A Computational Approach to Analyzing Language Change and Variation in the Constructed Language Toki Pona

Daniel Huang; Hyoun-A Joo

REVIEW 4 major objections 6 minor 14 references

A Computational Approach to Analyzing Language Change and Variation in the Constructed Language Toki Pona

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper argues that Toki Pona, a constructed language of about 120 words, shows the same kinds of sociolinguistic change and variation as natural languages when a real community uses it.

desk verdict A worthwhile first corpus study of Toki Pona with new empirical findings, but the unvalidated parser/heuristic means the trend lines need a gold-standard check before they carry the argument. read the letter →

arxiv 2508.10246 v1 pith:FY7ONFCX submitted 2025-08-14 cs.CL

Daniel Huang , Hyoun-A Joo This is my paper

classification cs.CL

keywords TokiPonaconstructedlanguagechangesociolinguisticvariationcorpuslinguisticspart-of-speechtaggingfluidwordclassestransitivity

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 asks whether a deliberately designed language with a tiny vocabulary, Toki Pona, undergoes the same kinds of change and variation as natural languages when a real community uses it. Using a custom Earley parser and part-of-speech tagger over two corpora, an informal Discord server with millions of sentences and a formal corpus of published works, the authors track how content words shift among syntactic positions over time and across registers. They find that body-part nouns such as luka 'hand' and uta 'mouth' are increasingly used as transitive verbs in casual use, that pu is losing ground as a standalone noun in favor of lipu pu, and that interjections such as wawa 'strong' have surged in conversational text but not in formal writing. The conclusion is that even a constructed language evolves through community use, shaped by processing efficiency, conformity to existing patterns, and identity signaling.

What carries the argument

The analysis rests on an Earley context-free parser that converts Toki Pona sentences into hierarchical phrase structures, making syntactic positions such as subject, predicate, and direct object explicit. A heuristic scoring module then resolves ambiguity, for example whether tawa is a preposition or a content verb, by preferring interpretations that match observed community usage. Finally, a part-of-speech tagger labels each content word as noun, modifier, intransitive verb, or transitive verb, and the counts are aggregated by year and corpus.

What would settle it

Hand-annotate a random sample of sentences from each year of both corpora using the paper's tag set, then compare the parser's transitive-verb rates for luka and uta with the hand labels; if the parser systematically over- or under-tags these body-part verbs, the upward trend in Figure 3 will not reproduce under gold-standard annotation.

Watch

Extended reading notes

Core claim

The central claim is that Toki Pona, despite being designed with fluid word classes and a minimal lexicon, behaves sociolinguistically like a natural language: content words show systematic diachronic shifts in their preferred syntactic positions, and usage differs by register. Concretely, the paper argues that in informal Toki Pona from 2020 to 2024, luka and uta have become noticeably more frequent as transitive verbs, while pu 'interacting with the official Toki Pona book' has declined as a standalone noun and increasingly appears in the phrase lipu pu, mirroring the language's name-construction pattern. The formal corpus does not lag behind on these innovations, which the authors attribu

Load-bearing premise

The trends stand on the unvalidated assumption that the parser-and-tagger pipeline assigns the correct part of speech to each token; because the ambiguity-resolving heuristic deliberately prefers interpretations that 'reflect observed usage patterns,' the frequency counts are partly shaped by the authors' prior expectations, and a systematic tagging bias would make the reported changes artifacts.

Editorial extensions

If this is right

If the reported trends are accurate, Toki Pona speakers are innovating within the existing grammar: body-part nouns luka and uta now routinely serve as transitive verbs in informal chat, a usage that is shorter than the periphrastic alternative.
pu as a standalone noun for the official book is declining, replaced by the pattern-conforming lipu pu, which suggests speakers normalize novel constructions to existing name patterns.
Register differences appear: pona is consistently more frequent as an interjection in informal chat, and wawa's rise as an interjection is confined to the informal corpus.
Innovations in the informal community are not delayed in formal written Toki Pona, pointing to a homogeneous, tightly connected speech community.
The finding generalizes: constructed languages with fluid word classes can serve as empirical evidence that sociolinguistic mechanisms operate independently of a language's historical depth.

