Pith. sign in

REVIEW 2 major objections 5 minor 62 references

A post-hoc energy decoder that scores whole label sets against the Schwartz circular continuum makes value predictions more theory-coherent without sacrificing F1.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 09:32 UTC pith:E4SOLSTN

load-bearing objection Clean controlled result: theory-fixed post-hoc decoding improves Schwartz coherence without F1 loss, and only for the true continuum—not random or empirical controls. the 2 major comments →

arxiv 2607.05052 v1 pith:E4SOLSTN submitted 2026-07-06 cs.CL cs.AIcs.CYcs.LG

Beyond Independent Labels: Schwartz-Geometry Decoding for Human Value Detection

V\'ictor Yeste , Paolo Rosso This is my paper
classification cs.CL cs.AIcs.CYcs.LG
keywords human value detectionSchwartz valuesmulti-label classificationstructured decodingenergy-based modelsoutput-space geometrytheory-aware metricscircular continuum
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Human value detection is usually treated as multi-label classification over 19 independent Schwartz values, even though Schwartz theory arranges those values on a circular continuum of compatible neighbors and conflicting opposites. This paper asks whether that continuum can be used as a soft output-space bias rather than a hard rule. On a fixed DeBERTa classifier, training-time geometry penalties give only limited gains that are no larger for the true continuum than for a random ordering. A post-hoc energy decoder that jointly scores whole label sets, by contrast, improves theory-aware coherence metrics the authors introduce while holding Macro-F1 and Micro-F1 fixed by its selection rule. The coherence gain is specific to the true Schwartz ordering: the same decoder with a random permutation or an empirical co-occurrence graph does not produce it. Prompting a large language model with the continuum shifts behavior but does not match the supervised decoder. The practical upshot is a lightweight, controllable way to make value detectors more faithful to the psychology of their label space without new architectures or accuracy trade-offs.

Core claim

On a fixed DeBERTa multi-label classifier for sentence-level detection of the 19 refined Schwartz values, a post-hoc Schwartz-aware energy decoder produces label sets more coherent with the circular motivational continuum—at no cost to Macro-F1 or Micro-F1—while the same structure injected into the training loss does not. The coherence gain is specific to the true continuum: it does not appear when the identical decoder is run with a random circular permutation or an empirical co-occurrence graph.

What carries the argument

The Schwartz-aware energy decoder: it selects the label set that maximizes a structured score combining the classifier’s per-label log-odds margins with pairwise rewards for co-selecting neighboring values on the continuum and penalties for co-selecting opposing values. The pairwise structure is fixed from theory rather than learned from data; decoder weights are chosen under a Pareto rule that preserves validation Macro-F1 while minimizing a composite geometry cost.

Load-bearing premise

The paper’s measure of “coherence with the continuum” is a hand-built composite of opposite-error rate, neighbor-error rate, and confusion-distance correlation under one particular angular layout and opposite threshold; if that composite does not track real theory faithfulness, the claimed gain dissolves.

What would settle it

If the identical energy decoder, under the same Pareto Macro-F1 constraint, produced equal or larger reductions in geometry cost when given a random circular permutation or an empirical co-occurrence graph as when given the true Schwartz order—on the same test sentences and seeds—the claim that the true continuum specifically drives the coherence gain would be falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Value detection systems can become more faithful to Schwartz theory by adding a tunable decoding step rather than redesigning the classifier or hierarchy.
  • Training-time geometry penalties are not a reliable way to inject this structure; the benefit appears at the final discrete decision step.
  • Control comparisons with random order and co-occurrence graphs are necessary to show that gains come from the theory itself, not from any structure.
  • Prompting large language models with the continuum is not a substitute for supervised structured decoding on this task.
  • Theory-aware coherence metrics give a measurable notion of faithfulness beyond standard F1.

