Algospeak, Hiding in the Open: The Trade-off Between Legible Meaning and Detection Avoidance

Jan Fillies; Jeffrey Hancock; Ronald E. Robertson

arxiv: 2605.06619 · v1 · submitted 2026-05-07 · 💻 cs.CL · cs.CY

Algospeak, Hiding in the Open: The Trade-off Between Legible Meaning and Detection Avoidance

Jan Fillies , Ronald E. Robertson , Jeffrey Hancock This is my paper

Pith reviewed 2026-05-08 09:59 UTC · model grok-4.3

classification 💻 cs.CL cs.CY

keywords algospeakcontent moderationlinguistic evasiondetection avoidancemeaning preservationdisinformationmajority understandable modulationCOVID-19

0 comments

The pith

Algospeak creates a trade-off where more evasion lowers both detection risk and human comprehension.

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

The paper formalizes how Algospeak, or linguistic changes to evade content moderation, produces a joint drop in detectability and understandability as the level of alteration rises. It defines Majority Understandable Modulation as the specific threshold where further changes improve evasion but cause most recipients to lose the core meaning. The authors built a reproducible framework that generates tunable Algospeak variants from base sentences, applied it to COVID-19 disinformation to create a 700-item dataset across five levels and seven strategies, and tested the variants with multiple language models on meaning recovery and classification tasks. Curve fitting on the results estimates the MUM point and reveals consistent relationships between modulation, comprehension, and detection.

Core claim

In a joint action model of evasion and detection, increases in Algospeak reduce both detectability and understandability. Majority Understandable Modulation is the modulation level at which additional evasive alteration increases detector evasion but loses comprehension for the majority of recipients. Evaluations on a dataset of modulated COVID-19 disinformation items confirm these relationships through meaning recovery and classification tasks across language models.

What carries the argument

Majority Understandable Modulation (MUM), the modulation threshold that marks the point of maximum detector evasion before majority comprehension collapses.

If this is right

Evasion strategies will concentrate near the MUM threshold to balance reach and safety.
Content detectors must incorporate meaning-preservation checks to avoid over-moderating understandable messages.
The tunable framework enables sensitivity testing of different evasion strategies against various models.
The same approach can map trade-offs in other disinformation domains beyond COVID-19.

Where Pith is reading between the lines

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

Platform rules may evolve toward targeting content near or beyond the MUM point rather than all altered language.
The trade-off implies that complete detection of evasive content will inevitably reduce the reach of legitimate discourse that uses similar modulations.
Extending the evaluations to non-English languages or live platform data would test whether MUM estimates hold outside the tested setting.

Load-bearing premise

The modulation framework produces variants that genuinely preserve core meaning at lower levels and that LLM performance on meaning recovery and classification tasks serves as a valid proxy for human comprehension and real-world detector behavior.

What would settle it

A direct comparison showing whether actual human readers retain high understanding at the modulation levels where the LLM models predict majority comprehension loss, or whether deployed detectors exhibit the predicted rise in evasion without corresponding meaning collapse.

Figures

Figures reproduced from arXiv: 2605.06619 by Jan Fillies, Jeffrey Hancock, Ronald E. Robertson.

**Figure 1.** Figure 1: Strategy “Code Word”: Shows the number of code words required for each model to fail view at source ↗

**Figure 2.** Figure 2: Example misinformation item with five levels of modulation. view at source ↗

**Figure 3.** Figure 3: Overview of the relationship between modulation and understandability. view at source ↗

**Figure 4.** Figure 4: Example results for two of the seven Algospeak strategies for the model GPT-4o-mini. view at source ↗