Reading between the lines

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

One testable extension the authors do not run: hold out a later time slice after 2025 and check whether the luka/uta transitive-verb trend continues, saturates, or reverses, since a genuine community innovation should show a logistic adoption curve rather than noise.
If the parser's heuristic scoring is nudged to favor different parses, the reported frequencies could shift; a natural check is to compare against a small hand-annotated gold standard or to perturb the heuristic weights and see how much the trends move.
The social mechanisms invoked, processing efficiency, analogy to name constructions, and identity signaling, are standard explanations in natural-language change; if they hold in a language with no native speakers, it suggests these mechanisms arise from general communication pressures rather than from particular language histories.
Because Toki Pona's lexicon has only about 120 words, a complete inventory of word classes over time is feasible, so this pipeline could be extended to map a trajectory for every content word, offering a near-exhaustive view of change in a small system.

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.

Desk Editor's Note

read the letter

Quick take: this is a modest, clearly written corpus study that reports genuinely new things about Toki Pona usage — body-part nouns sliding into transitive-verb slots, pu losing its standalone-noun use, and interjection use differing across the Discord and long-form corpora. The grammar is grounded in Lang (2014), the code and aggregated data are public, and the authors are upfront about the short time window and the written-only data. That is real credit: the work is reproducible in a way much corpus sociolinguistics is not.

The main soft spot is exactly where your stress-test note lands. The POS counts come from a custom Earley parser plus a heuristic ambiguity-resolution step that explicitly prefers parses "reflecting observed usage patterns." No gold-standard evaluation of either component is reported. If the heuristic biases toward or against certain readings in ways that correlate with year or formality, the trends in Figures 3–7 could be partly manufactured. The authors do not hide the heuristic, and the grammar itself is a simple, publicly inspectable CFG, but until the pipeline is validated against a manually annotated sample, the specific magnitudes and even some directions should be treated as suggestive rather than established. I'd also like token-count baselines and some significance testing for the diachronic claims; the post hoc word selection in §4.1 makes the absence of statistics more consequential. The pu-as-proper-name explanation is interesting but speculative.

That said, the central point — that a constructed language with a small, deliberately constrained grammar still shows sociolinguistically patterned change — does not rest entirely on the parser. The examples in the paper (e.g., luka/uta used as transitive verbs, the lipu pu construction) are real and community-discussed, and the cross-corpus comparison is a sensible design. The paper knows its limits and does not overclaim.

Who it is for: people working on language change, usage-based theory, and conlang communities. It is a solid subfield contribution, not a field reorientation. I'd send it to peer review asking for parser validation and a statistical pass; a gold-standard sample of a few hundred sentences should be feasible given the small lexicon and simple grammar. With that, the empirical claims would be much firmer.

Referee Report

4 major / 6 minor

Summary. This paper applies a corpus-based, computational pipeline to study diachronic and register variation in the constructed language Toki Pona. Using a Discord corpus (6.39M tokens) and the poki Lapo literary corpus, the authors filter with sona-toki, parse with a custom Earley-grammar implementation, resolve ambiguity with a hand-written heuristic scorer, and tag parts of speech. They report that luka and uta increasingly occur as transitive verbs in informal data (Fig. 3), that pu is decreasingly used as a standalone noun (Fig. 4), and that interjection use differs across registers (Figs. 5–6), concluding that Toki Pona changes like natural languages under sociolinguistic pressures.

