Pith. sign in

REVIEW 4 major objections 4 minor 41 references

Analysis of LLM as a grammatical feature tagger for African American English

T0 review · 4 major / 4 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 Rule-based taggers still beat LLMs at spotting AAE grammar

desk verdict Useful comparison: baselines do beat LLMs on AAE feature tagging, but the recency-bias analysis is endogenous and should not be reported as a finding without a proper experimental design. read the letter →

arxiv 2502.06004 v1 pith:NMCUTUWV submitted 2025-02-09 cs.CL cs.AIcs.LG

Rahul Porwal , Alice Rozet , Pryce Houck , Jotsna Gowda , Sarah Moeller , Kevin Tang This is my paper · ORCID
classification cs.CLcs.AIcs.LG
keywords AfricanAmericanEnglishgrammaticalfeaturetaggingHabitualBeMultipleNegationlargelanguagemodelsrecencybiasformalitylow-resourceNLP
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 off-the-shelf large language models can replace specialized tools for annotating two grammatical features of African American English: Habitual Be and Multiple Negation. On sentence-level binary classification, the answer is no: a rule-based tagger and a transformer trained for the task both outscore zero- and few-shot GPT and LLaMA models. The paper also isolates two systematic biases in the LLM outputs. Sentence order in a batch changes predictions (recency bias), and sentences whose informal, transcribed style deviates from standard written English are more likely to be called AAE features (formality bias). If these results hold, LLM-based annotation of AAE grammar is currently less reliable than existing small tools, and any pipeline that uses LLMs for such tagging should control for ordering and stylistic confounds.

What carries the argument

The machinery is a side-by-side evaluation plus a regression-based error analysis. The task is sentence-level binary classification: for each sentence, predict presence or absence of Habitual Be or Multiple Negation. Performance is measured by class-size-weighted F1 from a classification report. The Habitual Be baseline is a transformer trained on 10-fold cross-validated, augmented data with n-gram and part-of-speech windows around the verb 'be'; the Multiple Negation baseline is a rule-based tagger that counts negators within clause boundaries. To explain LLM errors, the paper fits logistic regressions with the model prediction as the dependent variable, ground truth as a control, and two hypothesized bias terms: recency (the fraction of the last five predictions that match the current prediction) and formality (a manual binary tag for deviation from formal written English). The regression coefficients are what carry the bias claims.

What would settle it

Repeat the same zero- and few-shot prompts several times on the same sentences and compute the distribution of weighted F1 for each model; if the confidence intervals around the LLM scores overlap the baseline scores, the claimed underperformance would not be stable. Separately, reorder a fixed batch so that the previous five predictions are inverted, and check whether the logistic-regression recency coefficient changes sign or loses significance.

Watch

Extended reading notes

Core claim

The paper's central claim is that rule-based and transformer-based models of AAE grammatical features outperform zero- and few-shot LLMs, and that LLM recognition of AAE is influenced by recency and by unrelated features of non-standard written conventions. The Habitual Be baseline, a transformer using n-gram windows and syntactic context around the verb 'be', reaches roughly 0.92 weighted F1 with 0.88 recall, while the Multiple Negation baseline, a small rule-based tagger that looks for two negators inside the same clause, reaches essentially perfect recall and 0.99 weighted F1. On the same tasks the LLMs land lower, with the gap in weighted F1 on the order of 0.16 to 0.18, and logistic regressions show that the proportion of matching predictions in the previous five sentences has a significant effect on the current prediction in most LLM configurations. Under few-shot prompting, a manually tagged informality variable (excessive punctuation, missing subjects, run-on sentences) significantly raises the probability of predicting an AAE feature. The paper releases the Multiple Negation tagger as a practical artifact.

Load-bearing premise

The strongest assumption is that one forward pass per sentence at temperature 0.7 gives a stable enough estimate of each LLM's accuracy that the reported F1 gaps are real differences rather than sampling noise.

