Pith. sign in

REVIEW 4 major objections 5 minor 28 references

Transparent NLP: Using RAG and LLM Alignment for Privacy Q&A

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Retrieval-augmented generation with a self-alignment module produces privacy answers that score better than plain RAG on most of 21 quality metrics, though still below human experts.

desk verdict A genuinely honest feasibility study of RAG+RAIN for GDPR privacy Q&A, weakened by a setup leak and no significance testing but worth refereeing. read the letter →

arxiv 2502.06652 v1 pith:VSAIU22O submitted 2025-02-10 cs.CL

Anna Leschanowsky , Zahra Kolagar , Erion Çano , Ivan Habernal , Dara Hallinan , Emanuël A. P. Habets , Birgit Popp This is my paper · ORCID
classification cs.CL
keywords RetrievalAugmentedGenerationRAINMultiGDPRtransparencyprivacyquestionansweringLLMalignmenttextevaluationmetricsreadability
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 asks whether a question-answering system can be built to meet the GDPR transparency principle, which demands answers that are precise and comprehensible. It compares a plain RAG system that retrieves privacy documents and generates answers with versions that add a self-alignment loop (RAIN, and a new multidimensional variant, MultiRAIN) that rewinds and regenerates text to satisfy preciseness and comprehensibility criteria. The central claim is that alignment-enhanced RAG beats plain RAG on 18 of 21 metrics, while no system matches the gold-standard human expert answers. The paper also argues that the 21 metrics do not cleanly separate into the two legal constructs, as shown by principal component analysis, and that metric refinement is needed before such systems can discharge legal obligations.

What carries the argument

The machinery is Rewindable Auto-regressive Inference (RAIN), a tree-search decoding procedure, plus MultiRAIN, a multidimensional version. RAIN treats the user query as the tree stem and generated words or partial sentences as leaves; a self-evaluation prompt scores candidate continuations on the target properties, low-scoring branches are rewound and regenerated, and an exploration term prevents premature convergence. MultiRAIN extends this by combining multiple metric scores with a function f that averages them and multiplies by a penalty factor p that drops below 1 when any score falls under a threshold. This lets a system optimise preciseness and comprehensibility at the same time without fine-tuning or extra training data.

What would settle it

Run a blind evaluation in which privacy law experts rank anonymised responses from VanillaRAG, RAIN, MultiRAIN, and human authors; if the experts do not systematically prefer the aligned systems over VanillaRAG, the claim that alignment improves transparency would be refuted despite the metric scores.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that adding RAIN-style self-alignment to a RAG pipeline is a feasible route to GDPR transparency: in 18 of 21 metrics, most alignment-based systems outperform VanillaRAG, and at least one aligned system beats the baseline on every metric. MultiRAIN, the proposed extension, generalises RAIN's rewindable tree search from one optimisation target to several by combining metric values through a function f that penalises any score below a threshold, so preciseness and comprehensibility can be optimised jointly without retraining. The results also show that no implementation reaches full agreement with human-written gold-standard answers, and that the alignment gains do not generalise across models or prompt definitions; Experiment 3, using deterministic metrics, confirms the alignment mechanism works as intended while exposing how far it is from perfect.

Load-bearing premise

The results stand on the assumption that 21 technical metrics, and the unpublished expert answers used as thresholds and baselines, actually capture the GDPR constructs of preciseness and comprehensibility.

Editorial extensions

If this is right

  • Adding a self-alignment module to a RAG pipeline is a cheap, training-free way to push generated answers closer to human-quality transparency on most metrics.
  • MultiRAIN demonstrates joint optimisation of multiple criteria within one decoding pass, so systems can target preciseness and comprehensibility simultaneously.
  • Because alignment gains do not generalise across models or metric definitions, deployment in legal settings needs per-system, per-metric validation.
  • The PCA result implies that metric choice changes conclusions: some LLM-as-judge and deterministic metrics behave differently than their labels suggest.
  • Current runtime (20-58 hours for 42 answers on one GPU) means the approach is not yet practical for real-time transparency answers, a limitation the authors state.

