pith. machine review for the scientific record. sign in

arxiv: 2605.14380 · v1 · submitted 2026-05-14 · 💻 cs.CL

Recognition: no theorem link

Mitigating Data Scarcity in Psychological Defense Classification with Context-Aware Synthetic Augmentation

Hoang-Thuy-Duong Vu , Quoc-Cuong Pham , Huy-Hieu Pham
Authors on Pith no claims yet

Pith reviewed 2026-05-15 02:31 UTC · model grok-4.3

classification 💻 cs.CL
keywords psychological defense mechanismssynthetic data augmentationlow-resource classificationhybrid NLP modelmental health text analysisclinical NLPdata scarcity
0
0 comments X

The pith

Context-aware synthetic augmentation with hybrid modeling lifts psychological defense mechanism classification to 58.26% accuracy under data scarcity.

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

The paper addresses the problem of classifying unconscious psychological defense mechanisms from text, a clinically useful task blocked by extreme data scarcity and class imbalance. It proposes generating synthetic training examples by prompting language models with precise clinical definitions of the mechanisms, then feeding both synthetic and real data into a hybrid classifier that merges contextual embeddings with basic clinical features. Experiments on the PsyDefDetect shared task show that higher-quality definitions produce better synthetic data and higher final accuracy, with the full pipeline outperforming the prior DMRS Co-Pilot baseline by more than 40 points in accuracy and 15 points in macro-F1. The result supplies the first strong, reproducible baseline for psychologically grounded defense-mechanism classification in low-resource conditions.

Core claim

Prompting with defense-mechanism definitions produces synthetic examples whose quality directly determines downstream performance; when these examples are combined with 150 annotated items in a hybrid model of contextual language representations and clinical features, the system reaches 58.26% accuracy and 24.62% macro-F1 on the PsyDefDetect task, surpassing the DMRS Co-Pilot by 40.25 and 15.99 points respectively and thereby establishing a strong baseline for low-resource psychological defense classification.

What carries the argument

The context-aware synthetic augmentation framework that generates examples grounded in clinical defense-mechanism definitions, integrated with a hybrid classification model using contextual language representations plus basic clinical features.

If this is right

  • Definition quality in the prompt directly governs generation fidelity and therefore final accuracy.
  • Hybrid models that fuse deep contextual features with domain clinical indicators outperform purely data-driven baselines in this setting.
  • Targeted synthetic augmentation outperforms generic generative methods when clinical grounding is required.
  • The pipeline supplies a reproducible baseline that future shared tasks on psychological text classification can build upon.

Where Pith is reading between the lines

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

  • The same definition-grounded prompting approach could be tested on other scarce clinical NLP tasks such as emotion or symptom detection.
  • Adding an automatic fidelity filter before training might further reduce any residual noise introduced by the generator.
  • Scaling the method to multilingual or longitudinal clinical text could reveal whether the same augmentation logic holds across languages or time.

Load-bearing premise

Synthetic examples created by prompting with defense-mechanism definitions keep enough psychological fidelity to improve classification without adding label noise or artifacts.

What would settle it

An expert review that finds the synthetic examples frequently violate clinical definitions of defense mechanisms, or an ablation showing that removing the synthetic data leaves performance unchanged or higher, would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.14380 by Hoang-Thuy-Duong Vu, Huy-Hieu Pham, Quoc-Cuong Pham.

Figure 1
Figure 1. Figure 1: Overview of the multi-stage research pipeline. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) The PSYDEFCONV official test set label distribution and (b) row-normalized confusion matrix of our official leaderboard submission (×8). Label 7 dominates both the distribution (243/472 instances) and predictions, absorbing errors from all other classes. bels 0 and 7 exceed F1 > 0.70, while all remaining classes fall below 0.30, with four classes below 0.15. This implies that the accuracy (0.55-0.58) s… view at source ↗
Figure 5
Figure 5. Figure 5: Correlation between magnitude and speed of [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 3
Figure 3. Figure 3: Class distribution across defense levels in the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: Distribution of turns at which seekers exhibit [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 4
Figure 4. Figure 4: Defense level trajectory across turns in di [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Composite Disclosure Index (CDI) across normalized dialogue progression. Disclosure peaks at the 10–20% mark then stabilizes, suggesting defensive activation intensifies after initial vulnerability [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Seeker response time per defense label. Label [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distribution of number of turns per dialogue. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 12
Figure 12. Figure 12: Mean NLI-inferred DMRS mechanism ac￾tivation per defense class (log-entailment scores). All values are negative due to log-probability scaling. Dif￾ferential gradients on Autistic Fantasy, Undoing, and Affiliation provide discriminative signal for the hybrid fusion model despite uniformly low absolute scores [PITH_FULL_IMAGE:figures/full_fig_p009_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Off-diagonal misclassification counts (best [PITH_FULL_IMAGE:figures/full_fig_p009_13.png] view at source ↗
read the original abstract