Editorial extensions

If this is right

  • For annotating AAE grammar today, rule-based and trained transformer taggers should be preferred over zero- or few-shot LLMs unless the LLM is fine-tuned or otherwise corrected.
  • Few-shot prompting is not uniformly better than zero-shot: it helps for Habitual Be but hurts for Multiple Negation, so prompt selection should depend on the feature's structural complexity.
  • Because sentence order influences LLM predictions, batching matters; randomizing or interleaving sentences could reduce recency-driven misclassification.
  • Because informal or transcribed style triggers false positives, preprocessing that repairs punctuation and segments run-on sentences may improve LLM accuracy on non-standard written or spoken text.
  • The release of a rule-based Multiple Negation tagger gives downstream researchers a transparent, strong baseline to beat.

Reading between the lines

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

  • If the formality-bias result generalizes beyond oral histories, LLM feature taggers for AAE on social-media text would over-predict AAE features whenever the text is nonstandard in any way, including typos or stylistic informality unrelated to AAE.
  • A testable extension would be to run the same prompts with repeated sampling at the same temperature and report variance; the current F1 gaps could shrink or reorder once sampling noise is measured.
  • The recency-bias coefficients being negative (the model avoids repeating its recent prediction) rather than positive invites a targeted experiment: craft batches where the previous five labels are all the opposite of the true label, and measure how much the prediction flips.
  • Because the strongest comparisons pit a purpose-trained transformer against general-purpose LLMs, a natural next test is whether instruction-tuned or fine-tuned LLMs with AAE data close the gap; the paper does not claim they would not.
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 / 4 minor

Summary. The paper evaluates whether general-purpose LLMs (gpt-4o-mini and LLaMA 3-8B-Instruct, both at temperature 0.7) can tag two AAE grammatical features — Habitual Be and Multiple Negation — at the sentence level, comparing zero- and few-shot prompting against a rule-based Multiple Negation tagger and a Transformer-based Habitual Be tagger. The headline finding is that the specialized baselines outperform the LLMs on weighted F1 (e.g., Habitual Be F1w drops by about 0.18 from the baseline to the best LLM setting; Multiple Negation drops by about 0.16). The paper then presents logistic regressions intended to show that LLM predictions are influenced by recency of prior predictions and by formality-related deviations from written standard text. Data and code are released. The central benchmark is not circular — the baselines are external and independently evaluated — but the recency analysis is, as detailed below, because its key regressor is a deterministic function of the model's own predictions.

Significance. If the baseline-comparison result is robust, the paper provides a useful empirical benchmark: for the two AAE features studied, specialized and interpretable taggers still beat general-purpose zero- and few-shot LLMs. That result is relevant to the design of annotation pipelines for low-resource language varieties and is reported with a clear task setup and released data/code. The paper's second contribution — that LLM AAE recognition is influenced by recency and by unrelated non-standard conventions — is currently not supported by the presented analysis because the recency regressor is endogeneous and the formality analysis rests on a single fold with a subjective annotation threshold. The evaluation also relies on single stochastic forward passes, so the magnitudes of F1 gaps should be treated with caution.