**Figure 5.** Figure 5: Classifications of language based on modulation and understandability. view at source ↗

**Figure 6.** Figure 6: Classifications of language based on modulation and understandability. view at source ↗

**Figure 7.** Figure 7: Heatmap detection adjusted R2 by strategy and model view at source ↗

**Figure 8.** Figure 8: Heatmap understanding adjusted R2 by strategy and model. [2] Sarah Kreps, R Miles McCain, and Miles Brundage. All the news that’s fit to fabricate: Aigenerated text as a tool of media misinformation. Journal of experimental political science, 9(1):104–117, 2022. [3] Samuel C Woolley. Automating power: Social bot interference in global politics. First Monday, 2016. [4] Franziska B Keller, David Schoch, Seb… view at source ↗

read the original abstract

As large language models (LLMs) increasingly mediate both content generation and moderation, linguistic evasion strategies known as Algospeak have intensified the coevolution between evaders and detectors. This research formalizes the underlying dynamics grounded in a joint action model: when Algospeak increases, detectability and understandability decrease. Further, the concept of Majority Understandable Modulation (MUM) is introduced and defined as the modulation level at which additional evasive alteration increases detector evasion but loses comprehension for the majority of recipients. To empirically probe this trade-off, we introduce a reproducible framework that can be used to create meaning-preserving, Algospeak-style variants, based on an existing taxonomy and with tunable modulation levels. Using COVID-19 disinformation as a first proof-by-example setting, we construct a reference dataset of 700 modulated items, drawn from twenty base sentences across five modulation levels and seven strategies. We then run two linked evaluations with seven different language models: one testing for interpretation through meaning recovery and one for disinformation detection through classification. Curve fitting over modulation levels yields an estimate of the Majority Understandable Modulation threshold and enables sensitivity analyses across strategies and models, see Figure 1. Results reveal the characteristic relationships between understandability and modulation. This study lays the groundwork for understanding the dynamics behind Algospeak and provides the framework, dataset, and experimental setups described.

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 paper defines a MUM threshold for Algospeak trade-offs and ships a taxonomy-based generator plus 700-item dataset, but its curve-fit results rest on LLM proxies with no human or real-detector checks.

read the letter

The main point is that they turn the Algospeak evasion-comprehension tension into a measurable threshold called Majority Understandable Modulation, estimated from curve fits on modulated text. They also give a clean, reproducible way to create those variants using an existing taxonomy at five tunable levels across seven strategies, all grounded in COVID disinformation examples. That framework and the linked dataset of 700 items are the concrete new pieces; prior work on Algospeak has been more descriptive than operational like this. The experiments run meaning-recovery and classification tasks through seven LLMs and show the expected monotonic drops in both quantities as modulation rises, which lines up with their joint-action model. The setup is systematic enough that sensitivity checks across strategies and models are feasible. The soft spot is the proxy assumption. All numbers come from LLM judgments on meaning and detection, with no human-subject validation or side-by-side tests against actual deployed moderators described. That makes the fitted MUM values and the claimed trade-off sensitive to whether LLMs actually track majority human comprehension or real-world detector behavior. The base material is also narrow—only twenty source sentences—so the generalizability of the curves is limited. No error bars or statistical tests are mentioned in the available description either. This is useful for people working on moderation systems or online safety who need a starting generator and a way to quantify the trade-off. A reader who wants to extend the framework or add human data would get direct value from the released pieces. It is coherent on its own terms and has enough structure to deserve peer review, though any referee should press on the proxy gap and ask for at least a small human check or explicit discussion of its limits.

Referee Report

2 major / 2 minor

Summary. The paper claims that Algospeak modulation creates a monotonic trade-off in which higher levels of evasive alteration simultaneously decrease detectability and understandability. It introduces the Majority Understandable Modulation (MUM) threshold, defined as the modulation level at which further evasion improves detector avoidance but causes the majority of recipients to lose core meaning. The claim is investigated via a reproducible generation framework applied to COVID-19 disinformation: a 700-item dataset is built from 20 base sentences using five modulation levels and seven strategies; two LLM-based evaluations (meaning recovery and disinformation classification) are run across seven models; and curve fitting on the resulting performance curves is used to estimate MUM values and conduct sensitivity analyses.