Where Pith is reading between the lines

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

  • The same post-hoc energy-decoding pattern could transfer to other circular or circumplex label spaces, such as emotion wheels, without retraining the base classifier.
  • Soft theory-derived pairwise terms may be preferable to hard hierarchical gates whenever real data legitimately expresses conflict that a hard constraint would forbid.
  • Because the decoder only reranks candidates the base model already surfaces, its gains will be limited by the classifier’s recall; better local evidence could compound them.
  • The Pareto selection rule itself is a general template for imposing soft structure at decode time without F1 cost.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper formulates sentence-level multi-label detection of the 19 refined Schwartz values as a circular output-space geometry (Eq. 1) and tests whether that geometry can serve as a soft inductive bias. On a fixed DeBERTa-v3-base classifier, training-time geometry-aware losses (GeoLoss, GeoSmooth) yield only limited, non-theory-specific gains, while a post-hoc energy decoder (Eq. 2) that rewards neighbor co-selection and penalizes opposite co-selection improves author-defined label-set coherence metrics at no cost to Macro-F1 or Micro-F1 (held fixed by a Pareto selection rule). The coherence gain is specific to the true continuum: it does not appear for a random circular permutation or an empirical co-occurrence graph through the identical decoder (significant in all five seeds). A bounded Qwen2.5-72B-Instruct diagnostic shows that prompting the continuum shifts geometry-aware behavior but does not match the supervised decoder.

Significance. If the controlled result holds, the paper supplies a lightweight, post-hoc recipe for making multi-label value detectors more faithful to a psychologically motivated label space without sacrificing standard F1. Strengths include five seeds, document-level splits, validation-only hyperparameter and threshold selection, paired bootstrap tests, and direct Schwartz-vs-random/empirical controls that isolate the true ordering. The explicit admission that F1 preservation is enforced by the Pareto rule rather than discovered, the null training-time result, and the LLM diagnostic further bound overclaim. The contribution is modest in absolute accuracy terms but methodologically clean for theory-aware structured prediction on this task family.

major comments (2)
  1. Section 4.3 Pareto rule and Section 5 decoder geometry cost: the decoder is selected to minimize a composite built from the same distance matrix D that supplies its pairwise weights N and O, under a Macro-F1 floor. This is partial circularity by construction. The random and empirical controls (Table 3, Section 6.3, Appendix C Table 8) adequately bound the claim as stated—only the true continuum lowers the cost through the identical decoder—but the manuscript should state more prominently in the abstract and introduction that the primary evidence is the controlled contrast, not the absolute reduction against thresholding, and should report the three component metrics separately as primary results rather than only the composite sum.
  2. Section 5 metrics and Section 8 limitations: the operationalization (equal angular spacing θ_k = 2πk/19, opposite threshold D > 0.75, two-step neighbor window) is one reasonable choice among several. Because the neighbor-error rate rises slightly under the Schwartz decoder (0.191 → 0.195) while opposite-error and confusion–distance improve, a sensitivity check over the opposite threshold and neighbor window (or an ablation that drops the neighbor term) would strengthen the claim that the gain is not an artifact of these cut-offs. The limitations section already flags the issue; a short appendix table would make it load-bearing rather than parenthetical.
minor comments (5)
  1. Table 2 vs. Table 3: theory-aware columns differ by design (expected circular error for probability models; opposite-error / geometry cost for discrete sets). A one-sentence reminder in the table captions would prevent readers from treating the columns as directly comparable.
  2. Section 6.2: the claim that the decoder changes only 2.45% of test sentences is useful; reporting the corresponding change in per-label precision/recall for the most frequent opposite pairs would make the surgical character of the edits more concrete.
  3. Figure 1 and Eq. (1): the figure is clear, but the text never states whether the canonical order is fixed from Schwartz et al. (2012) or re-derived; a citation to the exact source ordering would help replication.
  4. Appendix D Table 9: the qualitative examples are helpful; indicating which seed each row comes from would align them with the five-seed protocol used elsewhere.
  5. Typos / polish: abstract line 'no larger for the true continuum than for a random ordering' is slightly ambiguous (gains no larger vs. gains that are no larger); Section 7 'the neighbor-error rate needs care' is informal for a journal register.

Circularity Check

2 steps flagged

Decoder Pareto selection minimizes author-defined geometry cost built from the same D that supplies pairwise weights, making F1-preserving coherence gain true by construction; random/empirical controls bound the comparative claim.

specific steps
  1. fitted input called prediction [Section 4.3 (Pareto selection rule) and Section 5 (decoder geometry cost definition)]
    "The weights(α, β, γ)are tuned on validation under a Pareto criterion: among settings that retain Macro-F1 within a small tolerance of the best validation Macro-F1, we select the one that minimizes a validation geometry cost (a label-set coherence measure defined in Section 5). This is, by construction, the reason the decoder improves theory-aware coherence while preserving F1: F1 preservation is enforced by the Pareto constraint rather than discovered empirically."