Reading between the lines

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

  • A natural next experiment would test whether alignment gains survive when retrieval returns partially wrong documents, since RAIN only optimises generation and inherits retrieval errors.
  • The same alignment-plus-RAG template could be applied to other regulated domains, such as financial disclosures or medical consent, where preciseness and comprehensibility are also legal requirements.
  • The paper's cost categorisation hints that low-cost deterministic metrics may substitute for expensive LLM-as-judge metrics in real-time alignment, an implication the paper does not assert directly.
  • Because none of the systems beat human answers, a plausible regulatory stance is that such tools should assist human drafters rather than replace them; the paper leaves this policy implication implicit.
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

4 major / 5 minor

Summary. The paper investigates whether Retrieval Augmented Generation (RAG) systems enhanced with alignment modules (RAIN and the proposed MultiRAIN) can produce privacy-related answers that better satisfy GDPR transparency requirements, operationalized as 'preciseness' and 'comprehensibility'. Using a dataset of 42 privacy questions with expert-generated gold-standard answers, the authors compare nine system variants across three experiments, evaluating outputs with 21 metrics. The central claim, stated in the abstract and Section 4.1, is that alignment-based RAG systems outperform a vanilla RAG baseline on most metrics, though none fully match human expert answers. The paper also presents a PCA of the metrics, arguing that the metric space is more complex than the assumed two-construct taxonomy.

Significance. If the main claim holds, the paper offers a practical demonstration that lightweight self-alignment (RAIN-style) can be plugged into a RAG pipeline to improve the precision and readability of privacy answers, a step toward using LLMs in legally sensitive transparency contexts. The paper also contributes a useful discussion of the difficulty of mapping legal constructs to NLP metrics, and the PCA analysis is a valuable caution about assuming that metrics cluster according to their a priori labels. However, the empirical foundation is weakened by two issues that affect the headline comparison: alignment thresholds in Experiments 2 and 3 are derived from the same gold-standard answers used as the evaluation baseline, and the quantitative results are reported as descriptive percentages without measures of uncertainty. These issues are addressable but require re-analysis or additional experiments.

major comments (4)
  1. [Section 4.1] The alignment thresholds for Experiments 2 and 3 are computed from the same human gold-standard answers that are later used as the evaluation baseline. Specifically, Section 3.4 states thresholds of 90.74 (readability) and 78.64 (correctness) in Experiment 2, and 62.69 (readability) and 0.312 (BERT) in Experiment 3, all 'based on our human-generated answers.' Section 3.6 then defines the primary evaluation metric as the percentage of responses 'at least as good as' the same human answers, with thresholds derived from the human score distribution (e.g., meeting or exceeding the lowest human score for 'bigger is better' metrics). This creates a circularity: MultiRAIN is explicitly optimized to clear thresholds that are defined by the evaluation target, while VanillaRAG receives no such tuning. The observed advantage of alignment systems on the headline metric could therefore be an artifact of threshold information rather than a genuine improvement in preciseness or comprehensibility. The authors should either use a held-out subset of the gold-standard answers for threshold selection, or demonstrate that the comparative results are robust to thresholds chosen without access to the evaluation answers.
  2. [Section 3.1 / References] The central claim that alignment-based systems outperform VanillaRAG on 18 of 21 metrics rests on descriptive percentages computed over only 42 questions, with no confidence intervals, significance tests, or effect-size measures. With n=42, the difference between, say, 60% and 70% 'at least as good as human' is easily within sampling noise. The authors should provide uncertainty quantification (e.g., bootstrap confidence intervals) or statistical tests for the system comparisons, or temper the claim accordingly. Without this, the headline result is not yet supported.
  3. [Section 3.2.2] The gold-standard dataset (Leschanowsky et al., 2025) is described as an unpublished manuscript, and the paper does not include the data or a detailed account of its construction. This limits reproducibility and makes it impossible for readers to assess the quality of the human answers that are central to both threshold setting and evaluation. The authors should either release the dataset or provide a comprehensive description of its creation, inter-annotator agreement, and access conditions.
  4. [Section 4.2] The paper introduces MultiRAIN as a multidimensional extension of RAIN, but the results show that 'none of the MultiRAIN implementations outperformed RAIN on both preciseness and comprehensibility evaluation' (Section 4.1). This weakens the novelty claim: the reader is left with the impression that the proposed method is not yet competitive with the existing RAIN baseline. The authors should clarify whether the contribution of MultiRAIN is the algorithm design alone (with the negative result acknowledged) or whether there is any setting where MultiRAIN provides a measurable advantage over RAIN.
minor comments (5)
  1. [Section 3.2.2] The text around Equation (1) has a typo: 'The valueu(Yi:j; Y1:i−1)' should read 'The value u(Yi:j; Y1:i−1)' with a space.
  2. [Section 3.7] Equation (1) is numbered, but the subsequent formulation for f is labeled 'Eq. B'. Please use a consistent numbering scheme.
  3. [Appendix C] The section title uses 'Principle component analysis' while the text and Section 4.2 use 'Principal component analysis'. Please correct the spelling.
  4. [Appendix C] In Algorithm 1, line 16 refers to 'Equation (2) as in Li (2023)'. The reference should presumably be to Li et al. (2024), consistent with the rest of the paper.
  5. [Conclusion] The reference to Leschanowsky et al. (2025) is given as 'Unpublished manuscript, submitted for publication at CHI Late Breaking Papers 2025, Privacy Track.' Please update if the paper has been published or accepted, and consider providing a link or preprint identifier.