major comments (4)
  1. [Table 1 / §7.2.2] The Multiple Negation dataset description is internally inconsistent. Table 1 reports 37 positive and 161 negative examples (total 198), but Section 7.2.2 says the dataset contains 198 samples, 'out of which 162 contain Multiple Negation and the remaining 36 do not.' These two descriptions cannot both be correct. This matters because the baseline's near-perfect performance (F1=0.99) and the LLMs' precision/recall trade-offs are interpreted in light of the class balance; the reader cannot tell whether the positive class has 18.7% or 81.8% of the data. Please correct the numbers and state which dataset was actually used in Tables 2–5.
  2. [§8.1.1, Table 3, Limitations] The recency-bias regression is endogenous. The independent variable is defined as the proportion of the last five predicted labels that match the current predicted label, and the dependent variable is that same current prediction. Any model whose outputs are autocorrelated — even one that ignores neighboring sentences — will produce a nonzero recency coefficient, so the significant negative coefficients in Tables 3 and 4 do not establish that sentence order causally affects predictions. The ground-truth control does not remove this mechanical link. The Limitations paragraph concedes that 'we cannot say for certain if it is in fact due to recency bias or due to the LLM correctly predicting the feature,' yet the Conclusion (Section 9) nonetheless states as a contribution that 'LLMs recognition of AAE is influenced by recency.' This conclusion is unsupported by the presented analysis; the recency claims in Section 8.1.2 and 8.1.3 should be substantially revised or the analysis redesigned (e.g., with a held-out input-order manipulation).
  3. [§5.1 and Table 2] All LLM results come from a single forward pass at temperature 0.7, with no repeated sampling and no variance reporting. Because the models are stochastic at this temperature, the reported F1 differences (e.g., GPT Habitual Be zero-shot F1w 0.41 vs. few-shot 0.53, or GPT vs. LLaMA Multiple Negation zero-shot F1w 0.95 vs. 0.76) may be within sampling noise. At minimum, give the number of runs, report means and standard errors or confidence intervals for the weighted F1, and state a decision rule for when a difference is treated as meaningful. This is a load-bearing issue because the paper's first contribution is a quantitative ranking (baselines > LLMs) based on these single-run numbers.
  4. [§8.2.1–8.2.2 and Limitations] The formality-bias analysis is conducted on only one fold (fold 2) of the Habitual Be dataset, chosen for being 'median' in the zero-shot GPT results, and the formality labels were assigned manually with a threshold ('excessive or unusual punctuation') that is not operationalized or reliability-checked. The claim in §8.2.2 that few-shot prompting 'suffers from a formality bias with both features' therefore rests on a single fold and a subjective annotation. The Limitations section also notes that the augmented Habitual Be data lacks natural sentence structure, which further compromises the formality annotation for that feature. Please report the analysis on all folds (or justify the single-fold design) and provide inter-annotator agreement or a codebook for the formality variable.
minor comments (4)
  1. [Abstract / §1] There are several typos and ungrammatical constructions: 'performs well' should be 'perform well' in the Introduction; 'Habitual Be and Multiple Negation is found' should be 'are found'; 'exmaple' in §3.1; 'LLaMA model is unable to detects Habitual Be poorly' in §7.2.1 is garbled; and 're- cency' in Table 4 splits the word across lines. A careful proofread is needed.
  2. [Table 4] The LLaMA few-shot ground-truth p-value for Multiple Negation is reported as 1.000 with a coefficient of 30.81, which is implausible on its face and suggests a numerical or reporting error. Please verify the regression outputs and report confidence intervals for the coefficients.
  3. [§5.2 / Table 1] The sentence 'Approximately 3,500 more Habitual Be sentences with the method described in Santiago et al. (2022) to provide an evenly balanced the dataset' is missing a verb ('... were augmented with the method ...'). Also, the table caption says 'the datasets' but the column structure is confusing: please make clear which rows correspond to positive and negative examples for each feature, and reconcile the counts with the text.
  4. [§8.1.3] The LLaMA few-shot recency analysis is described with a small set of ten prompts (five true 0s and five true 1s) that were previously failed, then reordered and repeated six times; this is a targeted robustness probe rather than a regression on the full dataset, and its presentation as 'additional testing' should be labeled as exploratory, with the sample size stated in the main text rather than implied.

Circularity Check

1 steps flagged · score 6.0 of 10

The benchmark comparison is externally valid, but the recency-bias conclusion is built on an endogenous regression whose regressor is a function of the outcome, so that central claim reduces by construction.

  1. self definitional [Section 8.1.1 (Methodology) and the Limitations paragraph]
    "The dependent variable is the model prediction (presence or absent of a feature). The independent variable of interest is the presence of a recency bias. It is operationalised as the proportion of the predicted values of the last five samples (N−5,...N−1) that match the predicted value of the given sample (N). ..."