Significance. If the reported trends are real, the study is a valuable empirical demonstration of language change in a constructed language, with a publicly available pipeline and corpora. Its strengths are the scale of the informal corpus, the explicit grammar, public code, and the comparison of two registers. However, the central quantitative claims rest on a parser and scoring heuristic that have not been validated against gold-standard annotations, and the analysis lacks inferential statistics; as a result, the findings are currently suggestive rather than established.

major comments (4)

[§3.3–§3.4, Figs. 3–7] The load-bearing assumption is that the Earley parser and heuristic scorer produce accurate POS counts. §3.4 says the scorer 'favors specific patterns' and that this 'prioritization ... reflects observed usage patterns'; this means the counts encode the author's prior, and no gold-standard evaluation or inter-annotator agreement is reported. If parse errors correlate with word, year, or formality, every trend in Figs. 3–7 could be an artifact. Report precision/recall per tag on a stratified annotated sample, or remove/redraft the causal and general claims.
[§4.1, Figs. 3–4] Words were chosen because they 'showed a high degree of change' and were 'referenced often' (quote), introducing selection bias. No significance tests, confidence intervals, or raw token counts are given; percentages normalize but do not show magnitude. Because the survey window is 2020–2025 and the sample is one community, the claim of general language change is not supported. Provide counts and a regression model (e.g., year as continuous predictor) and report uncertainty.
[§4.2.2, Fig. 7] The claim that adoption is 'not delayed' in the formal corpus rests on visual inspection of a single line plot. The figure has no error bars or fitted trajectories. Quantify the onset or slope difference between corpora (e.g., interaction of year and corpus in a model), or phrase the result as descriptive.
[§5] The conclusion extends the results to 'proper nouns may be used more as adjectives,' but the analysis only addresses pu as a modifier versus standalone noun; pu is not a proper name. The cognitive/sociolinguistic explanations (processing efficiency, identity) are plausible but not tested by the corpus data. Separate observational trends from explanatory hypotheses.

minor comments (6)

[Figures 3–7] Axis labels do not state units or whether the plotted value is a proportion of tokens, sentences, or messages. Include raw counts or a clear denominator; error bars or confidence bands would also help.
[Figure 1 / Table 1] The parse tree in Figure 1 appears inconsistent with the grammar in Table 1: the tree shows 'li' as part of a content-word constituent rather than as a daughter of S between the subject and predicate. Please correct the figure or clarify the tree notation.
[Introduction / References] Typographical error: 'T oki Pona' appears with an extra space in the Introduction and in the reference for Lang (2014).
[§3.4] The heuristic scoring algorithm is described only informally and via an external repository. Include a pseudocode or formal description in the paper, since this is the step that most affects the reported counts.
[§3.2] The sona-toki filtering configuration is said to be 'the same configuration' used by an external tool, but no specific parameters are listed. Provide the configuration or a pointer to the exact code version so the filtering step is reproducible.
[§5] The phrase 'proper nouns may be used more as adjectives' should be 'proper names may be used more as modifiers,' since Toki Pona names are not a distinct noun class and the paper's own analysis concerns pu, an adjective.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; heuristic tuning is a validation concern, not a circular derivation.

full rationale

The paper's derivation chain is corpus-based measurement: filter with sona-toki, parse with a grammar based on Lang (2014), resolve ambiguity with a heuristic, and count POS tags defined by syntactic contexts. The reported trends (luka/uta as transitive verbs, pu as standalone noun, interjection use) are not directly encoded in the heuristic; the heuristic's stated preference is for prepositional interpretations, which are not among the reported trends. The grammar and tag definitions are externally grounded in Lang (2014). There are no author self-citations used as load-bearing support; cited tools and references (sona-toki, poki Lapo, Lang, Bybee, Labov) are independent. The lack of gold-standard parser validation is a methodological limitation, but it does not make any specific result reduce by construction. No step of the derivation is equivalent to its input by definition, and no fitted parameter is renamed as a prediction. Therefore the paper exhibits no significant circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a custom parsing and tagging pipeline whose components are validated only by appeal to community-observed patterns, not by any reference corpus.

free parameters (2)

heuristic scoring preferences = none (hand-chosen)
§3.4: The parser prefers prepositional interpretations when available and other unlisted patterns; these preferences were chosen to reflect observed usage, directly shaping POS counts.
sona-toki filtering configuration = not specified
§3.2: Uses 'the same configuration' as Danielson (2024b) with thresholds for Toki Pona sentence scoring; exact values not given in the paper.