Significance. If the central trade-off and MUM estimates hold under stronger validation, the work supplies a reusable dataset, generation framework, and experimental protocol that could support systematic study of linguistic evasion in LLM-mediated moderation. The multi-model, multi-strategy design permits sensitivity checks, and the explicit reproducibility emphasis is a clear strength. The practical significance remains limited, however, by the exclusive reliance on LLM proxies whose fidelity to human readers and deployed detectors is untested.

major comments (2)

[Abstract and Experimental Setup] Abstract and Experimental Setup: The MUM definition and the claimed trade-off are defined in terms of 'majority of recipients' losing comprehension and real detector evasion, yet both quantities are measured exclusively via LLM meaning-recovery and classification accuracy. No human-subject validation, inter-annotator agreement with humans, or comparison against non-LLM moderation systems is reported; this proxy assumption is load-bearing for the central empirical claims.
[Results and Curve Fitting] Results and Curve Fitting: The abstract states that curve fitting over modulation levels yields MUM estimates, but the manuscript provides no error bars, confidence intervals, goodness-of-fit statistics, or tests for robustness to functional form. Without these, it is impossible to assess whether the reported inflection points are statistically distinguishable from noise or sensitive to modeling choices.

minor comments (2)

The generation framework would be easier to replicate if the main text included one or two concrete examples of base sentences transformed at each modulation level.
[Abstract] The abstract's reference to 'Figure 1' and sensitivity analyses would benefit from a brief textual summary of the key patterns observed across the seven models and seven strategies.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the proxy measures and statistical reporting in our work. We address each major point below, indicating planned revisions where appropriate.

read point-by-point responses

Referee: [Abstract and Experimental Setup] The MUM definition and the claimed trade-off are defined in terms of 'majority of recipients' losing comprehension and real detector evasion, yet both quantities are measured exclusively via LLM meaning-recovery and classification accuracy. No human-subject validation, inter-annotator agreement with humans, or comparison against non-LLM moderation systems is reported; this proxy assumption is load-bearing for the central empirical claims.

Authors: We acknowledge that the MUM threshold and trade-off claims rest on LLM proxies for both meaning recovery and disinformation classification. The study is explicitly framed as a reproducible, LLM-centric proof-of-concept in the setting of LLM-mediated moderation, where such models serve as both generators and potential detectors. We agree that direct human validation would strengthen ecological validity. In the revised manuscript we will add an explicit limitations subsection that discusses the proxy assumption, notes the absence of human inter-annotator agreement or comparisons to non-LLM systems, and outlines future work needed to confirm MUM estimates with human recipients. revision: yes
Referee: [Results and Curve Fitting] The abstract states that curve fitting over modulation levels yields MUM estimates, but the manuscript provides no error bars, confidence intervals, goodness-of-fit statistics, or tests for robustness to functional form. Without these, it is impossible to assess whether the reported inflection points are statistically distinguishable from noise or sensitive to modeling choices.

Authors: We agree that the curve-fitting results require additional statistical detail for proper evaluation. The current version does not report error bars, confidence intervals, or goodness-of-fit metrics. In the revision we will update the results section and Figure 1 to include (i) error bars showing standard error across the seven models, (ii) 95% confidence intervals on the estimated MUM parameters, (iii) R² and residual diagnostics for the fitted curves, and (iv) a brief sensitivity check across alternative functional forms (linear, logistic, and piecewise) to demonstrate robustness of the inflection points. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper introduces a joint action model and defines MUM conceptually as the modulation level where further evasion trades off against majority comprehension. It then generates a 700-item dataset via a tunable framework, measures understandability and detectability through LLM tasks, and estimates the MUM threshold by curve fitting. These steps constitute an empirical observation of the trade-off rather than any reduction of the claimed relationships to fitted parameters or self-referential definitions by construction. No equations, self-citations, or imported uniqueness results are shown to make the central results tautological with the inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The central claim depends on the assumption that systematic modulation preserves meaning at lower levels and that LLM tasks proxy human and detector responses; MUM itself is a newly defined threshold estimated from data.

free parameters (1)

modulation levels
Five discrete tunable levels chosen to generate variants from each base sentence.

axioms (2)

domain assumption Algospeak strategies drawn from an existing taxonomy can be applied at graduated levels while preserving core meaning
This underpins the construction of the 700-item dataset and the claim that modulation trades off evasion against understandability.
domain assumption LLM performance on meaning recovery and disinformation classification tasks reflects human comprehension and detector behavior
Used to run the two linked evaluations and to estimate the MUM threshold via curve fitting.

