Pith. sign in

REVIEW 4 major objections 4 minor 56 references

Mitigating Confounding in Speech-Based Dementia Detection through Weight Masking

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Masking gender-linked weights yields fairer dementia screening

desk verdict A useful extension of confounding filters to transformers, but the reported gains may be inflated by train/test speaker overlap from the resampling protocol. read the letter →

arxiv 2506.05610 v1 pith:H2A7UE2G submitted 2025-06-05 cs.CL

Zhecheng Sheng , Xiruo Ding , Brian Hur , Changye Li , Trevor Cohen , Serguei Pakhomov This is my paper
classification cs.CL
keywords weightmaskinggenderconfoundingdementiadetectionBERTfairnessdistributionshiftspeechtranscripts
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 argues that transformer models fine-tuned to detect dementia from speech transcripts learn to lean on speaker gender, and that this dependence can be weakened by zeroing out the specific weights that react strongly to gender. It proposes two weight-masking methods, Extended Confounding Filter (ECF) and Dual Filter (DF), which locate gender-associated weights either by probing layers one by one or by comparing the most-changed weights from a dementia model and a gender model. On two dementia speech datasets, masking those weights shrinks the false-positive-rate gap and statistical-parity gap between male and female speakers while keeping dementia-detection AUPRC close to the unmasked model, across several simulated confounding shifts. The practical stake is that a screening model whose predictions do not track gender would be less likely to misdiagnose one group because of demographic language patterns rather than cognitive decline.

What carries the argument

The load-bearing object is the weight mask, derived from confounder-reactive parameter changes in a transformer. ECF (Extended Confounding Filter) tracks normalized per-batch update magnitudes while the already dementia-tuned model is trained toward gender with layers sequentially unfrozen, then applies a per-matrix threshold to zero the most reactive entries; the original Confounding Filter only tunes and masks the classification head. DF (Dual Filter) fine-tunes two identical BERT-base checkpoints, one on dementia and one on gender, and compares the sets of top-k% most-changed weights; the chosen mask can be the intersection ($M_I$), the difference ($M_D$), or their union, and is applied to the dementia model before evaluation. The mask does the work: it removes the specific parameters responsible for gender-linked prediction patterns without retraining or added loss terms.

What would settle it

Retest the same ECF and DF masks on a strictly participant-disjoint split (all transcripts of one participant confined to exactly one of train, validation, or test). If the fairness gains and stable AUPRC vanish or reverse under that split, the central claim that masking gender-related weights deconfounds the classifier would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that a BERT-base classifier tuned on dementia narratives stores gender information in identifiable weight entries, and that surgically ablating those entries produces a deconfounded classifier. ECF retrains the fine-tuned model toward predicting the confounder (gender) while sequentially unfreezing layers, records which weights change most, and masks the top fraction in each layer. DF instead fine-tunes two models from the same checkpoint, one toward dementia and one toward gender, and masks the top-k% most-changed weights from the gender model, or the intersection or difference of the two top-k sets. Under confounding shifts created by varying $\alpha = P(\text{dementia}|\text{female}) / P(\text{dementia}|\text{male})$ between training and test, both methods reduce the absolute FPR difference between genders and the statistical-parity gap while keeping AUPRC largely intact; for instance, at $\alpha=0.2$ on DementiaBank, a 10% intersection mask drops AUPRC from 0.83 to 0.80 while the FPR gap falls from 0.23 to 0.03. The trade-off is a slight overall loss in dementia detection performance, concentrated most heavily when embedding-layer weights are ablated.

Load-bearing premise

The resampled training, validation, and test transcripts never come from the same person; the paper does not describe a participant-level split, and with only 290 and 394 unique participants across the two corpora, repeated transcripts are unavoidable. If that assumption fails, reported gains from masking could reflect memorization instead of genuine deconfounding.

Editorial extensions

If this is right

  • If the claim holds, weight masking gives clinicians a post-hoc fairness patch for an already fine-tuned transformer, with no new training objectives or architectural modules.
  • The fairness/performance trade-off is controllable through the masking ratio, letting an operator tune how much gender parity to buy against dementia-detection accuracy.
  • The framework is stated to be model-agnostic, so the same recipe should transfer to other transformer encoders and to non-binary confounders reformulated as multi-class targets.
  • The layer-wise 'ladder' results imply that lower layers and the embedding carry dementia-critical information, so future debiasing should avoid ablating those layers aggressively.
  • Under confounding shift, the intact model's FPR gap worsens, so these methods are most valuable exactly when training and deployment populations differ by gender composition.