Psychological defense mechanisms (PDMs) are unconscious cognitive processes that modulate how individuals perceive and respond to emotional distress. Automatically classifying PDMs from text is clinically valuable but severely hindered by data scarcity and class imbalance, challenges which generative augmentation alone cannot resolve without psychological grounding. In this work, we address these challenges in the PsyDefDetect shared task (BioNLP@ACL 2026) by proposing a context-aware synthetic augmentation framework combined with a hybrid classification model. Our hybrid model integrates contextual language representations with basic clinical features, along with 150 annotated defense items. Experiments demonstrate that definition quality in prompting directly governs generation fidelity and downstream performance. Our method surpasses DMRS Co-Pilot, reaching an accuracy of 58.26% (+40.25%) and a macro-F1 of 24.62% (+15.99%), thereby establishing a strong baseline for psychologically grounded defense mechanism classification in low-resource settings. Source code is available at: https://github.com/htdgv/CASA-PDC.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper claims to mitigate data scarcity and class imbalance in psychological defense mechanism (PDM) classification from text by introducing a context-aware synthetic augmentation framework. Synthetic examples are generated via prompting with defense-mechanism definitions, combined with a hybrid classifier that integrates contextual embeddings and basic clinical features, trained alongside 150 annotated items. On the PsyDefDetect shared task, the approach reportedly surpasses the DMRS Co-Pilot baseline, achieving 58.26% accuracy (+40.25%) and 24.62% macro-F1 (+15.99%), and positions itself as a strong baseline for psychologically grounded classification in low-resource settings.

Significance. If the central results hold after verification, the work would supply a useful empirical baseline for low-resource psychological text classification tasks. The emphasis on definition quality governing generation fidelity offers a concrete direction for future augmentation methods in clinical NLP, and the public code release supports reproducibility.

major comments (2)
  1. [Abstract] Abstract: the headline gains (58.26% accuracy, +40.25%; 24.62% macro-F1, +15.99%) are presented without any experimental protocol, data splits, statistical tests, error bars, or ablation studies, leaving open whether the improvements are robust or attributable to synthetic label noise.
  2. [Context-aware synthetic augmentation framework] Context-aware synthetic augmentation framework: the claim that definition quality in prompting governs generation fidelity is load-bearing for attributing performance gains to the proposed method, yet no expert review, inter-annotator agreement, or quantitative fidelity metric is reported on the generated examples; in an imbalanced low-resource setting this risks confounding the hybrid model's results with artifacts.
minor comments (2)
  1. [Hybrid classification model] The hybrid model's integration of contextual embeddings with clinical features is described at a high level; a concrete feature list or ablation isolating their contribution would improve clarity.
  2. [Abstract] The manuscript states source code is available at the cited GitHub link; confirming that the repository includes the exact data splits and generation prompts used for the reported numbers would strengthen reproducibility claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive review of our manuscript on context-aware synthetic augmentation for psychological defense mechanism classification. We address each major comment in detail below and have made revisions to improve the clarity and robustness of the presentation.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline gains (58.26% accuracy, +40.25%; 24.62% macro-F1, +15.99%) are presented without any experimental protocol, data splits, statistical tests, error bars, or ablation studies, leaving open whether the improvements are robust or attributable to synthetic label noise.