invented entities (1)

Majority Understandable Modulation (MUM) no independent evidence
purpose: To mark the specific modulation level at which further evasion gains cause majority loss of comprehension
Newly introduced concept whose value is estimated from curve fitting on the experimental data.

pith-pipeline@v0.9.0 · 5548 in / 1566 out tokens · 62439 ms · 2026-05-08T09:59:28.712089+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

[1]

Next-generation phishing: How llm agents empower cyber attackers

Khalifa Afane, Wenqi Wei, Ying Mao, Junaid Farooq, and Juntao Chen. Next-generation phishing: How llm agents empower cyber attackers. In2024 IEEE International Conference on Big Data (BigData), pages 2558–2567, 2024. 11 Figure 7: Heatmap detection adjustedR 2 by strategy and model. Figure 8: Heatmap understanding adjustedR 2 by strategy and model

work page 2024
[2]

All the news that’s fit to fabricate: Ai- generated text as a tool of media misinformation.Journal of experimental political science, 9(1):104–117, 2022

Sarah Kreps, R Miles McCain, and Miles Brundage. All the news that’s fit to fabricate: Ai- generated text as a tool of media misinformation.Journal of experimental political science, 9(1):104–117, 2022

work page 2022
[3]

Automating power: Social bot interference in global politics.First Monday, 2016

Samuel C Woolley. Automating power: Social bot interference in global politics.First Monday, 2016

work page 2016
[4]

Political astroturfing on twitter: How to coordinate a disinformation campaign.Political communication, 37(2):256–280, 2020

Franziska B Keller, David Schoch, Sebastian Stier, and JungHwan Yang. Political astroturfing on twitter: How to coordinate a disinformation campaign.Political communication, 37(2):256–280, 2020. 12 Table 5: Estimated thresholdx 0 (IMUM@0.5, detection) by strategy and model Strategy Claude GPT-4o mini GPT-4o Llama Mistral Qwen xAI Unknown word 5.09988 5.10...

work page 2020
[5]

Steen, K

E. Steen, K. Yurechko, and D. Klug. You can (not) say what you want: Using algospeak to contest and evade algorithmic content moderation on tiktok.Social Media + Society, 9(3), 2023. Original work published 2023

work page 2023
[6]

barbiecore

Sophie Curtis. How tiktok is changing the way we speak: Phrases like “barbiecore”, “quiet quitting” and “le dollar bean” that originated on the social media app have crossed over into the mainstream - so how many do you know?, Sep 2022

work page 2022
[7]

Leg booty? panoramic? seggs? how tiktok is changing language, Nov 2022

Melina Delkic. Leg booty? panoramic? seggs? how tiktok is changing language, Nov 2022

work page 2022
[8]

Tiktok: Wie gartenzwerge die grenzen des sagbaren ver- schieben, Nov 2023

Una Titz and Theresa Lehmann. Tiktok: Wie gartenzwerge die grenzen des sagbaren ver- schieben, Nov 2023

work page 2023
[9]

You can (not) say what you want: Using algospeak to contest and evade algorithmic content moderation on tiktok.Social Media + Society, 9(3):20563051231194586, 2023

Ella Steen, Kathryn Yurechko, and Daniel Klug. You can (not) say what you want: Using algospeak to contest and evade algorithmic content moderation on tiktok.Social Media + Society, 9(3):20563051231194586, 2023

work page 2023
[10]

How algorithm awareness impacts algospeak use on tiktok

Daniel Klug, Ella Steen, and Kathryn Yurechko. How algorithm awareness impacts algospeak use on tiktok. InCompanion Proceedings of the ACM Web Conference 2023, WWW ’23 Companion, page 234–237, New York, NY , USA, 2023. Association for Computing Machinery

work page 2023
[11]

Simple llm based approach to counter algospeak

Jan Fillies and Adrian Paschke. Simple llm based approach to counter algospeak. InProceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024), pages 136–145, 2024

work page 2024
[12]

I am borrowing ya mixing ?