Reading between the lines

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

  • Not tested in the paper, a natural extension is applying the same masking to other confounders, such as age, education, or recording site, in any clinical text task with labeled attributes.
  • The appendix's entanglement analysis suggests gender and dementia signals overlap most in value and output matrices at lower layers; a cheaper variant of DF could restrict masking to those matrices rather than the full network.
  • For real deployment, the masking ratio would need to be selected without access to the test shift; a validation-based calibration procedure would make the method more directly usable.
  • Because the reported performance is measured on resampled sets with repeated transcripts, the practical effect size on genuinely new patients remains the key open question this paper does not settle.
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 proposes two weight-masking methods, Extended Confounding Filter (ECF) and Dual Filter (DF), to reduce gender confounding in BERT-base dementia detection from speech transcripts. ECF fine-tunes the dementia model toward the gender label while sequentially unfreezing layers and masks the top 15% of changed weights per matrix; DF trains a second model on gender from the same checkpoint and masks the intersection or difference of the top-k changed weights in the two models. The methods are evaluated on DementiaBank and the Carolinas Conversation Collection under synthetic confounding shifts obtained by resampling with alpha_train and alpha_test=1/alpha_train. The reported metrics are AUPRC, absolute FPR gap between genders (DeltaFPR), and statistical parity gap (DeltaSP), with comparisons to a naive Confounding Filter and two adapter baselines. The authors conclude that both methods improve gender parity while maintaining comparable dementia-detection performance.

Significance. If the evaluation supports the claims, this is a useful and practical contribution: it is model-agnostic, requires no additional training objective, and the code is released. The idea of identifying confounder-reactive weights by fine-tuning on the confounder label and then ablating them is plausible, and the paper goes beyond a single unfairness metric by reporting AUPRC, DeltaFPR, and DeltaSP. However, the current quantitative support is weakened by train/test overlap risk from the resampling protocol, an ambiguous statement about healthy-only evaluation, single-run curves without uncertainty quantification, and an unspecified operating-point selection for the baseline comparison. These issues are fixable, but they are load-bearing for the central claim.

major comments (4)
  1. [Section 5.2 and Limitations] The train/validation/test splits are not described as participant-disjoint. With DementiaBank reduced to 290 unique transcripts and CCC to 394 transcripts from 70 interviewees, sampling 480 training, 120 validation, and 150 test examples necessarily reuses transcripts. If the same participant's transcript appears in both training and test, a BERT-base model can memorize surface forms, inflating AUPRC and artificially narrowing FPR gaps. The Limitations section acknowledges 'a significant amount of duplicates' but does not state that splits are grouped by participant. Please re-run with participant-level disjoint splits, or at minimum quantify the train/test overlap and show results on a truly held-out participant set.
  2. [Section 5.2] The sentence 'Consequently, only healthy cases are used in the evaluation' contradicts the reported AUPRC and the balanced test set used for DeltaSP. If only healthy cases were evaluated, AUPRC would be undefined; if the intended meaning is that only healthy cases are used in the gender-probing phase, the text must say so. This ambiguity affects the interpretation of every quantitative claim because DeltaFPR is defined among healthy cases only and AUPRC requires both classes.
  3. [Section 6, Figures 4-6] The main ECF/DF results appear to be single runs without confidence intervals or significance tests, in contrast to Table 1, which uses 5-fold cross-validation with 3 repeats. Given the small data and stochastic fine-tuning, the specific claim that at alpha=0.2 the FPR gap drops from 0.23 to 0.03 while AUPRC drops only from 0.83 to 0.80 cannot be assessed. Please provide multiple seeds/repeats and confidence intervals, or at least error bars on the curves.
  4. [Section 6.3, Figure 7] The operating point for each method in the AUPRC-DeltaFPR comparison is not specified. If the point on each curve is selected post hoc on the test set to produce a favorable trade-off, the comparison with baselines is optimistic and selection-biased. Please specify how the points were chosen (e.g., fixed k, validation-based selection) and report the corresponding operating points for all methods.
minor comments (4)
  1. [Algorithm 1] The selection step should state that the sum is over absolute values of weight changes, since raw changes can have mixed signs; otherwise 'most changed' is undefined.
  2. [Figures 4-6] The row label 'MD MD' appears in several subplots; this is likely a typo for 'MD' only.
  3. [Table S2] The row 'Data Size' reports Train 480, Validation 120, Test 150, but the table does not state how many unique participants are in each split; adding this would clarify the leakage question discussed above.
  4. [Section 3.2] The sentence 'we restrict Delta_phi_i to each W in this probing procedure' should specify whether the normalization and the threshold tau are computed per weight matrix or globally; the current text is ambiguous.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; masks are chosen by gender-training weight changes and evaluated on held-out fairness and performance metrics, with only a minor self-cited evaluation framework.

