Pith. sign in

REVIEW 4 major objections 6 minor 44 references

Reasoning-Based Approach with Chain-of-Thought for Alzheimer's Detection Using Speech and Large Language Models

T0 review · 4 major / 6 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 Supervised fine-tuning with chain-of-thought prompts and cue counts raises Alzheimer's detection accuracy to 83.3% (from 75.0%) for ASR transcripts, and to 87.5% for ground-truth transcripts, on the ADReSS test set.

desk verdict Plausible idea, fragile evidence: the CoT gain on 48 subjects could be noise, and the abstract compares apples to oranges. read the letter →

arxiv 2506.01683 v1 pith:BNG6ZXZY submitted 2025-06-02 cs.AI

Chanwoo Park , Anna Seo Gyeong Choi , Sunghye Cho , Chanwoo Kim This is my paper
classification cs.AI
keywords Alzheimer'sdetectionchain-of-thoughtpromptinglargelanguagemodelsautomaticspeechrecognitiondementiascreeningLoRAfine-tuningcookietheftpicturedescriptioncue-countfeatures
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that chain-of-thought prompting improves LLM-based Alzheimer's detection from spontaneous speech. It couples Whisper speech-to-text with a small Llama model fine-tuned using LoRA, and prompts the model to reason about how often a speaker mentions twelve hand-picked visual cues from the cookie theft picture. On the ADReSS test set the CoT recipe reaches 83.33% accuracy and F1 for ASR transcripts versus 75.00% for fine-tuning without CoT, and 87.50% for ground-truth transcripts with CoT. The paper reports this as a 16.7% relative gain over no-CoT fine-tuning and as the best CoT-based result it knows of.

What carries the argument

The central object is the CoT prompt built from cue counts: a fixed list of twelve cookie-theft picture items (stool, sink, dish, wash, jar, cookie, child, mother, window, cabinet, kitchen, water) whose mention rates in each transcript are fed into the prompt alongside instructions to reason step by step before labelling AD or non-AD. The classification head is a linear layer on top of Llama3.2-1B-Instruct, adapted with LoRA so the base weights stay frozen. The cue counts make the reasoning partially inspectable, but the design does not separate the contribution of the cue-count feature from the contribution of the CoT reasoning text.

What would settle it

Run the same 108/48 split with the reasoning text removed from the CoT prompt while keeping the cue counts, repeated over several seeds: if accuracy stays near 83.3%, the CoT claim reduces to cue-count features; if bootstrapped confidence intervals on the original comparison contain 75.0%, the gain is not established.

Watch

Extended reading notes

Core claim

The central claim is that telling the model to count and reason about a fixed set of picture cues, rather than just presenting the transcript, changes the classification outcome. The pipeline transcribes audio with Whisper large-v2, computes the proportion of the twelve cues each participant produced, inserts those counts into a CoT prompt, and fine-tunes Llama3.2-1B-Instruct with a linear classification head using LoRA. Against the baseline that fine-tunes on the same transcripts without CoT prompts, the CoT version improves ASR-based accuracy from 75.0% to 83.3% and ground-truth accuracy from 83.3% to 87.5%. The paper interprets this as evidence that intermediate reasoning steps, guided by cues, add signal that plain supervised fine-tuning misses.

Load-bearing premise

The claimed gain rests on one run over a 48-person test set with no confidence interval or significance test, and on a hand-picked 12-word cue list whose effect is never separated from the reasoning prompt.

Editorial extensions

If this is right

  • CoT prompting could improve LLM-based dementia screening without adding a second modality or a larger model.
  • The pipeline is cheap to reproduce: a 1B-parameter LLM, a single consumer GPU, and the public ADReSS split are sufficient.
  • With ground-truth transcripts the CoT model reaches 87.5% accuracy, so transcription errors are the main remaining bottleneck for the ASR version.
  • Because the prompt names concrete cues, the model's decisions are more auditable than a black-box classifier, which matters for clinical acceptance.