    The recency regressor is computed from the model's own predictions, which are exactly the dependent variable of the logistic regression. For any classifier with class imbalance or temporally clustered correct labels, the proportion of the last five predictions that match the current prediction is mechanically correlated with the current prediction even when sentence order has no causal effect. The ground-truth control does not remove this endogeneity because the lagged predictions are themselves functions of the ground truth. Therefore the significant recency coefficients in Tables 3 and 4 do not establish an ordering bias; the paper's own Limitations paragraph concedes this ambiguity, yet the Conclusion still lists 'influenced by recency' as a main contribution.

full rationale

The paper's primary benchmark comparison is not circular: the rule-based tagger, the Transformer baseline, and the LLM predictions are evaluated against human gold labels, and the baselines are independently specified systems. The formality-bias analysis is also not circular, since its predictor is a manually assigned property of the sentence rather than a function of the model output. The circularity is concentrated in the recency-bias regression of Section 8.1. There, the independent variable is defined as the proportion of the last five model predictions that match the current prediction, while the dependent variable is that same current prediction. This self-referential construction guarantees a mechanical association under any autocorrelated or class-imbalanced prediction stream, so the significant coefficients cannot be interpreted as evidence of order-based recency bias. The paper's own Limitations text acknowledges the confound but the Conclusion nevertheless asserts recency as a finding. Because this is one of the paper's two headline contributions, the overall circularity score is moderate: the benchmark result stands, but the recency conclusion partially reduces to the way its regressor was constructed.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities. Its claims rest on standard NLP assumptions (gold annotations, valid baselines) and on two under-specified modeling choices: the recency window size and the few-shot example set. The subjectivity of the formality tags is an additional assumption that is not independently validated.

free parameters (3)
  • Recency window size = 5
    The regression uses the last 5 predictions to define the recency variable (Section 8.1.1); no justification for this window length is given.
  • Few-shot example set
    The number and selection of training examples in few-shot prompts is not specified (Section 6), so the comparison is under-determined.
  • Formality annotation threshold
    Manual binary tags for excessive punctuation, run-ons, or missing subjects (Section 8.2.1) are subjective; no inter-annotator agreement is reported.
assumptions (4)
  • domain assumption Human annotations are gold standard
    Section 4 says 'The labels assigned by human annotators are treated as gold standard'; all model scores are computed against these labels, so annotation errors propagate.
  • domain assumption Baseline model validity
    The rule-based Multiple Negation tagger and the transformer baseline from Previlon et al. (2024) are accepted as correct references without independent verification in this paper.
  • domain assumption Negative examples contain a key component
    Section 4 defines negative examples as sentences containing a key component ('be' or a negator) but not the feature, which limits the task and shapes baseline and LLM behavior.
  • ad hoc to paper Augmented Habitual Be data is natural enough
    Section 5.2 describes adding about 3,500 augmented sentences to balance the Habitual Be dataset; the authors later note in Limitations that the augmented data made formality assessment harder, so this assumption is load-bearing for the formality analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Analysis of LLM as a grammatical feature tagger for African American English." pith.science (2026). https://pith.science/paper/NMCUTUWV

@misc{pith2026250206004,
  author       = {Pith},
  title        = {Pith review of: Analysis of LLM as a grammatical feature tagger for African American English},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NMCUTUWV}},
  note         = {Machine review of arXiv:2502.06004}
}
read the original abstract

African American English (AAE) presents unique challenges in natural language processing (NLP). This research systematically compares the performance of available NLP models--rule-based, transformer-based, and large language models (LLMs)--capable of identifying key grammatical features of AAE, namely Habitual Be and Multiple Negation. These features were selected for their distinct grammatical complexity and frequency of occurrence. The evaluation involved sentence-level binary classification tasks, using both zero-shot and few-shot strategies. The analysis reveals that while LLMs show promise compared to the baseline, they are influenced by biases such as recency and unrelated features in the text such as formality. This study highlights the necessity for improved model training and architectural adjustments to better accommodate AAE's unique linguistic characteristics. Data and code are available.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 21 canonical work pages

  1. [1]

    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...