    Authors: We agree that the abstract would benefit from additional context on the evaluation protocol. In the revised manuscript, we have updated the abstract to briefly note the use of 5-fold cross-validation on the 150 annotated items from the PsyDefDetect shared task, along with statistical significance testing via paired t-tests (p < 0.01) and the inclusion of error bars in the reported results. The full experimental details, data splits (80/10/10 train/validation/test), and ablation studies (Table 3) demonstrating the contribution of synthetic augmentation versus baseline components remain in Section 4. These ablations show consistent gains across folds, indicating the improvements are not attributable to synthetic label noise. revision: yes

  2. Referee: [Context-aware synthetic augmentation framework] Context-aware synthetic augmentation framework: the claim that definition quality in prompting governs generation fidelity is load-bearing for attributing performance gains to the proposed method, yet no expert review, inter-annotator agreement, or quantitative fidelity metric is reported on the generated examples; in an imbalanced low-resource setting this risks confounding the hybrid model's results with artifacts.

    Authors: We acknowledge the value of direct validation metrics for the synthetic examples. The original experiments included ablations that varied prompt definition quality and measured downstream effects on classification performance, supporting the claim that higher-fidelity generations improve results. In the revised version, we have added a quantitative fidelity assessment using average cosine similarity between sentence embeddings of generated and real examples (reported in Section 3.2), along with representative generation examples in the appendix. While a full expert review and inter-annotator agreement study were not feasible within the low-resource shared-task constraints, the ablation results isolate the effect of definition quality and show performance degradation with lower-quality prompts, reducing the risk of confounding artifacts. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical augmentation and classification results stand on experimental comparison

full rationale

The paper describes a context-aware synthetic data generation pipeline followed by training a hybrid classifier on the augmented set, then reports accuracy and macro-F1 against an external baseline (DMRS Co-Pilot). No equations, parameter-fitting steps, or derivations are present that reduce any claimed result to its own inputs by construction. The load-bearing assumption (that definition-prompted generations preserve psychological fidelity) is treated as an empirical hypothesis tested via downstream metrics rather than asserted tautologically or justified solely by self-citation. The work is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on the domain assumption that LLM-generated text prompted by psychological definitions can serve as faithful training data; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)
  • domain assumption Synthetic data generated from psychological definitions can faithfully augment real annotated data for classification.
    Invoked to justify the augmentation strategy that addresses data scarcity.

pith-pipeline@v0.9.0 · 5484 in / 1193 out tokens · 61918 ms · 2026-05-15T02:31:21.429442+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages

  1. [1]

    2024 , url =

    Llama 3 Model Card , author=. 2024 , url =

    work page 2024

  2. [2]

    Focal Loss for Dense Object Detection , year=

    Lin, Tsung-Yi and Goyal, Priya and Girshick, Ross and He, Kaiming and Dollár, Piotr , booktitle=. Focal Loss for Dense Object Detection , year=

    work page

  3. [3]

    The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval , pages =

    Zhu, Yaoming and Lu, Sidi and Zheng, Lei and Guo, Jiaxian and Zhang, Weinan and Wang, Jun and Yu, Yong , title =. The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval , pages =. 2018 , isbn =. doi:10.1145/3209978.3210080 , abstract =

    work page doi:10.1145/3209978.3210080 2018

  4. [4]

    2020 , eprint=

    Supervised Multimodal Bitransformers for Classifying Images and Text , author=. 2020 , eprint=

    work page 2020

  5. [5]

    Educational and Psychological Measurement , year=

    A Coefficient of Agreement for Nominal Scales , author=. Educational and Psychological Measurement , year=

    work page

  6. [6]

    Brown, Tom B. and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and Herbert-Voss, Ariel and Krueger, Gretchen and Henighan, Tom and Child, Rewon and Ramesh, Aditya and Ziegler, Daniel M. and Wu, Jeffrey and W...

    work page 2020

  7. [7]

    Benchmarking Zero-shot Text Classification: Datasets, Evaluation and Entailment Approach

    Yin, Wenpeng and Hay, Jamaal and Roth, Dan. Benchmarking Zero-shot Text Classification: Datasets, Evaluation and Entailment Approach. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019. doi:10.18653/v1/D19-1404

    work page doi:10.18653/v1/d19-1404 2019

  8. [8]

    International Conference on Learning Representations , year=

    Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year=

    work page

  9. [9]