Reading between the lines

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

  • The reported gain might come from the cue counts alone: a classifier fed only the twelve cue proportions could match the CoT numbers, since the ablation never removes cues while keeping the reasoning text.
  • The single 48-person test set with no confidence intervals means the 8.3-point ASR gain could be sampling noise; repeated random seeds or cross-validation over all 156 participants would settle this.
  • If the cue counts are the real driver, the LLM and CoT prompt could be replaced by a transparent feature-based model; if the reasoning text matters, the prompt should still work when the picture description or cue list changes.
  • A direct transfer test on a second dementia speech corpus, with the same fixed cue list and prompt, would show whether the reasoning step generalizes or is tuned to the cookie theft picture.
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 / 6 minor

Summary. The paper proposes an Alzheimer's disease (AD) detection pipeline on the ADReSS benchmark: Whisper ASR transcribes the cookie-theft speech, Llama3.2-1B is fine-tuned with LoRA and a linear classification head, and a Chain-of-Thought (CoT) prompt is used during supervised fine-tuning. The CoT prompt includes counts of 12 hand-picked cue words computed from the same transcript. The authors report that CoT achieves 83.33% accuracy on ASR transcripts and 87.50% on ground-truth transcripts, compared to 75.00% for the ASR non-CoT baseline and 83.33% for the ground-truth non-CoT baseline. They claim a 16.7% relative improvement over methods without CoT and state state-of-the-art performance among CoT approaches. An ablation study and a comparison with prior ADReSS results are provided.

Significance. If the reported gains are reproducible, the paper would demonstrate a simple, parameter-efficient way to inject explicit reasoning cues into an LLM-based dementia screener, evaluated on a widely used public benchmark, with source code released via a DOI. The within-modality comparison (ASR no-CoT 75.0% vs. ASR CoT 83.3%) is the correct basis for the central claim and is suggestive. However, the current evidence is not yet adequate: every accuracy figure comes from a single training run, the decisive differences involve only two to four test participants, and the CoT condition is confounded with the addition of a hand-selected cue-count feature. The paper also contains internal inconsistencies in the reported dataset sizes and in the headline improvement figure.

major comments (4)
  1. [Abstract and §5] The accuracy differences that support the central CoT claim are within sampling noise on this 48-participant test set. One misclassified participant changes accuracy by about 2.08 percentage points. The ASR comparison (75.00% vs. 83.33%) is a four-subject difference, and the ground-truth comparison (83.33% vs. 87.50%) is a two-subject difference. All numbers come from a single training run with no standard deviations, confidence intervals, or significance tests. Under the most favorable discordance pattern, McNemar's exact test gives two-sided p≈0.125 for the ASR comparison and p≈0.5 for the ground-truth comparison. Please report multiple seeds (mean±std) and a paired significance test, and state whether hyperparameters and the cue list were fixed before the test set was used. Without this, the headline improvement could easily be sampling noise.
  2. [Abstract and §5] The abstract's central quantitative claim, a '16.7% relative performance improvement', is not an apples-to-apples comparison. The value is (87.50−75.00)/75.00, which compares the ground-truth CoT result against the ASR no-CoT baseline. The comparable within-modality relative gains are 11.1% for ASR (83.33 vs. 75.00) and 5.0% for ground truth (87.50 vs. 83.33). Please revise the abstract and conclusions to report the proper within-modality comparisons, or explicitly state which baselines are being compared.
  3. [§3.3.1 and Table 4] The CoT condition differs from the non-CoT baselines in two ways: it uses a reasoning prompt and it injects the counts of the 12 hand-picked cue words into the prompt. The baseline conditions have neither. The ablation in Table 4 therefore conflates the effect of the reasoning prompt with the effect of the cue-count feature. To support the claim that CoT reasoning itself helps, please include an ablation that keeps cue counts fixed across conditions (e.g., baseline SFT with cue counts, and CoT without cue counts) or otherwise decomposes the two contributions.
  4. [§3.1] The dataset description is internally inconsistent. The text states a total of 156 participants (78 non-AD and 78 AD), then says the training set has 108 participants consisting of 48 AD and 48 non-AD participants, and the test set has 48 participants (24 AD and 24 non-AD). These numbers sum to 144 participants, not 156, and the listed training participants sum to 96, not 108. Section 4.1 also refers to 'the original split of 108 participants for the SFT training set.' Please reconcile these numbers; the discrepancy is material because the claimed training-set size affects the credibility of the comparison with prior ADReSS work.
minor comments (6)
  1. [Abstract] The phrase 'an 16.7% relative performance improvement' should read 'a 16.7% relative performance improvement.'
  2. [Table 3] The final row for 'Ours (Llama3.2-1B)' appears to combine two experimental conditions (with and without speech) into one row with two result columns. Please split this into two clearly labeled rows, e.g., 'Ours (ASR)' and 'Ours (GT).'
  3. [§4.2 and Figure 2] The text says 'The CoT prompts we used are shown in Figure 2,' but Figure 2 shows only an example. Providing the full prompt template in an appendix or in the code repository would improve reproducibility.
  4. [§3.3.1] The abbreviation 'CTD' is used in 'generates important cues from the CTD' without being re-introduced after its definition in §3.1; consider defining it at first use in §3.3.1.
  5. [§5] The conclusion cites an '11.1% improvement in accuracy,' which matches the ASR CoT relative gain but not the abstract's 16.7%. Align these claims after revising the abstract and results sections.
  6. [Table 4] The row label 'Baseline (ASR CoT)' is potentially confusing because Table 2's 'Baseline' is the ASR condition without CoT. Use distinct labels such as 'ASR without CoT' and 'ASR with CoT' throughout the tables for consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: predictions are not defined in terms of labels, the evaluation uses an external benchmark, and no load-bearing self-citation is present.