Kalika Bali, Jatin Sharma, Monojit Choudhury, and Yogarshi Vyas. “I am borrowing ya mixing ?” an analysis of English-Hindi code mixing in Facebook. In Mona Diab, Julia Hirschberg, Pascale Fung, and Thamar Solorio, editors,Proceedings of the First Workshop on Computational Approaches to Code Switching, pages 116–126, Doha, Qatar, October 2014. Association ...

work page 2014
[13]

Code mixing: A challenge for language identification in the language of social media

Utsab Barman, Amitava Das, Joachim Wagner, and Jennifer Foster. Code mixing: A challenge for language identification in the language of social media. In Mona T. Diab, Julia Hirschberg, Pascale Fung, and Thamar Solorio, editors,Proceedings of the First Workshop on Computational Approaches to Code Switching@EMNLP 2014, Doha, Qatar , October 25, 2014, pages ...

work page 2014
[14]

Detecting offensive tweets in Hindi-English code-switched language

Puneet Mathur, Rajiv Shah, Ramit Sawhney, and Debanjan Mahata. Detecting offensive tweets in Hindi-English code-switched language. In Lun-Wei Ku and Cheng-Te Li, editors,Proceedings of the Sixth International Workshop on Natural Language Processing for Social Media, pages 18–26, Melbourne, Australia, July 2018. Association for Computational Linguistics. 13

work page 2018
[15]

A dataset of Hindi-English code-mixed social media text for hate speech detection

Aditya Bohra, Deepanshu Vijay, Vinay Singh, Syed Sarfaraz Akhtar, and Manish Shrivastava. A dataset of Hindi-English code-mixed social media text for hate speech detection. In Malvina Nissim, Viviana Patti, Barbara Plank, and Claudia Wagner, editors,Proceedings of the Second Workshop on Computational Modeling of People’s Opinions, Personality, and Emotion...

work page 2018
[16]

Mixed-code text analysis for the detection of online hidden propaganda

Andrea Tundis, Gaurav Mukherjee, and Max Mühlhäuser. Mixed-code text analysis for the detection of online hidden propaganda. InProceedings of the 15th International Conference on Availability, Reliability and Security, ARES ’20, New York, NY , USA, 2020. Association for Computing Machinery

work page 2020
[17]

Deobfuscating leetspeak with deep learning to improve spam filtering

Iñaki Vélez de Mendizabal, Xabier Vidriales Mazorriaga, Iñigo Ezpeleta, and Urko Zurutuza. Deobfuscating leetspeak with deep learning to improve spam filtering. 2023

work page 2023
[18]

Bae: Bert-based adversarial examples for text classification

Siddhant Garg and Goutham Ramakrishnan. Bae: Bert-based adversarial examples for text classification. InProceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), pages 6174–6181, 2020

work page 2020
[19]

Humanizing machine-generated content: evading ai-text detection through adversarial attack

Ying Zhou, Ben He, and Le Sun. Humanizing machine-generated content: evading ai-text detection through adversarial attack. InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 8427–8437, 2024

work page 2024
[20]

Brennan (1991) grounding in communication

Herbert H Clark. Brennan (1991) grounding in communication. 1991

work page 1991
[21]

Cambridge university press, 1996

Herbert H Clark.Using language. Cambridge university press, 1996

work page 1996
[22]

Referring as a collaborative process.Cognition, 22(1):1–39, 1986

Herbert H Clark and Deanna Wilkes-Gibbs. Referring as a collaborative process.Cognition, 22(1):1–39, 1986

work page 1986
[23]

How readability shapes social media engagement.Journal of consumer psychology, 29(2):262–270, 2019

Ethan Pancer, Vincent Chandler, Maxwell Poole, and Theodore J Noseworthy. How readability shapes social media engagement.Journal of consumer psychology, 29(2):262–270, 2019

work page 2019
[24]

Limitations

Felix A Wichmann and N Jeremy Hill. The psychometric function: I. fitting, sampling, and goodness of fit.Perception & psychophysics, 63(8):1293–1313, 2001. 14 NeurIPS Paper Checklist 1.Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The abstract ...