    Data Augmentation using Pre-trained Transformer Models

    Kumar, Varun and Choudhary, Ashutosh and Cho, Eunah. Data Augmentation using Pre-trained Transformer Models. Proceedings of the 2nd Workshop on Life-long Learning for Spoken Language Systems. 2020. doi:10.18653/v1/2020.lifelongnlp-1.3

    work page doi:10.18653/v1/2020.lifelongnlp-1.3 2020

  10. [10]

    Wei, Jason and Zou, Kai , editor =. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , month = nov, year =. doi:10.18653/v1/D19-1670 , url =

    work page doi:10.18653/v1/d19-1670 2019

  11. [11]

    Do Not Have Enough Data? Deep Learning to the Rescue!

    Anaby-Tavor, Ateret and Carmeli, Boaz and Goldbraich, Esther and Kantor, Amir and Kour, George and Shlomov, Segev and Tepper, Naama and Zwerdling, Naama. Do Not Have Enough Data? Deep Learning to the Rescue!. Proceedings of the AAAI Conference on Artificial Intelligence. 2020. doi:10.1609/aaai.v34i05.6233

    work page doi:10.1609/aaai.v34i05.6233 2020

  12. [12]

    BERT: Pre-training of deep bidirectional transformers for language understanding

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , booktitle =. 2019 , address =. doi:10.18653/v1/N19-1423 , url =

    work page doi:10.18653/v1/n19-1423 2019

  13. [13]

    A computational approach to understanding empathy expressed in text-based mental health support

    A Computational Approach to Understanding Empathy Expressed in Text-Based Mental Health Support , author =. Proceedings of EMNLP 2020 , year =. doi:10.18653/v1/2020.emnlp-main.425 , url =

    work page doi:10.18653/v1/2020.emnlp-main.425 2020

  14. [14]

    Towards emotional support dialog systems

    Towards Emotional Support Dialog Systems , author =. Proceedings of ACL-IJCNLP 2021 , year =. doi:10.18653/v1/2021.acl-long.269 , url =

    work page doi:10.18653/v1/2021.acl-long.269 2021

  15. [15]

    Archives of General Psychiatry , year =

    An Empirical Study of Self-Rated Defense Style , author =. Archives of General Psychiatry , year =

    work page

  16. [16]

    M ental BERT : Publicly Available Pretrained Language Models for Mental Healthcare

    Ji, Shaoxiong and Zhang, Tianlin and Ansari, Luna and Fu, Jie and Tiwari, Prayag and Cambria, Erik. M ental BERT : Publicly Available Pretrained Language Models for Mental Healthcare. Proceedings of the Thirteenth Language Resources and Evaluation Conference. 2022

    work page 2022

  17. [17]

    Christopher , journal =

    Di Giuseppe, Mariagrazia and Perry, J. Christopher , journal =. The Hierarchy of Defense Mechanisms: Assessing Defensive Functioning with the Defense Mechanisms Rating Scales. 2021 , volume =

    work page 2021

  18. [18]

    Enhancing Low-Resource

    Patwa, Parth and Filice, Simone and Chen, Zhiyu and Castellucci, Giuseppe and Rokhlenko, Oleg and Malmasi, Shervin , booktitle =. Enhancing Low-Resource. 2024 , address =

    work page 2024

  19. [19]

    ACM Computing Surveys , year =

    Survey of Hallucination in Natural Language Generation , author =. ACM Computing Surveys , year =

    work page

  20. [20]

    and Perry, J

    Skodol, Andrew E. and Perry, J. Christopher , journal =. Should an Axis for Defense Mechanisms Be Included in. 1993 , volume =

    work page 1993

  21. [21]

    Jiang, Zhiying and Yang, Matthew and Tsirlin, Mikhail and Tang, Raphael and Dai, Yiqin and Lin, Jimmy , booktitle =. ``. 2023 , address =. doi:10.18653/v1/2023.findings-acl.426 , url =

    work page doi:10.18653/v1/2023.findings-acl.426 2023

  22. [22]

    2023 , eprint =

    Self-Instruct: Aligning Language Models with Self-Generated Instructions , author =. 2023 , eprint =

    work page 2023

  23. [23]