full rationale

I traced the claimed derivation chain: ASR transcription, extraction of hand-selected cue counts from the transcript, construction of a CoT prompt, and supervised fine-tuning of a LoRA-adapted Llama model with a linear classification head trained on AD/non-AD labels. The cue-count features are inputs computed from the test transcript, not fitted parameters, and the label prediction comes from a head trained on held-out labels; no equation or construction makes the output equal to the input. The comparison tables use the external ADReSS test set, and the related-work comparisons are against prior published systems rather than self-citations. The paper cites standard external references for Whisper, Llama, LoRA, and CoT, and no uniqueness theorem or prior author result is invoked to make the method forced. The abstract's 16.7% relative improvement figure compares ground-truth CoT accuracy against the ASR no-CoT baseline, which is a reporting inconsistency rather than a circular reduction; likewise, the small test set and single-run results are statistical robustness concerns, not evidence that any result is true by definition. Therefore no specific circular step can be exhibited.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the ADReSS labels, the fidelity of Whisper transcriptions, the hand-picked cue list, and the single-run evaluation on a 48-sample test set. No new physical entities are introduced. The cue list is a hand-engineered feature, not a fitted parameter in the mathematical sense, but it is an unvalidated design choice that the performance gain may depend on.

free parameters (2)
  • Hand-picked cue list (12 words from the cookie theft picture) = stool, sink, dish, wash, jar, cookie, child, mother, window, cabinet, kitchen, water
    The CoT prompt is built around counts of these cue words. The list is chosen by the authors, not derived from data or prior literature, and the paper does not validate the choice.
  • LoRA and training hyperparameters = rank 16, alpha 16, dropout 0.01, batch size 8, learning rate 1e-4, weight decay 0.001
    Chosen by hand with no reported validation-set search; a single configuration is used for all runs. These are standard choices but not justified by ablations.
assumptions (3)
  • domain assumption ADReSS labels (AD/non-AD) are correct ground truth
    The DementiaBank Pitt corpus labels are treated as gold without independent verification; used throughout Section 3.1 and in all evaluations.
  • domain assumption Whisper large-v2 transcriptions adequately preserve AD-relevant linguistic content
    The whole ASR pipeline depends on this assumption; stated in Section 3.2. If ASR drops or distorts diagnostic cues, the classification in the ASR setting is affected.
  • ad hoc to paper The proportion of hand-picked cue words in a transcript is a valid indicator of AD
    The CoT prompt feeds these proportions to the LLM as reasoning guidance (Section 3.3.1). This feature is designed by the authors, not derived from theory or validated independently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reasoning-Based Approach with Chain-of-Thought for Alzheimer's Detection Using Speech and Large Language Models." pith.science (2026). https://pith.science/paper/BNG6ZXZY

@misc{pith2026250601683,
  author       = {Pith},
  title        = {Pith review of: Reasoning-Based Approach with Chain-of-Thought for Alzheimer's Detection Using Speech and Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BNG6ZXZY}},
  note         = {Machine review of arXiv:2506.01683}
}
read the original abstract

Societies worldwide are rapidly entering a super-aged era, making elderly health a pressing concern. The aging population is increasing the burden on national economies and households. Dementia cases are rising significantly with this demographic shift. Recent research using voice-based models and large language models (LLM) offers new possibilities for dementia diagnosis and treatment. Our Chain-of-Thought (CoT) reasoning method combines speech and language models. The process starts with automatic speech recognition to convert speech to text. We add a linear layer to an LLM for Alzheimer's disease (AD) and non-AD classification, using supervised fine-tuning (SFT) with CoT reasoning and cues. This approach showed an 16.7% relative performance improvement compared to methods without CoT prompt reasoning. To the best of our knowledge, our proposed method achieved state-of-the-art performance in CoT approaches.

