arxiv: 2509.17292 · v3 · submitted 2025-09-22 · 💻 cs.CL · cs.AI

Multi-View Attention Multiple-Instance Learning Enhanced by LLM Reasoning for Cognitive Distortion Detection

Jun Seo Kim , Hyemi Kim , Woo Joo Oh , Hongjin Cho , Hochul Lee , Hye Hyeon Kim This is my paper

Pith reviewed 2026-05-18 15:30 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords cognitive distortion detectionmultiple instance learninglarge language modelsmental health NLPmulti-view attentionELB decompositionsalience scores

0 comments

The pith

Decomposing utterances into emotion, logic, and behavior lets LLMs propose distortion instances with salience scores that a multi-view attention MIL model then classifies more accurately.

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

The paper sets out to solve the problem of detecting cognitive distortions in text, which are tied to mental health problems but difficult to identify automatically because of contextual ambiguity, overlap, and co-occurrence. It proposes breaking each utterance into Emotion, Logic, and Behavior components, then asking an LLM to generate multiple candidate distortion instances, each labeled with a type, the specific expression, and a model-assigned salience score. These candidates are pooled through a Multi-View Gated Attention layer inside a Multiple-Instance Learning classifier. Experiments on a Korean dataset and an English therapist-dialogue dataset show that the added ELB decomposition and salience scores raise overall classification performance, with the largest gains on distortions that are most open to interpretation. A reader would care if this hybrid LLM-MIL route really delivers more reliable and explainable detection than standard classifiers.

Core claim

The central claim is that LLM reasoning over ELB-decomposed utterances produces multiple distortion instances carrying salience scores, and that feeding these instances into a Multi-View Gated Attention MIL architecture yields higher classification accuracy than baselines, especially on high-ambiguity distortions, as measured on the KoACD Korean and Therapist QA English datasets.

What carries the argument

The Multi-View Gated Attention mechanism that pools LLM-generated distortion instances, each carrying a salience score derived from Emotion-Logic-Behavior decomposition of the original utterance.

If this is right

Classification performance rises when ELB decomposition and LLM-inferred salience scores are included in the MIL pipeline.
Gains are largest for cognitive distortions that carry high interpretive ambiguity.
The framework supplies expression-level reasoning that improves interpretability of the final classification.
The same pipeline generalizes from Korean to English therapy dialogues without language-specific retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach could be extended to live chat-based mental-health screening tools where quick, instance-level explanations matter.
If the salience scores prove stable across different LLMs, the method might reduce dependence on large human-annotated corpora for similar mental-health NLP tasks.
A natural next test would measure whether the model-assigned salience scores correlate with independent clinician ratings of distortion severity.

Load-bearing premise

The LLM-generated distortion instances and salience scores are accurate and psychologically valid enough to be used directly as inputs to the MIL classifier.

What would settle it

A side-by-side comparison in which human psychologists rate the same set of utterances and show low agreement with the LLM-proposed distortion types or salience rankings on a substantial fraction of cases would undermine the reliability of the generated inputs.

Figures

Figures reproduced from arXiv: 2509.17292 by Hochul Lee, Hongjin Cho, Hye Hyeon Kim, Hyemi Kim, Jun Seo Kim, Woo Joo Oh.

**Figure 2.** Figure 2: ELB-Based Psychological Decomposition of [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Multi-View MIL Architecture for Cognitive Distortion Classification [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Comparative Missing Rates of Cognitive Dis [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

read the original abstract

Cognitive distortions have been closely linked to mental health disorders, yet their automatic detection remains challenging due to contextual ambiguity, co-occurrence, and semantic overlap. We propose a novel framework that combines Large Language Models (LLMs) with a Multiple-Instance Learning (MIL) architecture to enhance interpretability and expression-level reasoning. Each utterance is decomposed into Emotion, Logic, and Behavior (ELB) components, which are processed by LLMs to infer multiple distortion instances, each with a predicted type, expression, and model-assigned salience score. These instances are integrated via a Multi-View Gated Attention mechanism for final classification. Experiments on Korean (KoACD) and English (Therapist QA) datasets demonstrate that incorporating ELB and LLM-inferred salience scores improves classification performance, especially for distortions with high interpretive ambiguity. Our results suggest a psychologically grounded and generalizable approach for fine-grained reasoning in mental health NLP. The dataset and implementation details are publicly accessible.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The ELB breakdown plus LLM-generated instances fed into multi-view gated attention MIL is the actual new piece, but the complete absence of human validation on those LLM labels is the part that needs fixing before the gains can be trusted.