    Journal of Personality , year =

    The Development of Defense Mechanisms , author =. Journal of Personality , year =

    work page

  24. [24]

    Advances in Psychology , year =

    Studying Defense Mechanisms in Psychotherapy Using the Defense Mechanism Rating Scales , author =. Advances in Psychology , year =

    work page

  25. [25]

    Journal of Abnormal Psychology , year =

    Ego Mechanisms of Defense and Personality Psychopathology , author =. Journal of Abnormal Psychology , year =

    work page

  26. [26]

    Overview of the PsyDefDetect Shared Task at BioNLP 2026: Detecting Levels of Psychological Defense Mechanisms in Supportive Conversations

    Na, Hongbin and Wang, Zimu and Chen, Zhaoming and Hua, Yining and Gao, Rena and Yang, Kailai and Chen, Ling and Wang, Wei and Ji, Shaoxiong and Torous, John and Ananiadou, Sophia. Overview of the PsyDefDetect Shared Task at BioNLP 2026: Detecting Levels of Psychological Defense Mechanisms in Supportive Conversations. Proceedings of the 25th Workshop on Bi...

    work page 2026

  27. [27]

    A Survey of Large Language Models in Psychotherapy: Current Landscape and Future Directions

    Na, Hongbin and Hua, Yining and Wang, Zimu and Shen, Tao and Yu, Beibei and Wang, Lilin and Wang, Wei and Torous, John and Chen, Ling. A Survey of Large Language Models in Psychotherapy: Current Landscape and Future Directions. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.385

    work page doi:10.18653/v1/2025.findings-acl.385 2025

  28. [28]

    You Never Know a Person, You Only Know Their Defenses: Detecting Levels of Psychological Defense Mechanisms in Supportive Conversations

    Na, Hongbin and Wang, Zimu and Chen, Zhaoming and Zhou, Peilin and Hua, Yining and Zhou, Grace Ziqi and Zhang, Haiyang and Shen, Tao and Wang, Wei and Torous, John and Ji, Shaoxiong and Chen, Ling. You Never Know a Person, You Only Know Their Defenses: Detecting Levels of Psychological Defense Mechanisms in Supportive Conversations. Findings of the Associ...

    work page 2026

  29. [29]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , pages =

    Transformers: State-of-the-Art Natural Language Processing , author =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , pages =

    work page 2020

  30. [30]

    2024 , url =

    Llama 3 Model Card , author =. 2024 , url =

    work page 2024

  31. [31]

    Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages =

    Data Augmentation for Low-Resource Neural Machine Translation , author =. Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages =

    work page

  32. [32]

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages =

    Understanding Back-Translation at Scale , author =. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages =

    work page 2018

  33. [33]

    Christopher and Kardos, Marianne E

    Perry, J. Christopher and Kardos, Marianne E. and Pagano, Christopher J. , journal =. The Study of Defenses in Psychotherapy Using the Defense Mechanism Rating Scales (. 1993 , publisher =

    work page 1993

  34. [34]

    Advances in Neural Information Processing Systems , volume =

    Attention Is All You Need , author =. Advances in Neural Information Processing Systems , volume =

    work page

  35. [35]

    arXiv preprint arXiv:2410.12896 , year =

    A Survey on Data Synthesis and Augmentation for Large Language Models , author =. arXiv preprint arXiv:2410.12896 , year =

    work page arXiv

  36. [36]

    Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing , pages =

    A Large Annotated Corpus for Learning Natural Language Inference , author =. Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing , pages =

    work page 2015

  37. [37]

    1992 , publisher =

    Ego Mechanisms of Defense: A Guide for Clinicians and Researchers , author =. 1992 , publisher =

    work page 1992

  38. [38]

    and Pennebaker, James W

    Tausczik, Yla R. and Pennebaker, James W. , journal =. The Psychological Meaning of Words:. 2010 , publisher =

    work page 2010

  39. [39]

    , title =

    Jurafsky, Daniel and Martin, James H. , title =. 2026 , url =

    work page 2026

  40. [40]

    ACM Computing Surveys , volume =

    Survey of Hallucination in Natural Language Generation , author =. ACM Computing Surveys , volume =. 2023 , publisher =

    work page 2023