Figures

Figures reproduced from arXiv: 2506.01683 by the authors.

Figure 1
Figure 1. The process of LLM for dementia detection. (a) The preprocessing stage involves speech recognition of participant responses using a speech model and generation of important visual cues. (b) In the supervised fine-tuning (SFT) stage, utilize the pre-trained language model Llama to classify between AD and non-AD cases. system (Amyloid-β, Tau, Neurodegeneration) are reshaping Alzheimer’s disease diagnostics toward prec… view at source ↗
Figure 2
Figure 2. Example of the reasoning guidance Chain-of-Thought (CoT) prompts provided to the model for diagnosing dementia. 4.3. Comparison of results with other studies The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 35 canonical work pages

  1. [1]

    It significantly impacts patients’ ability to perform daily activities, thereby severely affecting their quality of life

    Introduction Dementia is a neurodegenerative disorder characterized by the progressive decline of cognitive functions. It significantly impacts patients’ ability to perform daily activities, thereby severely affecting their quality of life. According to the World Health Organization (WHO), approximately 50 million people worldwide are living with dementia...

    work page

  2. [2]

    Diagnostic criteria from the DSM-IV/V and ICD- 10 remain foundational, though middle-income countries of- ten lack resources for advanced neuroimaging or biomarker testing [6]

    Related Works Existing research on dementia diagnosis emphasizes the widespread use of cognitive screening tools like the mini- mental state examination (MMSE) and clinical dementia rat- ing (CDR) in primary care settings, particularly in high-income countries [6]. Diagnostic criteria from the DSM-IV/V and ICD- 10 remain foundational, though middle-income...

    arXiv 2025

  3. [3]

    Methodology 3.1. Dataset Our experiment utilized data from the cookie theft picture de- scription task of the Boston Diagnostic Aphasia Exam (BDAE) [21–23], which is part of DementiaBank’s Pitt Corpus [2]. The transcripts, annotated using the CHAT coding system [24], were acoustically enhanced through static noise removal. Audio vol- ume was normalized ac...

    work page

  4. [4]

    Experiment results and analysis 4.1. Experimental Setup For our experiments, we trained the model using the original split of 108 participants for the SFT training set, and con- ducted evaluations using the remaining 48 participants in the test set. We trained the model on a NVIDIA GeForce RTX 4090 GPU. For the hyperparameter configuration, we used a batc...

    work page

  5. [5]

    Conclusions Advancements in AI technology, particularly LLM and multi- modal approaches, are driving transformative changes in de- mentia research and treatment. Models that combine auto- matic speech recognition (ASR) have demonstrated a 11.1% Method Acc (%) F1 (%) Baseline 83.33 83.22(ASR CoT) ground truth (SFT) 83.33 83.30 ground truth (CoT)87.50 87.48...

    work page

  6. [6]

    Acknowledgment This work was supported by the IITP(Institute of Informa- tion & Coummunications Technology Planning & Evaluation)- ITRC(Information Technology Research Center) grant funded by the Korea government(Ministry of Science and ICT)(IITP- 2025-RS-2024-00436857), IITP grant funded by the Korea government(MSIT) (No. RS-2019-II190079, Artificial Int...

    work page 2025

  7. [7]

    A clinicopathological approach to the diagnosis of dementia,

    F. M. Elahi and B. L. Miller, “A clinicopathological approach to the diagnosis of dementia,”Nature Reviews Neurology, vol. 13, no. 8, pp. 457–476, 2017

    work page 2017

  8. [8]

    Dementia prevention, interven- tion, and care: 2020 report of the lancet commission,

    G. Livingston, J. Huntley, A. Sommerlad, D. Ames, C. Ballard, S. Banerjee, and N. Mukadam, “Dementia prevention, interven- tion, and care: 2020 report of the lancet commission,”The Lancet, vol. 396, no. 10248, pp. 413–446, 2020

    work page 2020

Show all 44 references
  1. [9]

    Alzheimer’s dementia recognition through spontaneous speech,

    S. Luz, F. Haider, S. de la Fuente Garcia, D. Fromm, and B. MacWhinney, “Alzheimer’s dementia recognition through spontaneous speech,”Frontiers in Computer Science, vol. 3, p. 780169, 2021

    2021

  2. [10]

    Parameter-efficient fine-tuning of large-scale pre-trained lan- guage models,

    N. Ding, Y . Qin, G. Yang, F. Wei, Z. Yang, Y . Su, and M. Sun, “Parameter-efficient fine-tuning of large-scale pre-trained lan- guage models,”Nature Machine Intelligence, vol. 5, no. 3, pp. 220–235, 2023