read the letter

The paper's main move is to split each utterance into Emotion, Logic, and Behavior parts, let an LLM produce several candidate distortion instances with salience scores, and then classify the bag of instances with a multi-view gated attention MIL model. That specific pipeline for cognitive distortion detection looks new relative to the MIL and LLM baselines they cite. Experiments on the Korean KoACD set and the English Therapist QA set report gains, especially on the more ambiguous distortion types. The attempt to ground the model in psychologically motivated components and to surface explicit instances is a clear positive; it gives the output more structure than a plain classifier or a single LLM prompt would. The public release of data and code is also useful for anyone who wants to test the approach directly. The central weakness is the missing check on the LLM outputs themselves. The work trains the MIL attention directly on the inferred instances and salience scores without reporting human review, inter-annotator agreement, or error rates on either dataset. If the LLM tends to over- or under-assign certain distortions, the reported improvements could simply reflect that bias rather than genuine clinical signal. The abstract also leaves out baseline details and significance numbers, though the full paper may contain them. This is the kind of paper that belongs in a mental-health NLP reading group or a clinical text workshop. Readers who build tools for therapy support or screening would get concrete ideas from the decomposition and attention design, even if they end up adding their own validation layer. I would send it to peer review so referees can examine the full experimental tables and decide whether the LLM step holds up under scrutiny.

Referee Report

2 major / 1 minor

Summary. The paper proposes a framework combining Large Language Models (LLMs) with Multiple-Instance Learning (MIL) for cognitive distortion detection. Utterances are decomposed into Emotion, Logic, and Behavior (ELB) components; LLMs then infer multiple distortion instances each with a type, expression, and salience score. These are integrated via a Multi-View Gated Attention mechanism for final classification. Experiments on the Korean KoACD and English Therapist QA datasets claim performance improvements, especially for high-ambiguity distortions, with the dataset and code released publicly.

Significance. If the results hold after addressing validation gaps, the work could advance interpretable mental health NLP by grounding detection in psychologically motivated ELB decomposition and expression-level LLM reasoning. The public release of the dataset and implementation details is a clear strength supporting reproducibility.

major comments (2)

[§4 (Experiments)] §4 (Experiments) and abstract: Performance gains are reported from ELB decomposition plus LLM-inferred salience scores, yet no statistical significance tests, detailed baseline descriptions, or error analysis are provided. This is load-bearing for the central claim that improvements are robust and especially pronounced for high-ambiguity distortions.
[§3 (Method)] §3 (Method): The pipeline trains the MIL attention directly on LLM-generated distortion instances and salience scores without any reported human validation, inter-annotator agreement, or error rates on either KoACD or Therapist QA. This is load-bearing because the claim that these inputs improve classification for interpretive ambiguity rests on the unverified assumption that the LLM outputs are accurate psychological proxies rather than artifacts.

minor comments (1)

[Abstract] Abstract: The phrase 'psychologically grounded' would benefit from a brief explicit link to established cognitive-behavioral frameworks to clarify the ELB alignment.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We appreciate the recognition of the framework's potential contribution to interpretable mental health NLP and the value of releasing the dataset and code. We address the two major comments below and will revise the manuscript accordingly to strengthen the empirical support for our claims.

read point-by-point responses

Referee: §4 (Experiments) and abstract: Performance gains are reported from ELB decomposition plus LLM-inferred salience scores, yet no statistical significance tests, detailed baseline descriptions, or error analysis are provided. This is load-bearing for the central claim that improvements are robust and especially pronounced for high-ambiguity distortions.

Authors: We agree that statistical significance testing, expanded baseline details, and error analysis are necessary to robustly support the central claims. In the revised manuscript we will add McNemar's tests (or bootstrap confidence intervals) comparing our full model against ablations and baselines on both KoACD and Therapist QA. We will expand the baseline descriptions to include exact architectures, training procedures, and hyperparameter choices. We will also insert a dedicated error analysis subsection that stratifies results by interpretive ambiguity level, highlighting cases where the multi-view attention and salience scores yield gains and where they do not. revision: yes
Referee: §3 (Method): The pipeline trains the MIL attention directly on LLM-generated distortion instances and salience scores without any reported human validation, inter-annotator agreement, or error rates on either KoACD or Therapist QA. This is load-bearing because the claim that these inputs improve classification for interpretive ambiguity rests on the unverified assumption that the LLM outputs are accurate psychological proxies rather than artifacts.