Circularity Check

2 steps flagged · score 6.0 of 10

Experiment 2 and 3 alignment thresholds are calibrated to the same human gold-standard answers that define the evaluation's 'at least as good as human' criterion, so the reported advantage of aligned systems over VanillaRAG on target metrics is partly by construction.

  1. fitted input called prediction [Section 3.4 (Experiment 2) and Section 3.6 (Evaluation procedure)]
    "Based on our human-generated answers, we applied a threshold of 90.74 for readability and 78.64 for correctness. ... For all other metrics (“bigger is better”), responses had to meet or exceed the lowest human threshold."

    The threshold used by the MultiRAIN alignment gate is computed from the same human gold-standard answers that later define the evaluation criterion 'at least as good as human.' Since Experiment 2 uses similar definitions (i.e., prompts) for alignment and evaluation metrics, a response that clears the alignment penalty threshold is, by construction, counted as meeting or exceeding the lowest human threshold in the evaluation. VanillaRAG receives no such threshold, so the reported advantage on correctness and readability reflects optimization against the evaluation target rather than an independent measurement of transparency.

  2. self definitional [Section 3.3 (Experiment 3), Section 3.4 (Experiment 3), and Section 3.6 (Evaluation procedure)]
    "Experiment 3: Uses statistical metrics for alignment, with similar definitions (i.e., the same implementations) for optimized and evaluated metrics. ... Similar to Experiment 2, we applied thresholds for the MultiRAIN implementation of 62.69 and 0.312 for readability and BERT, respectively. ... responses had to meet or exceed the lowest human threshold."

    In Experiment 3 the alignment and evaluation metrics are literally the same implementations of Flesch-Kincaid Readability and BERT. The thresholds 62.69 and 0.312 were derived from human gold-standard answers, and the evaluation defines 'at least as good as human' as meeting or exceeding the lowest human threshold on the same metrics. Therefore, MultiRAIN is gating and stopping on the very criterion used to declare success; the resulting 'percentage at least as good as human' for these metrics measures whether the search terminated above its own threshold, not an independent property. This feeds the abstract's 'outperform baseline RAG systems on most metrics' claim.