    2023

  3. [11]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  4. [12]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” inAdvances in Neural Information Processing Systems, vol. 35, 2022, pp. 24 824–24 837

    2022

  5. [13]

    Diagnosing dementia and cognitive dysfunction in the el- derly in primary health care: a systematic review,

    L. N. Pelegrini, G. M. Mota, C. F. Ramos, E. Jesus, and F. A. Vale, “Diagnosing dementia and cognitive dysfunction in the el- derly in primary health care: a systematic review,”Dementia & Neuropsychologia, vol. 13, pp. 144–153, 2019

    2019

  6. [14]

    Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,

    W. N. Hsu, B. Bolte, Y . H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,”IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

    2021

  7. [15]

    Alzheimer’s dis- ease: A review on the current trends of the effective diagnosis and therapeutics,

    A. S. Abdul Manap, R. Almadodi, S. Sultana, M. G. Sebastian, K. S. Kavani, V . E. Lyenouq, and A. Shankar, “Alzheimer’s dis- ease: A review on the current trends of the effective diagnosis and therapeutics,”Frontiers in Aging Neuroscience, vol. 16, p. 1429211, 2024

    2024

  8. [16]

    The dementia diagnosis: a literature review of information, understanding, and attributions,

    L. Stokes, H. Combes, and G. Stokes, “The dementia diagnosis: a literature review of information, understanding, and attributions,” Psychogeriatrics, vol. 15, no. 3, pp. 218–225, 2015

    2015

  9. [17]

    Noninvasive automatic detection of alzheimer’s disease from spontaneous speech: a re- view,

    X. Qi, Q. Zhou, J. Dong, and W. Bao, “Noninvasive automatic detection of alzheimer’s disease from spontaneous speech: a re- view,”Frontiers in Aging Neuroscience, vol. 15, p. 1224723, 2023

    2023

  10. [18]

    An assessment of paralin- guistic acoustic features for detection of alzheimer’s dementia in spontaneous speech,

    F. Haider, S. De La Fuente, and S. Luz, “An assessment of paralin- guistic acoustic features for detection of alzheimer’s dementia in spontaneous speech,”IEEE Journal of Selected Topics in Signal Processing, vol. 14, no. 2, pp. 272–281, 2019

    2019

  11. [19]

    Automatic detection of alzheimer’s disease using spontaneous speech only,

    J. Chen, J. Ye, F. Tang, and J. Zhou, “Automatic detection of alzheimer’s disease using spontaneous speech only,” inInter- speech, vol. 2021. NIH Public Access, August 2021, p. 3830

    2021

  12. [20]

    wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” inAdvances in neural information processing systems, vol. 33, 2020, pp. 12 449–12 460

    2020

  13. [21]

    The assessment of aphasia and related disorders,

    I. T. Draper, “The assessment of aphasia and related disorders,” Journal of Neurology, Neurosurgery, and Psychiatry, vol. 36, no. 5, p. 894, 1973

    1973

  14. [22]

    Exploring multi- task learning and data augmentation in dementia detection with self-supervised pretrained models,

    M. Chen, C. Miao, J. Ma, S. Wang, and J. Xiao, “Exploring multi- task learning and data augmentation in dementia detection with self-supervised pretrained models,” inProc. INTERSPEECH, vol. 2023, 2023, pp. 5037–5041

    2023

  15. [23]

    Llama 2: Open foundation and fine- tuned chat models,

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, and T. Scialom, “Llama 2: Open foundation and fine- tuned chat models,”arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  16. [24]

    Mistral 7b,

    A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chap- lot, D. D. L. Casas, and W. E. Sayed, “Mistral 7b,”arXiv preprint arXiv:2310.06825, 2023

    2023 arXiv

  17. [25]

    Optimizing and evaluating pre-trained large language models for alzheimer’s disease detection,

    F. Casu, E. Grosso, A. Lagorio, and G. A. Trunfio, “Optimizing and evaluating pre-trained large language models for alzheimer’s disease detection,” in2024 32nd Euromicro International Con- ference on Parallel, Distributed and Network-Based Processing (PDP). IEEE, March 2024, p...