work page 2001
[25]

All experiments used LLMs only

Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

work page

[1] [1]

Next-generation phishing: How llm agents empower cyber attackers

Khalifa Afane, Wenqi Wei, Ying Mao, Junaid Farooq, and Juntao Chen. Next-generation phishing: How llm agents empower cyber attackers. In2024 IEEE International Conference on Big Data (BigData), pages 2558–2567, 2024. 11 Figure 7: Heatmap detection adjustedR 2 by strategy and model. Figure 8: Heatmap understanding adjustedR 2 by strategy and model

work page 2024

[2] [2]

All the news that’s fit to fabricate: Ai- generated text as a tool of media misinformation.Journal of experimental political science, 9(1):104–117, 2022

Sarah Kreps, R Miles McCain, and Miles Brundage. All the news that’s fit to fabricate: Ai- generated text as a tool of media misinformation.Journal of experimental political science, 9(1):104–117, 2022

work page 2022

[3] [3]

Automating power: Social bot interference in global politics.First Monday, 2016

Samuel C Woolley. Automating power: Social bot interference in global politics.First Monday, 2016

work page 2016

[4] [4]

Political astroturfing on twitter: How to coordinate a disinformation campaign.Political communication, 37(2):256–280, 2020

Franziska B Keller, David Schoch, Sebastian Stier, and JungHwan Yang. Political astroturfing on twitter: How to coordinate a disinformation campaign.Political communication, 37(2):256–280, 2020. 12 Table 5: Estimated thresholdx 0 (IMUM@0.5, detection) by strategy and model Strategy Claude GPT-4o mini GPT-4o Llama Mistral Qwen xAI Unknown word 5.09988 5.10...

work page 2020

[5] [5]

Steen, K

E. Steen, K. Yurechko, and D. Klug. You can (not) say what you want: Using algospeak to contest and evade algorithmic content moderation on tiktok.Social Media + Society, 9(3), 2023. Original work published 2023

work page 2023

[6] [6]

barbiecore

Sophie Curtis. How tiktok is changing the way we speak: Phrases like “barbiecore”, “quiet quitting” and “le dollar bean” that originated on the social media app have crossed over into the mainstream - so how many do you know?, Sep 2022

work page 2022

[7] [7]

Leg booty? panoramic? seggs? how tiktok is changing language, Nov 2022

Melina Delkic. Leg booty? panoramic? seggs? how tiktok is changing language, Nov 2022

work page 2022

[8] [8]

Tiktok: Wie gartenzwerge die grenzen des sagbaren ver- schieben, Nov 2023

Una Titz and Theresa Lehmann. Tiktok: Wie gartenzwerge die grenzen des sagbaren ver- schieben, Nov 2023

work page 2023

[9] [9]

You can (not) say what you want: Using algospeak to contest and evade algorithmic content moderation on tiktok.Social Media + Society, 9(3):20563051231194586, 2023

Ella Steen, Kathryn Yurechko, and Daniel Klug. You can (not) say what you want: Using algospeak to contest and evade algorithmic content moderation on tiktok.Social Media + Society, 9(3):20563051231194586, 2023

work page 2023

[10] [10]

How algorithm awareness impacts algospeak use on tiktok

Daniel Klug, Ella Steen, and Kathryn Yurechko. How algorithm awareness impacts algospeak use on tiktok. InCompanion Proceedings of the ACM Web Conference 2023, WWW ’23 Companion, page 234–237, New York, NY , USA, 2023. Association for Computing Machinery

work page 2023

[11] [11]

Simple llm based approach to counter algospeak

Jan Fillies and Adrian Paschke. Simple llm based approach to counter algospeak. InProceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024), pages 136–145, 2024

work page 2024

[12] [12]

I am borrowing ya mixing ?