  2. [2]

    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...

  3. [3]

    Soumya Barikeri, Anne Lauscher, Ivan Vulic, and Goran Glavas. 2021. https://doi.org/10.18653/v1/2021.acl-long.151 Redditbias: A real-world resource for bias evaluation and debiasing of conversational language models . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on N...

    doi:10.18653/v1/2021.acl-long.151 2021

  4. [4]

    O'Connor

    Su Lin Blodgett, Lisa Green, and Brendan T. O'Connor. 2016. https://doi.org/10.18653/v1/d16-1120 Demographic dialectal variation in social media: A case study of african-american english . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016 , pages 1119--1130. The As...

    doi:10.18653/v1/d16-1120 2016

  5. [5]

    Su Lin Blodgett, Johnny Wei, and Brendan O ' Connor. 2018. https://doi.org/10.18653/v1/P18-1131 T witter U niversal D ependency parsing for A frican- A merican and mainstream A merican E nglish . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1415--1425, Melbourne, Australia. Assoc...

    doi:10.18653/v1/p18-1131 2018

  6. [6]

    Lu Cheng, Nayoung Kim, and Huan Liu. 2022. https://aclanthology.org/2022.coling-1.110 Debiasing word embeddings with nonlinear geometry . In Proceedings of the 29th International Conference on Computational Linguistics, COLING 2022, Gyeongju, Republic of Korea, October 12-17, 2022 , pages 1286--1298. International Committee on Computational Linguistics

    work page 2022

  7. [7]

    Christian Clark, Byung-Doh Oh, and William Schuler. 2025. https://aclanthology.org/2025.coling-main.517/ Linear recency bias during training improves transformers' fit to reading times . In Proceedings of the 31st International Conference on Computational Linguistics, pages 7735--7747, Abu Dhabi, UAE. Association for Computational Linguistics

    2025

  8. [8]

    Jamell Dacon. 2022. https://doi.org/10.48550/arXiv.2206.08978 Towards a deep multi-layered dialectal language analysis: A case study of A frican- A merican E nglish . Proceedings of the S econd W orkshop on B ridging H uman-- C omputer I nteraction and N atural L anguage P rocessing

    work page Pith review arXiv doi:10.48550/arxiv.2206.08978 2022

Show all 41 references
  1. [9]

    Ralph W. Fasold. 1969. https://doi.org/10.2307/412334 Tense and the Form Be in Black English . Language, 45(4):763--776. Publisher: Linguistic Society of America

    1969 doi

  2. [10]

    Aparna Garimella, Rada Mihalcea, and Akhash Amarnath. 2022. https://aclanthology.org/2022.aacl-short.38 Demographic-aware language model fine-tuning as a bias mitigation technique . In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computa...

    2022

  3. [11]

    Lisa J. Green. 2002. https://doi.org/10.1017/CBO9780511800306 African American English : A Linguistic Introduction . Cambridge University Press

    2002 doi

  4. [12]

    Xiaobo Guo and Soroush Vosoughi. 2024. https://arxiv.org/abs/2406.15981 Serial position effects of large language models . Preprint, arXiv:2406.15981

    2024 arXiv

  5. [13]

    Camille Harris, Matan Halevy, Ayanna Howard, Amy Bruckman, and Diyi Yang. 2022. https://doi.org/10.1145/3531146.3533144 Exploring the role of grammar and word choice in bias toward African American English (AAE) in hate speech classification . In 2022 ACM Conference on Fairnes...

    2022

  6. [14]

    Frey, and Kathleen McKeown

    Alyssa Hwang, William R. Frey, and Kathleen McKeown. 2020. https://aclanthology.org/2020.vardial-1.15/ Towards augmenting lexical resources for slang and A frican A merican E nglish . In Proceedings of the 7th Workshop on NLP for Similar Languages, Varieties and Dialects, page...