Authors: We acknowledge that the current manuscript does not report human validation of the LLM-generated ELB instances or salience scores. While the observed performance lift on held-out test sets provides indirect evidence of utility, we agree that direct verification would better ground the assumption that these outputs serve as reliable psychological proxies. In the revision we will add a human evaluation on a stratified sample of generated instances from both datasets, reporting inter-annotator agreement (Cohen's kappa) and error rates for distortion type, expression, and salience. We will also discuss potential LLM artifacts and how the gated attention mechanism is intended to down-weight unreliable instances. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain; empirical results are independent of inputs

full rationale

The paper proposes an empirical pipeline that decomposes utterances into ELB components, uses LLMs to generate distortion instances and salience scores, and feeds them into a Multi-View Gated Attention MIL classifier. Central claims rest on reported classification improvements on the KoACD and Therapist QA datasets rather than any closed-form derivation, fitted parameter renamed as prediction, or self-citation that reduces the outcome to the input by construction. No equations or uniqueness theorems are invoked that would create self-definitional or load-bearing circularity. The framework is self-contained against external benchmarks via standard train/test evaluation, making this the normal non-circular finding for an applied ML paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach assumes LLMs can reliably extract psychologically meaningful ELB components and distortion instances; no free parameters or invented entities are described in the abstract.

axioms (1)

domain assumption LLM inferences on ELB components produce valid distortion instances and salience scores
Central to the pipeline but not validated in the provided abstract

pith-pipeline@v0.9.0 · 5712 in / 1114 out tokens · 26872 ms · 2026-05-18T15:30:58.700126+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Each utterance is decomposed into Emotion, Logic, and Behavior (ELB) components... integrated via a Multi-View Gated Attention mechanism
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean LogicNat unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

MIL framework... bag of LLM-generated instances... salience score

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 2 internal anchors

[1]

Stuart Andrews, Ioannis Tsochantaridis, and Thomas Hofmann. 2002. Support vector machines for multiple-instance learning. In Advances in Neural Information Processing Systems, volume 15

work page 2002
[2]

Anthropic. 2025. https://www.anthropic.com/claude/sonnet Claude 3.7 sonnet

work page 2025
[3]

Aaron T. Beck. 1979. Cognitive therapy and the emotional disorders. Penguin

work page 1979
[4]

Zhiyu Chen, Yujie Lu, and William Yang Wang. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.284 Empowering psychotherapy with large language models: Cognitive distortion detection through diagnosis of thought prompting . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4295--4304, Singapore. Association for Computatio...

work page doi:10.18653/v1/2023.findings-emnlp.284 2023
[5]

Dietterich, Richard H

Thomas G. Dietterich, Richard H. Lathrop, and Tom \'a s Lozano-P \'e rez. 1997. https://doi.org/10.1016/S0004-3702(96)00034-3 Solving the multiple instance problem with axis-parallel rectangles . Artificial Intelligence, 89(1-2):31--71

work page doi:10.1016/s0004-3702(96)00034-3 1997
[6]

Xiruo Ding, Kevin Lybarger, Justin Tauscher, and Trevor Cohen. 2022. https://doi.org/10.18653/v1/2022.naacl-srw.9 Improving classification of infrequent cognitive distortions: Domain-specific model vs. data augmentation . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Te...

work page doi:10.18653/v1/2022.naacl-srw.9 2022
[7]

Nada Elsharawi and Alia El Bolock. 2024. https://aclanthology.org/2024.lrec-main.286/ C-journal: A journaling application for detecting and classifying cognitive distortions using deep-learning based on a crowd-sourced dataset . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-...

work page 2024
[8]

Google DeepMind . 2024. https://deepmind.google/technologies/gemini/flash/ Gemini 2.0 flash

work page 2024
[9]

Maximilian Ilse, Jakub Tomczak, and Max Welling. 2018. Attention-based deep multiple instance learning. In Proceedings of the 35th International Conference on Machine Learning (ICML 2018), pages 2127--2136. PMLR

work page 2018
[10]

Zach Jorgensen, Yan Zhou, and Meador Inge. 2008. A multiple instance learning strategy for combating good word attacks on spam filters. Journal of Machine Learning Research, 9(6)

work page 2008
[11]