    Hyperparameters of the energy decoder (whose pairwise term is built from the same circular distance matrix D that defines neighbor/opposite relations) are chosen precisely to minimize the author-defined composite geometry cost (opposite-error rate + neighbor-error rate + confusion–distance correlation, all derived from D) subject to a Macro-F1 floor. Reporting lower geometry cost at essentially unchanged F1 is therefore true by the selection rule itself, not an independent empirical discovery of coherence. The paper acknowledges the construction; the comparative controls (identical decoder + cost with random/empirical D) prevent the circularity from fully determining the strongest claim.

  2. self definitional [Section 5 Metrics (decoder geometry cost) and Abstract/Section 6.2]
    "Thedecoder geometry costsums these three—two rates in[0,1]and a correlation in[−1,1], equally weighted—so a lower value is more coherent. ... the decoder makes label sets more coherent with the continuum—on theory-aware coherence metrics we introduce—at no cost to Macro-F1 or Micro-F1 (held fixed by its selection rule)."

    The primary success metric for the decoder (geometry cost) is defined by the authors from exactly the same circular geometry D that supplies the decoder's N and O matrices. Minimizing that metric under an F1 constraint therefore tautologically produces 'more coherent' sets with respect to the metric's own definition of coherence. This is not hidden, but it means the absolute improvement on the introduced metrics is definitional once selection occurs; only the differential vs. control geometries supplies non-circular content.

full rationale

The paper is largely self-contained and transparent: it explicitly states that F1 preservation is enforced by the Pareto rule rather than discovered, introduces the composite geometry cost as its own theory-aware metric family derived from D, and tests specificity via identical-decoder controls (random circular permutation and empirical co-occurrence) that share the same cost machinery, candidate pool, and selection rule. Only the true Schwartz ordering lowers the cost (significant in all five seeds; Table 3, Section 6.3, Appendix C). Training-time GeoLoss/GeoSmooth results and the LLM diagnostic are independent of this loop and show null or weaker effects. The residual circularity is therefore partial and local to the decoder's selection objective (not a self-citation chain, uniqueness import, or renaming of a known result). It does not collapse the central comparative claim, which remains falsifiable by the controls. Score 4 reflects that the reported coherence improvement under fixed F1 is definitionally selected rather than independently measured, while the decisive evidence (true vs. control geometries) is not.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 3 invented entities

The central claim rests on the Schwartz continuum being a valid soft inductive bias for presence (not stance), on an equal-angle circular embedding of 19 values, on author-chosen neighbor/opposite cut-offs, and on a composite coherence cost that the decoder is explicitly selected to minimize. Free parameters are the decoder weights and selection tolerances plus training hyperparameters. Invented entities are the energy decoder formulation specialized to theory-derived N/O matrices and the composite decoder geometry cost. No new physical or psychological entities are postulated beyond operationalizing an existing theory.