    2020

  7. [15]

    Suramya Jadhav, Abhay Shanbhag, Amogh Thakurdesai, Ridhima Sinare, and Raviraj Joshi. 2024. https://arxiv.org/abs/2411.17637 On limitations of llm as annotator for low resource languages . Preprint, arXiv:2411.17637

    2024 arXiv

  8. [16]

    Anna J rgensen, Dirk Hovy, and Anders S gaard. 2016. https://doi.org/10.18653/v1/n16-1130 Learning a POS tagger for aave-like language . In NAACL HLT 2016, The 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Techno...

    2016 doi

  9. [17]

    Amel Kallel. 2011. https://www.cambridgescholars.com/product/978-1-4438-2738-6 The Loss of Negative Concord in Standard E nglish: A Case of Lexical Reanalysis . Cambridge Scholars Publishing, Newcastle upon Tyne, UK

    2011

  10. [18]

    Tyler Kendall and Charlie Farrington. 2021. https://doi.org/10.7264/1AD5-6T35 The Corpus of Regional African American Language . Publisher: Online Resources for African American Language Project

    2021 doi

  11. [19]

    Mohammad

    Svetlana Kiritchenko and Saif M. Mohammad. 2018. https://doi.org/10.18653/v1/s18-2005 Examining gender and race bias in two hundred sentiment analysis systems . In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, *SEM@NAACL-HLT 2018, New Orle...

    2018 doi

  12. [20]

    Allison Koenecke, Andrew Nam, Emily Lake, Joe Nudell, Minnie Quartey, Zion Mengesha, Connor Toups, John R Rickford, Dan Jurafsky, and Sharad Goel. 2020. https://doi.org/10.1073/pnas.1915768117 Racial disparities in automated speech recognition . Proceedings of the National Aca...

    2020 doi

  13. [21]

    Bernd Kortmann. 2020. https://doi.org/10.1002/9781119540618.ch16 Syntactic Variation in E nglish , chapter 16. John Wiley & Sons, Ltd

    2020 doi

  14. [22]

    Ernesto William De Luca and Oliver Streiter. 2003. https://api.semanticscholar.org/CorpusID:169404662 Example-based NLP for minority languages: Tasks, resources and tools . In Proceedings of the Workshop on Natural Language Processing of Minority Languages with Few Computation...

    2003

  15. [23]

    Antonis Maronikolakis, Philip Baader, and Hinrich Sch \"u tze. 2022. https://doi.org/10.18653/v1/2022.gebnlp-1.1 Analyzing hate speech data along racial, gender and intersectional axes . In Proceedings of the 4th Workshop on Gender Bias in Natural Language Processing (GeBNLP),...

    2022 doi

  16. [24]

    Joshua L. Martin. 2022. https://www.proquest.com/dissertations-theses/automatic-speech-recognition-systems-spoken/docview/2676147443/se-2 Automatic Speech Recognition Systems, Spoken Corpora, and African American Language: An Examination of Linguistic Bias and Morphosyntactic ...

    2022

  17. [25]

    Martin and Kevin Tang

    Joshua L. Martin and Kevin Tang. 2020. https://doi.org/10.21437/Interspeech.2020-2893 Understanding Racial Disparities in Automatic Speech Recognition: The Case of Habitual “be” . In Proc. Interspeech 2020, pages 626--630

    2020 doi

  18. [26]

    Palacios Martínez

    Ignacio M. Palacios Martínez. 2003. http://www.jstor.org/stable/43343991 Multiple negation in modern english. a preliminary corpus-based study . Neuphilologische Mitteilungen, 104(4):477--498

    2003

  19. [27]

    Tessa Masis, Chloe Eggleston, Lisa J Green, Taylor Jones, Meghan Armstrong, and Brendan O'Connor. 2023. https://doi.org/10.7275/ZDG0-0914 Investigating morphosyntactic variation in A frican A merican E nglish on T witter . Proceedings of the Society for Computation in Linguist...