full rationale

The derivation chain is not circular. ECF and DF select masks from per-batch weight-update magnitudes accumulated while fine-tuning toward the gender label (Section 3.2 and Algorithm 1), not from the fairness metrics ΔFPR or ΔSP; the reported AUPRC, ΔFPR, and ΔSP are computed on a held-out test set sampled at α_test = 1/α_train (Section 5.2) and are not used to choose the mask or checkpoint, since early stopping uses validation AUPRC only. The statement that gender-related weights are masked is definitional in the narrow sense that the masks are constructed from weight changes during gender training, but the substantive claims—comparable dementia AUPRC and improved gender parity—are empirical and could in principle fail; indeed, the paper reports trade-offs. The one self-referential element is the confounding-shift testbed framed by Ding et al. (2024), whose authors overlap with the present paper; however, the α-parameterization is fully specified in Section 4 and is anchored to Landeiro and Culotta's formal definition of confounding shift, so the citation is not load-bearing. The acknowledged duplicate-transcript limitation (Limitations: 'the resultant dataset contains a significant amount of duplicates that may reduce the strength of the findings') is a validity and potential leakage risk, not a circularity, and the same applies to the gender-label inference caveat. No equation equates the mask to the outcome metric, and no unverified uniqueness theorem is imported. Score 2 reflects one minor non-load-bearing self-citation rather than any reduction of the central claim to its inputs.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The central empirical claims depend on the mask construction (top-k and top-15%, mask type selection), on a resampling regime that creates duplicate examples, and on the assumption that parameter update magnitude under gender fine-tuning marks gender-associated weights. These are not derived from theory; they are design choices. The invented_entities list is empty because the paper introduces no new physical or ontological entities.

free parameters (4)
  • ECF layer-wise masking ratio = 15% default; 5%, 25%, 35% in Appendix C.2
    Determines how many weights are zeroed per matrix; central to the reported fairness-performance trade-off.
  • Dual Filter top-k% threshold and mask choice = k swept 0-60%, step 1; mask type chosen from {MI, MD, MI∪MD}
    The reported favorable operating points depend on the selected k and mask type; no pre-registered selection rule is specified.
  • Resampling target prevalence = P(gender=1)=0.5, P(dementia=1)=0.5; alpha in {0.2,0.33,1,3,5}
    Creates the confounding-shift evaluation scenarios; central to all reported results and introduces duplicate examples.
  • Gender-probing training subset = non-dementia cases only
    The paper trains for gender only on healthy transcripts; this choice affects which weights are identified as gender-related.
assumptions (7)
  • domain assumption Gender labels in DB are a valid proxy for gender, inferred from biological sex.
    The paper acknowledges equating gender with biological sex is a strong assumption (Ethical Statement, Gender Clarification).
  • domain assumption Gender is a binary confounder with a causal backdoor path affecting both language and dementia status in these datasets.
    Section 4 formalizes confounding shift using P(X|Yc) and P(Yp|Yc); this causal model is assumed rather than tested.
  • domain assumption The magnitude of weight updates during fine-tuning toward a label measures how much a parameter encodes that label.
    Sections 3.2 and 3.3 track |Δφ| to rank importance; this is the core heuristic of both ECF and DF.
  • domain assumption Training the gender probe only on non-dementia cases yields gender-representative language without dementia contamination.
    Section 5.2 states only healthy cases are used to train toward gender; this assumes dementia-free text isolates gendered style.
  • domain assumption The resampled train/validation/test sets are patient-disjoint.
    Never stated; necessary for valid evaluation because the sample sizes exceed the number of unique participants.
  • standard math The confounding-shift simulation framework of Ding et al. (2024) correctly operationalizes subpopulation shift.
    Section 4 adopts this framework as a mathematical tool; it is cited from the same group but used as an external benchmark.
  • ad hoc to paper Sequentially unfreezing layers from classification head to embedding exposes all confounder-associated weights.
    ECF's probing schedule is a design choice, not a derived property; the paper itself notes other layer-freezing combinations are ignored (Limitations).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Confounding in Speech-Based Dementia Detection through Weight Masking." pith.science (2026). https://pith.science/paper/H2A7UE2G