    2024

  18. [26]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Ale- man, and B. McGrew, “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  19. [27]

    Alzheimer’s disease recognition from spontaneous speech using large language mod- els,

    J. U. Bang, S. H. Han, and B. O. Kang, “Alzheimer’s disease recognition from spontaneous speech using large language mod- els,”ETRI Journal, vol. 46, no. 1, pp. 96–105, 2024

    2024

  20. [28]

    Automated recognition of alzheimer’s dementia using bag-of-deep-features and model ensembling,

    Z. S. Syed, M. S. S. Syed, M. Lech, and E. Pirogova, “Automated recognition of alzheimer’s dementia using bag-of-deep-features and model ensembling,”IEEE Access, vol. 9, pp. 88 377–88 390, 2021

    2021

  21. [29]

    Goodglass and E

    H. Goodglass and E. Kaplan,Boston diagnostic aphasia exami- nation booklet. Lea & Febiger, 1983

    1983

  22. [30]

    Goodglass, E

    H. Goodglass, E. Kaplan, and S. Weintraub,BDAE: The Boston diagnostic aphasia examination. Philadelphia, PA: Lippincott Williams & Wilkins, 2001

    2001

  23. [31]

    The childes project part 1: The chat transcrip- tion format,

    B. Macwhinney, “The childes project part 1: The chat transcrip- tion format,” inThe CHILDES Project, 2009

    2009

  24. [32]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” inInternational Conference on Machine Learning. PMLR, July 2023, pp. 28 492–28 518

    2023

  25. [33]

    The llama 3 herd of models,

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Let- man, and R. Ganapathy, “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  26. [34]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhari- wal, and D. Amodei, “Language models are few-shot learners,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 1877–1901

    2020

  27. [35]

    Self-supervised learning: Generative or contrastive,

    X. Liu, F. Zhang, Z. Hou, L. Mian, Z. Wang, J. Zhang, and J. Tang, “Self-supervised learning: Generative or contrastive,”IEEE trans- actions on knowledge and data engineering, vol. 35, no. 1, pp. 857–876, 2021

    2021

  28. [36]

    Comparing acoustic-based approaches for alzheimer’s disease detection,

    A. Balagopalan and J. Novikova, “Comparing acoustic-based approaches for alzheimer’s disease detection,”arXiv preprint arXiv:2106.01555, 2021

    2021 arXiv

  29. [37]

    models explored transfer learning techniques for AD clas- sification and MMSE regression tasks. The transfer learning models were pre-trained on general large-scale datasets and fine-tuned and tested using the ADReSS dataset, achieving 82.9 ±1.56 and 82.9±1.86 for accuracy and...

  30. [38]

    A comparison of acoustic and linguistics methodologies for alzheimer’s dementia recognition,

    N. Cummins, Y . Pan, Z. Ren, J. Fritsch, V . S. Nallanthighal, H. Christensen, and A. H ¨arm¨a, “A comparison of acoustic and linguistics methodologies for alzheimer’s dementia recognition,” inInterspeech 2020. ISCA-International Speech Communication Association, October 2020,...

    2020

  31. [39]

    Exploring deep transfer learning techniques for alzheimer’s dementia detection,

    Y . Zhu, X. Liang, J. A. Batsis, and R. M. Roth, “Exploring deep transfer learning techniques for alzheimer’s dementia detection,” Frontiers in Computer Science, vol. 3, p. 624683, 2021

    2021

  32. [40]

    Decoupled weight decay regularization,

    I. Loshchilov, “Decoupled weight decay regularization,”arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  33. [41]

    Linear dis- criminant analysis,

    P. Xanthopoulos, P. M. Pardalos, and T. B. Trafalis, “Linear dis- criminant analysis,” inRobust data mining, 2013, pp. 27–33

    2013

  34. [42]

    Support vector machines,

    M. A. Hearst, S. T. Dumais, E. Osuna, J. Platt, and B. Scholkopf, “Support vector machines,”IEEE Intelligent Systems and their applications, vol. 13, no. 4, pp. 18–28, 1998

    1998

  35. [43]

    Attention-based bidirectional long short-term memory networks for relation classification,

    P. Zhou, W. Shi, J. Tian, Z. Qi, B. Li, H. Hao, and B. Xu, “Attention-based bidirectional long short-term memory networks for relation classification,” inProceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Vol- ume 2: Short Papers), August ...

    2016

  36. [44]

    Speechbert: An audio-and-text jointly learned language model for end-to-end spo- ken question answering,

    Y . S. Chuang, C. L. Liu, H. Y . Lee, and L. S. Lee, “Speechbert: An audio-and-text jointly learned language model for end-to-end spo- ken question answering,”arXiv preprint arXiv:1910.11559, 2019

    1910 arXiv

Pith tools

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