assumptions (4)

domain assumption Toki Pona grammar is accurately captured by the simplified CFG in Table 1 and Lang (2014).
Used throughout §3.3-3.5 to parse all sentences; any mismatch produces systematic tagging errors.
domain assumption sona-toki heuristics correctly distinguish Toki Pona from non-Toki-Pona sentences.
§3.2 filters 5.97M sentences down to 1.23M with these heuristics; no precision/recall evaluation is reported.
ad hoc to paper The heuristic scoring algorithm's parse preferences reflect true community usage probabilities.
§3.4: 'This prioritization of interpretations reflects observed usage patterns'; this makes the parser's biases an input rather than an output.
domain assumption The ma pona Discord corpus represents informal conversational Toki Pona and poki Lapo represents formal written Toki Pona.
§3.1: The research questions depend on this contrast, but both corpora are written and draw from an overlapping community (§4.2.2 acknowledges this).

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Computational Approach to Analyzing Language Change and Variation in the Constructed Language Toki Pona." pith.science (2026). https://pith.science/paper/FY7ONFCX

@misc{pith2026250810246,
  author       = {Pith},
  title        = {Pith review of: A Computational Approach to Analyzing Language Change and Variation in the Constructed Language Toki Pona},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FY7ONFCX}},
  note         = {Machine review of arXiv:2508.10246}
}

read the original abstract

This study explores language change and variation in Toki Pona, a constructed language with approximately 120 core words. Taking a computational and corpus-based approach, the study examines features including fluid word classes and transitivity in order to examine (1) changes in preferences of content words for different syntactic positions over time and (2) variation in usage across different corpora. The results suggest that sociolinguistic factors influence Toki Pona in the same way as natural languages, and that even constructed linguistic systems naturally evolve as communities use them.

Figures

Figures reproduced from arXiv: 2508.10246 by the authors.

**Figure 2.** Example of distribution of words across different content word classes for [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Change in usage of luka and uta as transitive verbs in ma pona pi toki pona from 2020 to 2024 as percentage of total content word usage. A possible explanation for this change is processing efficiency following Bybee’s (2015, p. 1) argumentation that “mental processes ... are ... main causes of change”: body-part terms can be used as nouns to express similar ideas with the verb pilin ’to feel,’ as in jan li pilin e… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Change in the use of pu as a modifier and noun in ma pona pi toki pona. This is likely due to pressure to conform with existing name constructions in the language, applying known patterns to new contexts and thereby reinforcing these patterns (cf. Bybee, 2015, p. 10). …

**Figure 5.** Figure 5: Usage of pona as a standalone interjection between the two corpora. As seen in [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

**Figure 6.** Figure 6: Usage of wawa as a standalone interjection between the two corpora. This pattern can be attributed to several factors. First, interjections may inherently appear less in non-conversational data because simpler statements like acknowledgements are less necessary. Seco…

**Figure 7.** Figure 7: Usage of luka as a transitive verb between the two corpora. One explanation for the lack of an apparent difference between the two corpora is the homogeneity of the Toki Pona community. A significant majority of authors with published works in Toki Pona are members of …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 13 canonical work pages

[1]

Bybee, J. (2015). Language Change. Cambridge University Press

work page 2015
[2]

(2020, June)

Chandra, K., & Radvan, T. (2020, June). nearley: A parsing toolkit for JavaScript . https://doi.org/10.5281/zenodo.3897993

work page doi:10.5281/zenodo.3897993 2020
[3]

Chomsky, N. (1956). Three models for the description of language.IRE Transactions on Information Theory , 2(3), 113–124. https://doi.org/10.1109/TIT.1956. 1056813 UGA Working Papers in Linguistics, Vol 8, 2025 Language Change and Variation in Toki Pona Page 14

doi:10.1109/tit.1956 1956
[4]

(2024, January)

Comrie, B., Haspelmath, M., & Bickel, B. (2024, January). Leipzig Glossing Rules . https://www.eva.mpg.de/lingua/resources/glossing-rules.php

work page 2024
[5]

(2024a, March)