@misc{pith2026250605610,
  author       = {Pith},
  title        = {Pith review of: Mitigating Confounding in Speech-Based Dementia Detection through Weight Masking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H2A7UE2G}},
  note         = {Machine review of arXiv:2506.05610}
}
abstract

Deep transformer models have been used to detect linguistic anomalies in patient transcripts for early Alzheimer's disease (AD) screening. While pre-trained neural language models (LMs) fine-tuned on AD transcripts perform well, little research has explored the effects of the gender of the speakers represented by these transcripts. This work addresses gender confounding in dementia detection and proposes two methods: the $\textit{Extended Confounding Filter}$ and the $\textit{Dual Filter}$, which isolate and ablate weights associated with gender. We evaluate these methods on dementia datasets with first-person narratives from patients with cognitive impairment and healthy controls. Our results show transformer models tend to overfit to training data distributions. Disrupting gender-related weights results in a deconfounded dementia classifier, with the trade-off of slightly reduced dementia detection performance.

Figures

Figures reproduced from arXiv: 2506.05610 by the authors.

Figure 1
Figure 1. (a) Illustration of the Extended Confounding Filter (ECF) Probing framework for weights identification. (b) Illustration of the Dual Filter (DF) procedure to find weights to mask. phase of training. The importance of each element in π is determined by their magnitude. A threshold function is then employed to get the mask: Mi = ( 0 if πi > τ 1 otherwise Here, τ is the k th percentile in π, where k is a hyperparameter… view at source ↗
Figure 2
Figure 2. Tracked weights in the transformer network [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. ECF filtering with 15% masking rate across different confounding shifts from two dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: ∆FPR (y-axis) on ECF and DF for different αtrain configurations. 0.00 0.05 0.10 MI train = 0.2 train = 0.33 train = 1.0 train = 3.0 train = 5.0 ECF DF 0.00 0.05 0.10 MD 0 5 10 15 0.00 0.05 0.10 MD MD 0 5 10 15 0 5 10 15 0 5 10 15 0 5 10 15 Ablation Ratio [PITH_FULL_IM…
Figure 6
Figure 6. Figure 6: ∆SP (y-axis) on ECF and DF for different αtrain configurations. The results in [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: AUPRC vs. ∆FPR among different methods for two datasets. removes a larger set of weights, exhibits compara￾ble or slightly reduced resilience compared to ECF across different α settings. This suggests that re￾moving all gender-related weights without consid￾ering their…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 29 canonical work pages

  1. [1]

    Ioana Baldini, Dennis Wei, Karthikeyan Natesan Ramamurthy, Mikhail Yurochkin, and Moninder Singh. 2022. https://arxiv.org/abs/2108.01250 Your fairness may vary: Pretrained language model fairness in toxic text classification . Preprint, arXiv:2108.01250

    work page Pith review arXiv 2022

  2. [2]

    Christopher R Beam, Cody Kaneshiro, Jung Yun Jang, Chandra A Reynolds, Nancy L Pedersen, and Margaret Gatz. 2018. Differences between women and men in incidence rates of dementia and alzheimer’s disease. Journal of Alzheimer's disease, 64(4):1077--1083

    work page 2018

  3. [3]

    JT Becker, F Boller, OL Lopez, J Saxton, and KL McGonigle. 1994. https://doi.org/10.1001/archneur.1994.00540180063015 The natural history of alzheimer's disease: Description of study cohort and accuracy of diagnosis . Archives of Neurology, 51(6):585--594

    arXiv 1994

  4. [4]

    Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. 2016. https://arxiv.org/abs/1607.06520 Man is to computer programmer as woman is to homemaker? debiasing word embeddings . Preprint, arXiv:1607.06520

    arXiv 2016

  5. [5]

    A Theory of Label Propagation for Subpopulation Shift

    Tianle Cai, Ruiqi Gao, Jason D. Lee, and Qi Lei. 2021. https://arxiv.org/abs/2102.11203 A theory of label propagation for subpopulation shift . Preprint, arXiv:2102.11203

    work page Pith review arXiv 2021

  6. [6]

    Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. 2019. Learning imbalanced datasets with label-distribution-aware margin loss. In Advances in Neural Information Processing Systems

    work page 2019

  7. [7]

    Pengyu Cheng, Weituo Hao, Siyang Yuan, Shijing Si, and Lawrence Carin. 2021. https://openreview.net/forum?id=N6JECD-PI5w Fairfil: Contrastive neural debiasing method for pretrained text encoders . In International Conference on Learning Representations

    2021

  8. [8]

    Trevor Cohen and Serguei Pakhomov. 2020. https://arxiv.org/abs/2005.03593 A tale of two perplexities: Sensitivity of neural language models to lexical retrieval deficits in dementia of the alzheimer's type . Preprint, arXiv:2005.03593

    work page Pith review arXiv 2020

Show all 56 references
  1. [9]

    Pierre Colombo, Pablo Piantanida, and Chlo \'e Clavel. 2021. https://doi.org/10.18653/v1/2021.acl-long.511 A novel estimator of mutual information for learning to disentangle textual representations . In Proceedings of the 59th Annual Meeting of the Association for Computation...

    2021 doi

  2. [10]

    Daniel de Vassimon Manela, David Errington, Thomas Fisher, Boris van Breugel, and Pasquale Minervini. 2021. https://arxiv.org/abs/2101.09688 Stereotype and skew: Quantifying gender bias in pre-trained and fine-tuned language models . Preprint, arXiv:2101.09688

    2021 arXiv

  3. [11]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associat...

    2019 doi

  4. [12]

    Xiruo Ding, Zhecheng Sheng, Meliha Yetişgen, Serguei Pakhomov, and Trevor Cohen. 2024. Backdoor adjustment of confounding by provenance for robust text classification of multi-institutional clinical notes. In AMIA ... Annual Symposium proceedings. AMIA Symposium, pages 923--932

    2024

  5. [13]

    Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Rich Zemel. 2011. https://arxiv.org/abs/1104.3913 Fairness through awareness . Preprint, arXiv:1104.3913

    2011 arXiv

  6. [14]

    Shahla Farzana and Natalie Parde. 2023. https://doi.org/10.18653/v1/2023.acl-long.668 Towards domain-agnostic and domain-adaptive dementia detection from spoken language . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Lon...

    2023 doi

  7. [15]

    Michael Gira, Ruisu Zhang, and Kangwook Lee. 2022. https://doi.org/10.18653/v1/2022.ltedi-1.8 Debiasing pre-trained language models via efficient fine-tuning . In Proceedings of the Second Workshop on Language Technology for Equality, Diversity and Inclusion, pages 59--69, Dub...

    2022 doi

  8. [16]

    Harold Goodglass and Edith Kaplan. 1983. Boston Diagnostic Aphasia Examination Booklet. Lea & Febiger, Philadelphia

    1983

  9. [17]

    Yue Guo, Changye Li, Carol Roan, Serguei Pakhomov, and Trevor Cohen. 2021. https://doi.org/10.3389/fcomp.2021.642517 Crossing the ``cookie theft'' corpus chasm: Applying what bert learns from outside data to the adress challenge dementia detection task . Frontiers in Computer ...

    2021

  10. [18]

    Yue Guo, Yi Yang, and Ahmed Abbasi. 2022. Auto-debias: Debiasing masked language models with automated biased prompts. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL 2022). Association for Computational Linguistics

    2022

  11. [19]

    Moritz Hardt, Eric Price, and Nathan Srebro. 2016. https://arxiv.org/abs/1610.02413 Equality of opportunity in supervised learning . Preprint, arXiv:1610.02413

    2016 arXiv

  12. [20]

    Lukas Hauzenberger, Shahed Masoudian, Deepak Kumar, Markus Schedl, and Navid Rekabsaz. 2023. https://doi.org/10.18653/v1/2023.findings-acl.386 Modular and on-demand bias mitigation with attribute-removal subnetworks . In Findings of the Association for Computational Linguistic...

    2023 doi

  13. [21]

    Sierra, and Jesus A

    Luis Hernandez-Dominguez, Samuel Ratté, Basilio A. Sierra, and Jesus A. Roche-Berges. 2018. https://doi.org/10.3233/JAD-180124 Computer-based evaluation of alzheimer's disease and mild cognitive impairment using lexical and syntactic information . Journal of Alzheimer's Diseas...