full rationale

The paper is not wholly circular: RAIN is an external method, MultiRAIN's multidimensional combination function is a genuine algorithmic extension, Experiment 1 uses a different construct and a different judge model than the evaluation, and the PCA analysis is self-contained empirical work. However, the central comparative claim—that alignment-based systems outperform VanillaRAG on most of the 21 metrics—is partially by construction in Experiments 2 and 3. In those experiments, the MultiRAIN penalty thresholds are fitted to the same human gold-standard answers that define the evaluation baseline, and the evaluation counts a response as 'at least as good as human' precisely when it meets or exceeds a human-derived threshold. For the target metrics (correctness, readability, BERT, Flesch-Kincaid), the alignment objective and the evaluation outcome are the same criterion, so the observed advantage over the unaligned VanillaRAG on those metrics is an artifact of optimizing against the evaluation target rather than independent evidence of GDPR transparency. The gold-standard dataset is an unpublished same-author artifact (Leschanowsky et al., 2025) used for question-answering, alignment, and evaluation, which further increases the burden, although this is primarily a data-provenance concern rather than a formal circular step. Because only a subset of metrics reduces this way and the abstract's claim also relies on non-target metrics, the circularity is partial rather than total, warranting a score of 6.

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

The central claims rest primarily on measurement choices (metric selection and gold-standard definition) rather than on fitted physical constants. The key free parameters are the alignment thresholds fitted to the gold-standard answers, which carry most of the circularity burden. MultiRAIN is an algorithm, not a postulated physical or conceptual entity.

free parameters (5)
  • Readability threshold (Experiment 2) = 90.74
    Chosen from human gold-standard answers in Section 3.4; used as a penalty threshold in MultiRAIN alignment.
  • Correctness threshold (Experiment 2) = 78.64
    Chosen from human gold-standard answers; used as a penalty threshold in MultiRAIN alignment.
  • Readability threshold (Experiment 3) = 62.69
    Chosen from human gold-standard answers; used as a penalty threshold for MultiRAIN alignment.
  • BERT threshold (Experiment 3) = 0.312
    Chosen from human gold-standard answers; multiplied by 100 before averaging, used as a penalty threshold.
  • RAIN regularization hyperparameter c
    Referenced in Equation 1 but value not reported; balances exploration and exploitation in the tree search.
assumptions (4)
  • domain assumption The 21 metrics capture GDPR 'preciseness' and 'comprehensibility'.
    Section 3.5 explicitly calls the mapping a 'good-faith attempt' and notes no standard exists.
  • domain assumption The human-generated gold-standard answers in Leschanowsky et al. (2025) are valid legal-linguistic references.
    The dataset is unpublished, authored by the same group, and not independently validated.
  • domain assumption GPT-4-based LLM-as-a-judge scores are reliable measures of correctness and readability.
    Correctness lacks validation; readability relies on a single cited correlation (Trott, 2024).
  • domain assumption The Mistral-7B self-evaluation used inside RAIN and MultiRAIN is a reliable reward signal.
    No calibration of self-evaluation scores is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transparent NLP: Using RAG and LLM Alignment for Privacy Q&A." pith.science (2026). https://pith.science/paper/VSAIU22O

@misc{pith2026250206652,
  author       = {Pith},
  title        = {Pith review of: Transparent NLP: Using RAG and LLM Alignment for Privacy Q&A},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VSAIU22O}},
  note         = {Machine review of arXiv:2502.06652}
}
read the original abstract

The transparency principle of the General Data Protection Regulation (GDPR) requires data processing information to be clear, precise, and accessible. While language models show promise in this context, their probabilistic nature complicates truthfulness and comprehensibility. This paper examines state-of-the-art Retrieval Augmented Generation (RAG) systems enhanced with alignment techniques to fulfill GDPR obligations. We evaluate RAG systems incorporating an alignment module like Rewindable Auto-regressive Inference (RAIN) and our proposed multidimensional extension, MultiRAIN, using a Privacy Q&A dataset. Responses are optimized for preciseness and comprehensibility and are assessed through 21 metrics, including deterministic and large language model-based evaluations. Our results show that RAG systems with an alignment module outperform baseline RAG systems on most metrics, though none fully match human answers. Principal component analysis of the results reveals complex interactions between metrics, highlighting the need to refine metrics. This study provides a foundation for integrating advanced natural language processing systems into legal compliance frameworks.