Danielson, G., III. (2024a, March). When was pu added to toki pona? Retrieved March 1, 2024, from https://web.archive.org/web/20240301100156/https: //mun.la/lipu/pu.html

work page arXiv 2024
[6]

(2025, April).sona-toki [GitHub repository]

Danielson, G., III. (2025, April).sona-toki [GitHub repository]. https://github.com/ gregdan3/sona-toki

work page 2025
[7]

Dyer, W., Futrell, R., Liu, Z., & Scontras, G. (2020). Predicting cross-linguistic adjec- tive order with information gain . arXiv: 2012.15263 [cs.CL]. https://arxiv. org/abs/2012.15263

work page Pith review arXiv 2020
[8]

Earley, J. (1970). An efficient context-free parsing algorithm.Commun. ACM, 13(2), 94–102. https://doi.org/10.1145/362007.362035

work page arXiv 1970

Show all 14 references

[9]

Glass, L. (2024). The red dress is cute: Why subjective adjectives are more often predicative. Corpus Linguistics and Lingustic Theory . https : / / doi . org / 10 . 1515/cllt-2024-0044 kala Asi, ijo vivi, jan Juwan, & jan Kita. (2025, April).poki Lapo [GitHub repository]. htt...

2024
[10]

Labov, W. (2010). Principles of linguistic change . John Wiley & Sons, Ltd. https : //doi.org/10.1002/9781444327496.ch9

2010 doi
[11]

Lang, S. (2014). T oki Pona: The Language of Good. Tawhid. ma pona pi toki pona [Discord server]. (2025, January). https://discord.gg/mapona

2014
[12]

McCulloch, G. (2020). Because Internet: Understanding how Language is Changing . Vintage

2020
[13]

Meulen, S. v. d. (2021, October). Request for New Language Code Element in ISO 639-3. https://iso639-3.sil.org/sites/iso639-3/files/change_requests/2021/ 2021-043_tok.pdf

2021
[14]

Peterson, D. J. (2015). The Art of Language Invention: From Horse-Lords to Dark Elves to Sand Worms, the Words Behind World-Building . Penguin Books. UGA Working Papers in Linguistics, Vol 8, 2025

2015

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Bybee, J. (2015). Language Change. Cambridge University Press

work page 2015

[2] [2]

(2020, June)

Chandra, K., & Radvan, T. (2020, June). nearley: A parsing toolkit for JavaScript . https://doi.org/10.5281/zenodo.3897993

work page doi:10.5281/zenodo.3897993 2020

[3] [3]

Chomsky, N. (1956). Three models for the description of language.IRE Transactions on Information Theory , 2(3), 113–124. https://doi.org/10.1109/TIT.1956. 1056813 UGA Working Papers in Linguistics, Vol 8, 2025 Language Change and Variation in Toki Pona Page 14

doi:10.1109/tit.1956 1956

[4] [4]

(2024, January)

Comrie, B., Haspelmath, M., & Bickel, B. (2024, January). Leipzig Glossing Rules . https://www.eva.mpg.de/lingua/resources/glossing-rules.php

work page 2024

[5] [5]

(2024a, March)

Danielson, G., III. (2024a, March). When was pu added to toki pona? Retrieved March 1, 2024, from https://web.archive.org/web/20240301100156/https: //mun.la/lipu/pu.html

work page arXiv 2024

[6] [6]

(2025, April).sona-toki [GitHub repository]

Danielson, G., III. (2025, April).sona-toki [GitHub repository]. https://github.com/ gregdan3/sona-toki

work page 2025

[7] [7]

Dyer, W., Futrell, R., Liu, Z., & Scontras, G. (2020). Predicting cross-linguistic adjec- tive order with information gain . arXiv: 2012.15263 [cs.CL]. https://arxiv. org/abs/2012.15263

work page Pith review arXiv 2020

[8] [8]

Earley, J. (1970). An efficient context-free parsing algorithm.Commun. ACM, 13(2), 94–102. https://doi.org/10.1145/362007.362035

work page arXiv 1970