    2018 doi

  14. [22]

    Ben Hutchinson, Vinodkumar Prabhakaran, Emily Denton, Kellie Webster, Yu Zhong, and Stephen Denuyl. 2020. https://arxiv.org/abs/2005.00813 Social biases in nlp models as barriers for persons with disabilities . Preprint, arXiv:2005.00813

    2020 arXiv

  15. [23]

    Masahiro Kaneko and Danushka Bollegala. 2021. https://arxiv.org/abs/2101.09523 Debiasing pre-trained contextualised embeddings . Preprint, arXiv:2101.09523

    2021 arXiv

  16. [24]

    Virgile Landeiro and Aron Culotta. 2018. https://doi.org/10.1613/jair.1.11248 Robust text classification under confounding shift . J. Artif. Int. Res., 63(1):391–419

    2018 doi

  17. [25]

    Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. 2020. https://doi.org/10.1093/bioinformatics/btz682 Biobert: a pre-trained biomedical language representation model for biomedical text mining . Bioinformatics, 36(4):1234--1240

    2020 doi

  18. [26]

    Namhoon Lee, Thalaiyasingam Ajanthan, and Philip H. S. Torr. 2019. https://arxiv.org/abs/1810.02340 Snip: Single-shot network pruning based on connection sensitivity . Preprint, arXiv:1810.02340

    2019 arXiv

  19. [27]

    Changye Li, David Knopman, Weizhe Xu, Trevor Cohen, and Serguei Pakhomov. 2022. https://doi.org/10.18653/v1/2022.acl-long.131 GPT - D : Inducing dementia-related linguistic anomalies by deliberate degradation of artificial neural language models . In Proceedings of the 60th An...

    2022 doi

  20. [28]

    Changye Li, Zhecheng Sheng, Trevor Cohen, and Serguei Pakhomov. 2024. https://doi.org/10.18653/v1/2024.findings-acl.380 Too big to fail: Larger language models are disproportionately resilient to induction of dementia-related linguistic anomalies . In Findings of the Associati...

    2024 doi

  21. [29]

    Yan Liu, Yu Liu, Xiaokang Chen, Pin-Yu Chen, Daoguang Zan, Min-Yen Kan, and Tsung-Yi Ho. 2024. https://openreview.net/forum?id=SQGUDc9tC8 The devil is in the neurons: Interpreting and mitigating social biases in language models . In The Twelfth International Conference on Lear...

    2024

  22. [30]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1907.11692 Roberta: A robustly optimized bert pretraining approach . Preprint, arXiv:1907.11692

    2019 arXiv

  23. [32]

    Saturnino Luz, Fasih Haider, Sofia de la Fuente, Davida Fromm, and Brian MacWhinney. 2020 b . https://arxiv.org/abs/2004.06833 Alzheimer's dementia recognition through spontaneous speech: The ADReSS Challenge . In Proceedings of INTERSPEECH 2020, Shanghai, China

    2020 arXiv

  24. [33]

    Brian MacWhinney. 2007. https://doi.org/10.1057/9780230223936_9 The talkbank project . In Creating and Digitizing Language Corpora, pages 163--180. Palgrave Macmillan, London

    2007 doi

  25. [34]

    Shahed Masoudian, Cornelia Volaucnik, Markus Schedl, and Navid Rekabsaz. 2024. https://aclanthology.org/2024.eacl-long.150 Effective controllable bias mitigation for classification and retrieval using gate adapters . In Proceedings of the 18th Conference of the European Chapte...

    2024

  26. [35]

    Shamila Nasreen, Julian Hough, and Matthew Purver. 2021. https://doi.org/10.18653/v1/2021.sigdial-1.32 Rare-class dialogue act tagging for A lzheimer ' s disease diagnosis . In Proceedings of the 22nd Annual Meeting of the Special Interest Group on Discourse and Dialogue, page...