Figures

Figures reproduced from arXiv: 2502.06652 by the authors.

Figure 1
Figure 1. Evaluation metrics presented as subplots: (a) LLM-as-Judge Metrics, (b) Statistical Metrics for Correctness [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. These 2D PCA projections show relationships [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Screenshots of a part of a conversation with [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Explained variance by principal components. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 17 canonical work pages

  1. [1]

    Wasi Ahmad, Jianfeng Chi, Yuan Tian, and Kai-Wei Chang. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.66 P olicy QA : A reading comprehension dataset for privacy policies . In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 743--749, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.findings-emnlp.66 2020

  2. [2]

    Article 29 Data Protection Working Party . 2018. https://ec.europa.eu/newsroom/article29/redirection/document/51025 Guidelines on Transparency under Regulation 2016/679 . Accessed: 2024-11-22

    work page 2018

  3. [3]

    Brown, Jack Clark, Sam McCandlish, Christopher Olah, and Jared Kaplan

    Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones, Nicholas Joseph, Benjamin Mann, Nova Dassarma, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, John Kernion, Kamal Ndousse, Catherine Olsson, Dario Amodei, Tom B. Brown, Jack Clark, Sam McCandlish, Christopher Olah, and Jared Kaplan. 2021. https://api.semanticsch...

    arXiv 2021

  4. [4]

    Rochelle A. Cadogan. 2004. An imbalance of power: the readability of internet privacy policies. Journal of Business & Economics Research (JBER), 2

    work page 2004

  5. [5]

    Daniel Cer, Mona Diab, Eneko Agirre, I \ n igo Lopez-Gazpio, and Lucia Specia. 2017. https://doi.org/10.18653/v1/S17-2001 S em E val-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation . In Proceedings of the 11th International Workshop on Semantic Evaluation ( S em E val) , pages 1--14, Vancouver, Canada. Association...

    doi:10.18653/v1/s17-2001 2017

  6. [6]

    Edgar Dale and Jeanne S. Chall. 1949. http://www.jstor.org/stable/41383594 The concept of readability . Elementary English, 26(1):19--26

    work page arXiv 1949

  7. [7]

    European Union . 2016. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX Official Journal of the European Union, L 119/1. Accessed: 2024-11-21

    work page 2016

  8. [8]

    Owain Evans, Owen Cotton-Barratt, Lukas Finnveden, Adam Bales, Avital Balwit, Peter Wills, Luca Righetti, and William Saunders. 2021. https://doi.org/10.48550/arXiv.2110.06674 Truthful AI : Developing and governing AI that does not lie

    arXiv doi:10.48550/arxiv.2110.06674 2021

Show all 28 references
  1. [9]

    Benjamin Fabian, Tatiana Ermakova, and Tino Lentz. 2017. https://doi.org/10.1145/3106426.3106427 Large-scale readability analysis of privacy policies

    2017

  2. [10]

    Forbes, Parth Katlana, and Zeydy Ortiz

    Grant C. Forbes, Parth Katlana, and Zeydy Ortiz. 2023. https://api.semanticscholar.org/CorpusID:264147021 Metric ensembles for hallucination detection . ArXiv, abs/2310.10495

    2023 arXiv

  3. [11]

    Robert Friel and Atindriyo Sanyal. 2023. https://api.semanticscholar.org/CorpusID:264590664 Chainpoll: A high efficacy method for llm hallucination detection . ArXiv, abs/2310.18344

    2023 arXiv

  4. [12]

    Galileo AI . 2024. https://docs.rungalileo.io/galileo/galileo-gen-ai-studio/guardrail-metrics/ Guardrail metrics . Accessed: 2024-11-25

    2024

  5. [13]