Kaplan, A.S

S.C. Kaplan, A.S. Morrison, P.R. Goldin, R.T. Olino, and J. Gross. 2017. https://doi.org/10.1007/s10608-017-9838-9 The cognitive distortions questionnaire (cd-quest): Validation in a sample of adults with social anxiety disorder . Cognitive Therapy and Research, 41:576--587

work page doi:10.1007/s10608-017-9838-9 2017
[12]

K \"u rsat Mustafa Karao g lan. 2024. https://doi.org/10.1016/j.neucom.2024.128263 Novel approaches for fake news detection based on attention-based deep multiple-instance learning using contextualized neural language models . Neurocomputing, 602:128263

work page doi:10.1016/j.neucom.2024.128263 2024
[13]

Jun Seo Kim and Hye Hyeon Kim. 2025. https://arxiv.org/abs/2505.00367 Koacd: The first korean adolescent dataset for cognitive distortion analysis . Preprint, arXiv:2505.00367

work page arXiv 2025
[14]

Dimitrios Kotzias, Misha Denil, Phil Blunsom, and Nando de Freitas. 2014. https://arxiv.org/abs/1411.3128 Deep multi-instance transfer learning . Preprint, arXiv:1411.3128

work page internal anchor Pith review Pith/arXiv arXiv 2014
[15]

Jiexi Liu, Dehan Kong, Longtao Huang, Dinghui Mao, and Hui Xue. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.546 Multiple instance learning for offensive language detection . In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 7387--7396, Abu Dhabi, UAE. Association for Computational Linguistics

work page doi:10.18653/v1/2022.findings-emnlp.546 2022
[16]

Kevin Lybarger, Justin Tauscher, Xiruo Ding, Dror Ben-Zeev, and Trevor Cohen. 2022. https://doi.org/10.18653/v1/2022.clpsych-1.11 Identifying distorted thinking in patient-therapist text message exchanges by leveraging dynamic multi-turn context . In Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology, pages 126--136, S...

work page doi:10.18653/v1/2022.clpsych-1.11 2022
[17]

Oded Maron and Aparna Lakshmi Ratan. 1998. Multiple-instance learning for natural scene classification. In Proceedings of the 15th International Conference on Machine Learning (ICML 1998), pages 341--349

work page 1998
[18]

McGrath, Ali Al-Hamzawi, Jordi Alonso, Yasmin Altwaijri, Laura H

John J. McGrath, Ali Al-Hamzawi, Jordi Alonso, Yasmin Altwaijri, Laura H. Andrade, Evelyn J. Bromet, Ronny Bruffaerts, Jos \'e Miguel Caldas-de Almeida, Stephanie Chardoul, Wai Tat Chiu, Louisa Degenhardt, Olga V. Demler, Finola Ferry, Oye Gureje, Josep Maria Haro, Elie G. Karam, Georges Karam, Salma M. Khaled, Viviane Kovess-Masfety, and 40 others. 2023....

work page doi:10.1016/s2215-0366(23)00193-1 2023
[19]

Nihan Mercan, Merve Bulut, and C i g dem Y \"u ksel. 2023. https://doi.org/10.1007/s12144-021-02251-z Investigation of the relatedness of cognitive distortions with emotional expression, anxiety, and depression . Current Psychology, 42:2176--2185

work page doi:10.1007/s12144-021-02251-z 2023
[20]

Morrison, Carrie M

Amanda S. Morrison, Carrie M. Potter, Matthew M. Carper, Dina G. Kinner, Dane Jensen, Laura Bruce, Judy Wong, Irismar Reis de Oliveira, Donna M. Sudak, and Richard G. Heimberg. 2015. https://doi.org/10.1521/ijct.2015.8.4.287 The cognitive distortions questionnaire (cd-quest): Psychometric properties and exploratory factor analysis . International Journal ...

work page doi:10.1521/ijct.2015.8.4.287 2015
[21]

OpenAI . 2023. https://arxiv.org/abs/2303.08774 Gpt-4 technical report . Preprint, arXiv:2303.08774

work page internal anchor Pith review Pith/arXiv arXiv 2023
[22]

Nikolaos Pappas and Andrei Popescu-Belis. 2014. https://doi.org/10.3115/v1/D14-1052 Explaining the stars: Weighted multiple-instance learning for aspect-based sentiment analysis . In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 455--466, Doha, Qatar. Association for Computational Linguistics

work page doi:10.3115/v1/d14-1052 2014
[23]