    2023 doi

  20. [28]

    Tessa Masis, Anissa Neal, Lisa Green, and Brendan O ' Connor. 2022. https://aclanthology.org/2022.fieldmatters-1.2/ Corpus-guided contrast sets for morphosyntactic feature detection in low-resource E nglish varieties . In Proceedings of the first workshop on NLP applications t...

    2022

  21. [29]

    Sarah Moeller, Alexis Davis, Wilermine Previlon, Michael Bottini, and Kevin Tang. 2024. https://doi.org/10.32714/ricl.12.02.04 Compiling a corpus of A frican A merican language from oral histories . Research in Corpus Linguistics, 12(2):45--79

    2024 doi

  22. [30]

    Sarah Moeller, Alexis Ciara Davis, Wilermine Previlon, and Kevin Tang. 2025. ALFAH -- A Toolkit for Annotating Linguistic Features in African American Oral Histories . https://osf.io/x9whn/. Accessed: 2025-02-08

    2025

  23. [31]

    Maja Pavlovic and Massimo Poesio. 2024. https://aclanthology.org/2024.nlperspectives-1.11/ The effectiveness of LLM s as annotators: A comparative overview and empirical analysis of direct representation . In Proceedings of the 3rd Workshop on Perspectivist Approaches to NLP (...

    2024

  24. [32]

    Fabian Pedregosa, Ga \"e l Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. 2011. Scikit-learn: Machine learning in python. Journal of machine learning research, 12(Oct):2825--2830

    2011

  25. [33]

    Alexander Peysakhovich and Adam Lerer. 2023. https://api.semanticscholar.org/CorpusID:263609111 Attention sorting combats recency bias in long context language models . ArXiv, abs/2310.01427

    2023 arXiv

  26. [34]

    Wilermine Previlon, Alice Rozet, Jotsna Gowda, Bill Dyer, Kevin Tang, and Sarah Moeller. 2024. https://aclanthology.org/2024.lrec-main.909/ Leveraging syntactic dependencies in disambiguation: The case of A frican A merican E nglish . In Proceedings of the 2024 Joint Internati...

    2024

  27. [35]

    Harrison Santiago, Joshua Martin, Sarah Moeller, and Kevin Tang. 2022. https://doi.org/10.18653/v1/2022.ltedi-1.9 Disambiguation of morpho-syntactic features of African American English – the case of habitual be . In Proceedings of the Second Workshop on Language Technology fo...

    2022 doi

  28. [36]

    Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. https://doi.org/10.18653/v1/p19-1163 The risk of racial bias in hate speech detection . In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, It...

    2019 doi

  29. [37]

    Jonathan Shaki, Sarit Kraus, and Michael Wooldridge. 2023. https://doi.org/10.3233/faia230505 Cognitive Effects in Large Language Models . IOS Press

    2023 doi

  30. [38]

    Andrew Silva, Pradyumna Tambwekar, and Matthew Gombolay. 2021. https://doi.org/10.18653/v1/2021.naacl-main.189 Towards a comprehensive understanding and accurate evaluation of societal biases in pre-trained transformers . In Proceedings of the 2021 Conference of the North Amer...

    2021 doi

  31. [39]

    Ian Stewart. 2014. https://doi.org/10.3115/v1/e14-3004 Now we stronger than ever: A frican- A merican E nglish syntax in T witter . In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2014, April 26-30, 2014, Got...

    2014 doi

  32. [40]

    Samson Tan, Shafiq Joty, Lav Varshney, and Min-Yen Kan. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.455 Mind your inflections! I mproving NLP for non-standard E nglishes with B ase- I nflection E ncoding . In Proceedings of the 2020 Conference on Empirical Methods in Nat...

    2020 doi

  33. [41]

    Caleb Ziems, William Held, Jingfeng Yang, Jwala Dhamala, Rahul Gupta, and Diyi Yang. 2023. https://doi.org/10.18653/v1/2023.acl-long.44 Multi- VALUE : A framework for cross-dialectal E nglish NLP . In Proceedings of the 61st Annual Meeting of the Association for Computational ...

    2023 doi

Pith tools

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