    Shin, and Karl Aberer

    Hamza Harkous, Kassem Fawaz, Kang G. Shin, and Karl Aberer. 2016. https://api.semanticscholar.org/CorpusID:12173813 Pribots: Conversational privacy with chatbots . In WSF@SOUPS

    2016

  6. [14]

    Yue Huang, Lichao Sun, Haoran Wang, Siyuan Wu, Qihui Zhang, Yuan Li, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, Xiner Li, Hanchi Sun, Zhengliang Liu, Yixin Liu, Yijue Wang, Zhikun Zhang, Bertie Vidgen, Bhavya Kailkhura, Caiming Xiong, Chaowei Xiao, Chunyuan Li, Eric P....

    2024

  7. [15]

    Peter Kincaid, Robert P

    J. Peter Kincaid, Robert P. Fishburne Jr. Fishburne, Richard L. Rogers Rogers, and Brad S. Chissom. 1975. Derivation of new readability formulas: (automated readability index, fog count and flesch reading ease formula) for navy enlisted personnel. Technical report, Chief of Na...

    1975

  8. [16]

    Anna Leschanowsky, Farnaz Salamatjoo, Zahra Kolagar, and Birgit Popp. 2025. Expert-generated privacy q&a dataset for conversational AI and user study insights. Unpublished manuscript, submitted for publication at CHI Late Breaking Papers 2025, Privacy Track

    2025

  9. [17]

    Li, FangyunWei, J

    Y. Li, FangyunWei, J. Zhao, C. Zhang, and H. Zhang. 2024. RAIN : Your language models can align themselves without finetuning. In International Conference on Learning Representations

    2024

  10. [18]

    LlamaIndex . 2024. https://docs.llamaindex.ai/en/stable/module_guides/evaluating/root.html Evaluating with LlamaIndex . Accessed: 2024-11-25

    2024

  11. [19]

    Eric Martínez, Francis Mollica, and Edward Gibson. 2023. https://doi.org/10.1073/pnas.2302672120 Even lawyers do not like legalese . Proceedings of the National Academy of Sciences, 120(23):e2302672120

    2023 doi

  12. [20]

    Saeed Mehrpour and Abdolmehdi Riazi. 2004. The impact of text length on reading comprehension in English as a second language. Asian EFL Journal, 3(6):1--14

    2004

  13. [21]

    Kanthashree Mysore Sathyendra, Shomir Wilson, Florian Schaub, Sebastian Zimmeck, and Norman Sadeh. 2017. https://doi.org/10.18653/v1/D17-1294 Identifying the provision of choices in privacy policy text . In Proceedings of the 2017 Conference on Empirical Methods in Natural Lan...

    2017 doi

  14. [22]

    Abhilasha Ravichander, Alan W Black, Shomir Wilson, Thomas Norton, and Norman Sadeh. 2019. https://doi.org/10.18653/v1/D19-1500 Question answering for privacy policies: Combining computational and legal perspectives . In Proceedings of the Conference on Empirical Methods in Na...

    2019 doi

  15. [23]

    Nils Reimers and Iryna Gurevych. 2019. https://doi.org/10.18653/v1/D19-1410 Sentence-BERT : Sentence embeddings using siamese BERT -networks . In Proceedings of the Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Nat...

    2019 doi

  16. [24]

    Reuters. 2024. https://www.reuters.com/technology/italy-fines-openai-15-million-euros-over-privacy-rules-breach-2024-12-20/ Italy fines openai over chatgpt privacy rules breach . Reuters. Accessed: 2025-01-22

    2024

  17. [25]

    readability

    Sean Trott. 2024. https://seantrott.substack.com/p/measuring-the-readability-of-texts Measuring the "readability" of texts with large language models

    2024

  18. [26]

    Fred Zenker and Kristopher Kyle. 2021. https://doi.org/10.1016/j.asw.2020.100505 Investigating minimum text lengths for lexical diversity indices . Assessing Writing, 47:100505

    2021

  19. [27]

    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...

  20. [28]

    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 8, 2026 · model on record in the stance chip above.