Elamur Research⌕
☰
home / papers / 2026 / Interpretable Semantic Gradients in SSD: A PCA Sweep Approach and a Case Study on AI Discourse

Interpretable Semantic Gradients in SSD: A PCA Sweep Approach and a Case Study on AI Discourse

Paper · 2026 · Hubert Plisiecki, Maria Leniarska, Jan Piotrowski et al.

Summary

Supervised Semantic Differential (SSD) builds a semantic gradient in embedding space by regressing a continuous trait on PCA-compressed text vectors. Until now, users chose the number of retained components (K) by hand, which adds unnecessary analytic flexibility. The authors propose a PCA sweep that fits SSD over a range of K values and picks the smallest K with the best combined score. The score has two parts: cluster interpretability (coherence plus alignment with the gradient, detrended against variance explained) and gradient stability across neighbouring K values. Predictive fit is deliberately left out of the criterion.

They demonstrate the procedure on 349 short posts about AI from Prolific participants who also completed Narcissistic Admiration and Rivalry scales. For Admiration, the sweep produced a significant and interpretable gradient: hopeful, cooperative views of AI at one end and suspicious, mocking ones at the other. Rivalry showed no reliable association. A high-dimensional comparison run produced blurry, poorly structured clusters. The sweep therefore makes SSD analyses more transparent while preserving their qualitative value.

Key points

  • How the sweep works:
    • It tracks three things per K: cumulative explained variance (used only for detrending), a cluster-size-weighted interpretability z-score, and gradient change (1 − cosine between consecutive gradients).
    • Both the interpretability and stability curves are smoothed with a local neighbourhood average (AUCK), so broad plateaus are favoured over isolated peaks.
  • Case study setup:
    • Corpus of N = 349 posts, about 30 words each, preprocessed with spaCy 3.8.7 (en_core_web_sm).
    • Posts were embedded with the 300-dimensional Dolma GloVe model, using SIF weighting (a = 10⁻³) and removal of the top principal component.
    • Whole posts were used instead of a concept lexicon, since all texts answered the same AI-focused prompt.
  • Sweep settings: K from 1 to 119 in steps of 2; cluster number chosen by silhouette within 2–5; 100 nearest neighbours per pole; AUCK radius of 3; median-smoothed gradient change with a window of 7.
  • Results:
    • Admiration: adjusted R² = .19, F = 6.32, p < 10⁻¹⁰.
    • Rivalry: adjusted R² = .03, p = .095, so it was not interpreted.
    • A counterfactual run at K = 120 gave diffuse clusters.
  • Limitations and resources:
    • The authors flag a small survey sample, purely correlational interpretation, and other open pipeline choices such as the embedding model and window size.
    • Code is available at github.com/hplisiecki/acl2026-ssd-pca-sweep.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of WarsawWarsaw · Poland7 works · 496 citations🎓Warsaw University of TechnologyWarsaw · Poland3 works · 256 citations

Also on the byline: VIZJA University.

 Fields, topics and concepts

📚 Shelves 8The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksDecision Sciences · 209 worksGeneral Social Sciences · 227 worksManagement Science and Operations Research · 108 works

🧭 Topics 1Research subjects and languages the work is about.Discourse analysis · 466 works

Cite

Cite as Hubert Plisiecki, Maria Leniarska, Jan Piotrowski and Marcin Zajenkowski. 2026. Interpretable Semantic Gradients in SSD: A PCA Sweep Approach and a Case Study on AI Discourse. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.446

@inproceedings{plisiecki2026interpretable,
  title = {{Interpretable Semantic Gradients in SSD: A PCA Sweep Approach and a Case Study on AI Discourse}},
  author = {Hubert Plisiecki, Maria Leniarska, Jan Piotrowski and Marcin Zajenkowski},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.446},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.446}
}

References

16 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sanjeev Arora, Yingyu Liang, and Tengyu Ma. 2017. A Simple but Tough-to-Beat Baseline for Sentence Embeddings. find
  2. Mitja D. Back. 2018. The Narcissistic Admiration and Rivalry Concept. In Anthony D. Hermann, Amy B. Brunell, and Joshua D. Foster, editors, Handbook of Trait Narcissism: Key Advances, Research Methods, and Controversies, pages 57–67. Springer International Publishing, Cham. find
  3. Mitja D. Back, Albrecht C. P. Küfner, Michael Dufner, Tanja M. Gerlach, John F. Rauthmann, and Jaap J. A. Denissen. 2013. Narcissistic admiration and rivalry: Disentangling the bright and dark sides of narcissism. Journal of Personality and Social Psychology, 105(6):1013–1037. find
  4. Riley Carlson, John Bauer, and Christopher D. Manning. 2025. A New Pair of GloVes. arXiv preprint. Version Number: 1. find
  5. Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. 2018. Word embeddings quantify 100 years of gender and ethnic stereotypes. Proceedings of the National Academy of Sciences, 115(16). find
  6. I. T. Jolliffe. 2002. Principal Component Analysis. Springer Series in Statistics. Springer-Verlag, New York. find
  7. Austin C. Kozlowski, Matt Taddy, and James A. Evans. 2019. The Geometry of Culture: Analyzing the Meanings of Class through Word Embeddings. American Sociological Review, 84(5):905–949. find
  8. Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient Estimation of Word Representations in Vector Space. arXiv preprint. Version Number: 3.
  9. David Mimno, Hanna Wallach, Edmund Talley, Miriam Leenders, and Andrew McCallum. 2011. Optimizing Semantic Coherence in Topic Models. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 262–272, Edinburgh, Scotland, UK. Association for Computational Lingu…
  10. Ines Montani, Matthew Honnibal, Adriane Boyd, Sofie Van Landeghem, and Henning Peters. 2023. explosion/spaCy: v3.7.2: Fixes for APIs and requirements. find
  11. Jiaqi Mu, Suma Bhat, and Pramod Viswanath. 2017. All-but-the-Top: Simple and Effective Postprocessing for Word Representations. arXiv preprint. Version Number: 2. find
  12. Charles Egerton Osgood, George J. Suci, and Percy H. Tannenbaum. 1978. The measurement of meaning. University of Illinois Press, Urbana-Champaign. find
  13. Ethan Perez, Sam Ringer, Kamilė Lukošiūtė, Karina Nguyen, Edwin Chen, Scott Heiner, Craig Pettit, Catherine Olsson, Sandipan Kundu, Saurav Kadavath, Andy Jones, Anna Chen, Ben Mann, Brian Israel, Bryan Seethor, Cameron McKinnon, Christopher Olah, Da Yan, Daniela Amodei, and 44 others. 2022. Disco…
  14. Hubert Plisiecki, Paweł Lenartowicz, Artur Pokropek, Kinga Małyska, and Maria Flakus. 2026. Measuring Individual Differences in Meaning: The Supervised Semantic Differential. find
  15. Joseph P. Simmons, Leif D. Nelson, and Uri Simonsohn. 2011. False-Positive Psychology: Undisclosed Flexibility in Data Collection and Analysis Allows Presenting Anything as Significant. Psychological Science, 22(11):1359–1366. find
  16. Virgil Zeigler-Hill, Jennifer K. Vrabel, Gillian A. Mc-Cabe, Cheryl A. Cosby, Caitlin K. Traeder, Kelsey A. Hobbs, and Ashton C. Southard. 2019. Narcissism and the pursuit of status. Journal of Personality, 87(2):310–327. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/plisiecki-2026-interpretable-semantic-gradients-ssd-pca/