    2021 doi

  27. [36]

    Anh Nguyen, Jason Yosinski, and Jeff Clune. 2015. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    2015

  28. [37]

    Judea Pearl. 2009. Causality, 2nd edition. Cambridge University Press, Cambridge, UK

    2009

  29. [38]

    Podcasy and C

    Jessica L. Podcasy and C. Neill Epperson. 2016. https://doi.org/10.31887/DCNS.2016.18.4/cepperson Considering sex and gender in alzheimer disease and other dementias . Dialogues in Clinical Neuroscience, 18(4):437--446. PMID: 28179815

    2016 doi

  30. [39]

    Charlene Pope and Boyd H. Davis. 2011. Finding a balance: The carolinas conversation collection. Corpus Linguistics and Linguistic Theory, 7(1):143--161

    2011

  31. [40]

    Rebecca Qian, Candace Ross, Jude Fernandes, Eric Smith, Douwe Kiela, and Adina Williams. 2022. https://arxiv.org/abs/2205.12586 Perturbation augmentation for fairer nlp . Preprint, arXiv:2205.12586

    2022 arXiv

  32. [41]

    Shauli Ravfogel, Michael Twiton, Yoav Goldberg, and Ryan Cotterell. 2022. https://arxiv.org/abs/2201.12091 Linear adversarial concept erasure . Preprint, arXiv:2201.12091

    2022 arXiv

  33. [42]

    Yaniv Romano, Stephen Bates, and Emmanuel J. Candès. 2020. https://arxiv.org/abs/2006.04292 Achieving equalized odds by resampling sensitive attributes . Preprint, arXiv:2006.04292

    2020 arXiv

  34. [43]

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2020. https://arxiv.org/abs/1910.01108 Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter . Preprint, arXiv:1910.01108

    2020 arXiv

  35. [44]

    Aili Shen, Xudong Han, Trevor Cohn, Timothy Baldwin, and Lea Frermann. 2021. https://arxiv.org/abs/2109.10645 Contrastive learning for fair representations . Preprint, arXiv:2109.10645

    2021 arXiv

  36. [45]

    Zico Kolter

    Mingjie Sun, Zhuang Liu, Anna Bair, and J. Zico Kolter. 2024. https://arxiv.org/abs/2306.11695 A simple and effective pruning approach for large language models . Preprint, arXiv:2306.11695

    2024 arXiv

  37. [46]

    Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2013. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199

    2013 arXiv

  38. [47]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30

    2017

  39. [48]

    Haohan Wang, Akshay Meghawat, Louis-Philippe Morency, and Eric P Xing. 2017 a . Select-additive learning: Improving generalization in multimodal sentiment analysis. In IEEE International Conference on Multimedia and Expo

    2017

  40. [49]

    Haohan Wang, Bhiksha Raj, and Eric P Xing. 2017 b . On the origin of deep learning. arXiv preprint arXiv:1702.07800

    2017 arXiv

  41. [50]

    Haohan Wang, Zhenglin Wu, and Eric P. Xing. 2019. Removing confounding factors associated weights in deep neural networks improves the prediction accuracy for healthcare applications. In Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing, volume 24, pages 54--65

    2019

  42. [51]

    Kellie Webster, Xuezhi Wang, Ian Tenney, Alex Beutel, Emily Pitler, Ellie Pavlick, Jilin Chen, Ed Chi, and Slav Petrov. 2021. https://arxiv.org/abs/2010.06032 Measuring and reducing gendered correlations in pre-trained models . Preprint, arXiv:2010.06032

    2021 arXiv

  43. [52]

    Boyi Wei, Kaixuan Huang, Yangsibo Huang, Tinghao Xie, Xiangyu Qi, Mengzhou Xia, Prateek Mittal, Mengdi Wang, and Peter Henderson. 2024. https://arxiv.org/abs/2402.05162 Assessing the brittleness of safety alignment via pruning and low-rank modifications . Preprint, arXiv:2402.05162

    2024 arXiv

  44. [53]

    Yuzhe Yang, Haoran Zhang, Dina Katabi, and Marzyeh Ghassemi. 2023. https://arxiv.org/abs/2302.12254 Change is hard: A closer look at subpopulation shift . Preprint, arXiv:2302.12254

    2023 arXiv

  45. [54]

    Xiongyi Zhang, Jan-Willem van de Meent, and Byron Wallace. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.60 Disentangling representations of text by masking transformers . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 778-...

    2021 doi

  46. [55]

    Beier Zhu, Yulei Niu, Saeil Lee, Minhoe Hur, and Hanwang Zhang. 2023. https://arxiv.org/abs/2301.12429 Debiased fine-tuning for vision-language models by prompt regularization . Preprint, arXiv:2301.12429

    2023 arXiv

  47. [56]

    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...

  48. [57]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.