia Li, Chenghao Mou, Carlos Muñoz Ferrandis, Yacine Jernite, Margaret Mitchell, Sean Hughes, Thomas Wolf, Dzmitry Bahdanau, Leandro von Werra, and Harm de Vries. 2022. The stack: 3 tb of permissively licensed source code. Preprint

    2022

  4. [12]

    LanguageTool. 2023. https://languagetool.org/insights/post/spelling-dialog-or-dialogue/ Dialogue vs. dialog—spelling in british & american english

    2023

  5. [13]

    Brockman, and Slav Petrov

    Yuri Lin, Jean-Baptiste Michel, Erez Aiden Lieberman, Jon Orwant, William S. Brockman, and Slav Petrov. 2012. https://api.semanticscholar.org/CorpusID:17707301 Syntactic annotations for the google books ngram corpus . In Annual Meeting of the Association for Computational Ling...

    2012

  6. [14]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://api.semanticscholar.org/CorpusID:198953378 Roberta: A robustly optimized bert pretraining approach . ArXiv, abs/1907.11692

    2019 arXiv

  7. [15]

    Gretchen McCulloch. 2019. Because Internet: Understanding the New Rules of Language. Riverhead Books, New York

    2019

  8. [16]

    Microsoft. 1986. https://archive.org/details/win103-sdk-manuals Windows 1.03 SDK User Manuals

    1986

  9. [17]

    University of Chicago Press, editor. 2017. The Chicago Manual of Style, 17th edition. University of Chicago Press, Chicago

    2017

  10. [18]

    Oxford English Dictionary . 2023. https://doi.org/10.1093/OED/3867334626 Dialogue | dialog

    2023

  11. [19]

    PyMuPDF Development Team . 2024. https://pymupdf.readthedocs.io/en/latest/ PyMuPDF Documentation

    2024

  12. [20]

    Skipper Seabold and Josef Perktold. 2010. statsmodels: Econometric and statistical modeling with python. In 9th Python in Science Conference

    2010

  13. [21]

    Mark Sebba. 2007. Spelling and Society: The Culture and Politics of Orthography Around the World. Cambridge University Press, Cambridge

    2007

  14. [22]

    Louisa Shen. 2023. Guiding the overwrite: Resurfacing lexicality in the xerox parc alto. Articulating Media, page 78

    2023

  15. [23]

    Richard L. Venezky. 1999. The American Way of Spelling: The Structure and Origins of American English Orthography. Guilford Press, New York

    1999

  16. [24]

    Gregor Wiedemann, Steffen Remus, Avi Chawla, and Chris Biemann. 2019. https://api.semanticscholar.org/CorpusID:202719403 Does bert make any sense? interpretable word sense disambiguation with contextualized embeddings . ArXiv, abs/1909.10430

    2019 arXiv

  17. [25]

    Writing Explained . 2024. https://writingexplained.org/dialog-vs-dialogue-difference Dialog vs. dialogue – what’s the difference?

    2024

  18. [26]

    Xerox. 1979. http://www.bitsavers.org/pdf/xerox/alto/Alto_Users_Handbook_Sep79.pdf Alto User's Handbook

    1979

  19. [27]

    Xerox. 1981. https://worrydream.com/refs/Xerox_1981_-_8010_Star_Information_System_Learner

    1981

  20. [28]

    https://github.com/shibing624/code-autocomplete code-autocomplete: Code autocomplete with gpt2 model

    Ming Xu. https://github.com/shibing624/code-autocomplete code-autocomplete: Code autocomplete with gpt2 model

  21. [29]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  22. [30]

    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 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.