Kalika Bali, Jatin Sharma, Monojit Choudhury, and Yogarshi Vyas. “I am borrowing ya mixing ?” an analysis of English-Hindi code mixing in Facebook. In Mona Diab, Julia Hirschberg, Pascale Fung, and Thamar Solorio, editors,Proceedings of the First Workshop on Computational Approaches to Code Switching, pages 116–126, Doha, Qatar, October 2014. Association ...

work page 2014

[13] [13]

Code mixing: A challenge for language identification in the language of social media

Utsab Barman, Amitava Das, Joachim Wagner, and Jennifer Foster. Code mixing: A challenge for language identification in the language of social media. In Mona T. Diab, Julia Hirschberg, Pascale Fung, and Thamar Solorio, editors,Proceedings of the First Workshop on Computational Approaches to Code Switching@EMNLP 2014, Doha, Qatar , October 25, 2014, pages ...

work page 2014

[14] [14]

Detecting offensive tweets in Hindi-English code-switched language

Puneet Mathur, Rajiv Shah, Ramit Sawhney, and Debanjan Mahata. Detecting offensive tweets in Hindi-English code-switched language. In Lun-Wei Ku and Cheng-Te Li, editors,Proceedings of the Sixth International Workshop on Natural Language Processing for Social Media, pages 18–26, Melbourne, Australia, July 2018. Association for Computational Linguistics. 13

work page 2018

[15] [15]

A dataset of Hindi-English code-mixed social media text for hate speech detection

Aditya Bohra, Deepanshu Vijay, Vinay Singh, Syed Sarfaraz Akhtar, and Manish Shrivastava. A dataset of Hindi-English code-mixed social media text for hate speech detection. In Malvina Nissim, Viviana Patti, Barbara Plank, and Claudia Wagner, editors,Proceedings of the Second Workshop on Computational Modeling of People’s Opinions, Personality, and Emotion...

work page 2018

[16] [16]

Mixed-code text analysis for the detection of online hidden propaganda

Andrea Tundis, Gaurav Mukherjee, and Max Mühlhäuser. Mixed-code text analysis for the detection of online hidden propaganda. InProceedings of the 15th International Conference on Availability, Reliability and Security, ARES ’20, New York, NY , USA, 2020. Association for Computing Machinery

work page 2020

[17] [17]

Deobfuscating leetspeak with deep learning to improve spam filtering

Iñaki Vélez de Mendizabal, Xabier Vidriales Mazorriaga, Iñigo Ezpeleta, and Urko Zurutuza. Deobfuscating leetspeak with deep learning to improve spam filtering. 2023

work page 2023

[18] [18]

Bae: Bert-based adversarial examples for text classification

Siddhant Garg and Goutham Ramakrishnan. Bae: Bert-based adversarial examples for text classification. InProceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), pages 6174–6181, 2020

work page 2020

[19] [19]

Humanizing machine-generated content: evading ai-text detection through adversarial attack

Ying Zhou, Ben He, and Le Sun. Humanizing machine-generated content: evading ai-text detection through adversarial attack. InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 8427–8437, 2024

work page 2024

[20] [20]

Brennan (1991) grounding in communication

Herbert H Clark. Brennan (1991) grounding in communication. 1991

work page 1991

[21] [21]

Cambridge university press, 1996

Herbert H Clark.Using language. Cambridge university press, 1996

work page 1996

[22] [22]

Referring as a collaborative process.Cognition, 22(1):1–39, 1986

Herbert H Clark and Deanna Wilkes-Gibbs. Referring as a collaborative process.Cognition, 22(1):1–39, 1986

work page 1986

[23] [23]

How readability shapes social media engagement.Journal of consumer psychology, 29(2):262–270, 2019

Ethan Pancer, Vincent Chandler, Maxwell Poole, and Theodore J Noseworthy. How readability shapes social media engagement.Journal of consumer psychology, 29(2):262–270, 2019

work page 2019

[24] [24]

Limitations

Felix A Wichmann and N Jeremy Hill. The psychometric function: I. fitting, sampling, and goodness of fit.Perception & psychophysics, 63(8):1293–1313, 2001. 14 NeurIPS Paper Checklist 1.Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The abstract ...

work page 2001

[25] [25]

All experiments used LLMs only

Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

work page