Hongzhi Qi, Qing Zhao, Jianqiang Li, Changwei Song, Wei Zhai, Dan Luo, Shuo Liu, Yi Jing Yu, Fan Wang, Huijing Zou, Bing Xiang Yang, and Guanghui Fu. 2023. https://doi.org/10.21203/rs.3.rs-3523508/v1 Supervised learning and large language model benchmarks on mental health datasets: Cognitive distortions and suicidal risks in chinese social media . Preprin...

work page doi:10.21203/rs.3.rs-3523508/v1 2023
[24]

Benjamin Shickel, Scott Siegel, Martin Heesacker, Sherry Benton, and Parisa Rashidi. 2020. https://doi.org/10.1109/BIBE50027.2020.00052 Automatic detection and classification of cognitive distortions in mental health text . In Proceedings of the 2020 IEEE 20th International Conference on Bioinformatics and Bioengineering (BIBE), pages 275--280. IEEE

work page doi:10.1109/bibe50027.2020.00052 2020
[25]

Sagarika Shreevastava and Peter W. Foltz. 2021. https://doi.org/10.18653/v1/2021.clpsych-1.17 Detecting cognitive distortions from patient-therapist interactions . In Proceedings of the Seventh Workshop on Computational Linguistics and Clinical Psychology: Improving Access, pages 151--158. Association for Computational Linguistics

work page doi:10.18653/v1/2021.clpsych-1.17 2021
[26]

Taetem Simms, Christopher Ramstedt, Marc Rich, Matthew Richards, Thomas Martinez, and Christophe Giraud-Carrier. 2017. https://doi.org/10.1109/ICHI.2017.39 Detecting cognitive distortions through machine learning text analytics . In Proceedings of the 2017 IEEE International Conference on Healthcare Informatics (ICHI), pages 508--512. IEEE

work page doi:10.1109/ichi.2017.39 2017
[27]

Strohmeier, Brad Rosenfield, Robert A

Craig W. Strohmeier, Brad Rosenfield, Robert A. DiTomasso, and J. Russell Ramsay. 2016. https://doi.org/10.1016/j.psychres.2016.02.034 Assessment of the relationship between self-reported cognitive distortions and adult adhd, anxiety, depression, and hopelessness . Psychiatry Research, 238:153--158

work page doi:10.1016/j.psychres.2016.02.034 2016
[28]

Putu Gede Harsa Suputra, Linawati, Ni Putu Sastra, Gede Sukadarmika, Ni Kadek Anindya Saraswati, Diah Purwitasari Er, and I Made Agus Setiawan

I. Putu Gede Harsa Suputra, Linawati, Ni Putu Sastra, Gede Sukadarmika, Ni Kadek Anindya Saraswati, Diah Purwitasari Er, and I Made Agus Setiawan. 2023. https://doi.org/10.1109/ICSGTEIS60500.2023.10424225 Detection and classification of cognitive distortions: A literature review . In Proceedings of the 2023 International Conference on Smart-Green Technolo...

work page doi:10.1109/icsgteis60500.2023.10424225 2023
[29]

Tauscher, Kevin Lybarger, Xiruo Ding, Ayesha Chander, William J

Justin S. Tauscher, Kevin Lybarger, Xiruo Ding, Ayesha Chander, William J. Hudenko, Trevor Cohen, and Dror Ben-Zeev. 2023. https://doi.org/10.1176/appi.ps.202100692 Automated detection of cognitive distortions in text exchanges between clinicians and people with serious mental illness . Psychiatric Services, 74(4):407--410

work page doi:10.1176/appi.ps.202100692 2023
[30]

Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. 2021. https://doi.org/10.18653/v1/2021.findings-acl.188 Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2140--2151. Association for Computational Linguistics

work page doi:10.18653/v1/2021.findings-acl.188 2021
[31]

World Health Organization . 2022. World mental health report: Transforming mental health for all

work page 2022
[32]

Goldman, Wei Yu, and Jason E

Qi Zhang, Sally A. Goldman, Wei Yu, and Jason E. Fritts. 2008. https://doi.org/10.1109/ICPR.2008.4761441 Em-dd: An improved multiple-instance learning algorithm . In Proceedings of the 19th IEEE International Conference on Pattern Recognition, pages 1--4. IEEE

work page doi:10.1109/icpr.2008.4761441 2008
[33]

Zhi Zhao, Yang Zhang, and Xindong Wu. 2017. Multiple instance learning via deep kernel embedding. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI 2017), pages 3442--3448. AAAI Press

work page 2017
[34]

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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

work page
[35]

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...

work page