free parameters (8)
  • decoder neighbor weight α = 0.1 (final active)
    Fixed component magnitude α=0.1 (with β=0.2) then selected under Pareto rule on validation; controls strength of neighbor reward.
  • decoder opposite weight β = 0.2 (final active)
    Fixed component magnitude β=0.2; controls strength of opposite penalty; active in final runs.
  • decoder cardinality weight γ = 0 (final)
    Tuned under Pareto; validation sets γ=0 so cardinality term is inactive in reported runs.
  • opposite distance threshold = 0.75
    Pairs with D>0.75 treated as opposing; author choice among possible cut-offs.
  • neighbor window = 2 steps
    Two-step circular neighborhood for compatibility matrix N; author choice.
  • Pareto Macro-F1 tolerance = 99% of best val Macro-F1
    Retain validation Macro-F1 within 99% of thresholding before minimizing geometry cost; enforces F1 preservation by construction.
  • per-label thresholds τ_k = per-label, validation-tuned
    Tuned on validation by 0.01 grid to maximize each label's F1; enter unary term of decoder.
  • GeoLoss λ / GeoSmooth τ = grid-searched
    Training-time geometry hyperparameters searched on validation; secondary to main decoder claim.
axioms (5)
  • domain assumption The 19 refined Schwartz values form a circular motivational continuum in which adjacent values are compatible and opposing values are in tension; this structure is a valid soft inductive bias for sentence-level presence prediction.
    Stated in Introduction and Section 3.2; drawn from Schwartz et al. (2012, 2017) and treated as soft rather than hard constraint.
  • ad hoc to paper Equal angular spacing θ_k = 2πk/19 and shorter-arc circular distance (Eq. 1) adequately operationalize the continuum for output-space geometry.
    Section 3.2; one reasonable choice among several, flagged in Limitations.
  • domain assumption Collapsing attained and constrained stance annotations into a single presence indicator preserves the geometry of value presence that the paper studies.
    Section 3.1; modeling choice that discards stance polarity.
  • ad hoc to paper Standard multi-label metrics (Macro-F1, Micro-F1, Macro-AUPRC) plus author-defined circular-error and label-set coherence metrics are appropriate evaluation targets.
    Section 5; composite geometry cost is introduced in this paper.
  • standard math Binary cross-entropy with independent per-label thresholding is a fair strong baseline against which geometry injection should be measured.
    Section 4.1; standard multi-label practice.
invented entities (3)
  • Schwartz-aware energy decoder (Eq. 2 with theory-derived N, O) no independent evidence
    purpose: Post-hoc structured scoring of whole label sets that rewards neighbor co-selection and penalizes opposite co-selection while generalizing independent thresholding when α=β=γ=0.
    Main method (Section 4.3); pairwise structure fixed from theory rather than learned from data.
  • Decoder geometry cost (composite of opposite-error rate, neighbor-error rate, confusion–distance correlation) no independent evidence
    purpose: Corpus-level theory-aware coherence metric used both for Pareto selection and as primary evidence of improved faithfulness.
    Introduced in Section 5; dimensionally heterogeneous composite the authors themselves define.
  • Expected circular error (gold-normalized distance-weighted probability mass) no independent evidence
    purpose: Probability-space theory-aware metric for supervised models, matching the GeoLoss penalty.
    Section 5; derived from D for continuous outputs.

pith-pipeline@v1.1.0-grok45 · 23933 in / 4082 out tokens · 32296 ms · 2026-07-11T09:32:40.984749+00:00 · methodology

0 comments
read the original abstract

Human value detection is commonly formulated as sentence-level multi-label classification over the 19 refined Schwartz values, typically predicted as independent labels. Schwartz theory, however, describes them as a circular motivational continuum, in which adjacent values are compatible and opposing values are in tension. We ask whether this structure can be operationalized as an explicit output-space geometry and used as a soft bias rather than a hard constraint. On a DeBERTa-v3-base classifier, we compare two ways of injecting it: training-time geometry-aware objectives and a post-hoc Schwartz-aware energy decoder that scores whole label sets jointly. Across five seeds, training-time geometry gives only limited gains-no larger for the true continuum than for a random ordering-whereas the decoder makes label sets more coherent with the continuum-on theory-aware coherence metrics we introduce-at no cost to Macro-F1 or Micro-F1 (held fixed by its selection rule). The gain is specific to the true Schwartz ordering: it does not appear for a random permutation or an empirical co-occurrence graph through the identical decoder. A bounded Qwen2.5-72B-Instruct diagnostic shows that supplying the continuum at inference shifts behavior but does not match supervised structured prediction. Theory-aware decoding thus offers a lightweight, controllable way to make value detection faithful to its label space.

Figures

Figures reproduced from arXiv: 2607.05052 by Paolo Rosso, V\'ictor Yeste.

Figure 1
Figure 1. Figure 1: The refined Schwartz value continuum used as the output-space geometry. The 19 values are placed [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

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

62 extracted references · 2 canonical work pages

  1. [1]

    Journal of Personality and Social Psychology , volume=

    Refining the theory of basic individual values , author=. Journal of Personality and Social Psychology , volume=. 2012 , publisher=

    2012

  2. [2]

    Values and Behavior: Taking a Cross Cultural Perspective , editor=

    The refined theory of basic values , author=. Values and Behavior: Taking a Cross Cultural Perspective , editor=. 2017 , publisher=

    2017

  3. [3]

    and Cieciuch, Jan , journal=

    Schwartz, Shalom H. and Cieciuch, Jan , journal=. Measuring the refined theory of individual values in 49 cultural groups: Psychometrics of the revised. 2022 , publisher=

    2022

  4. [4]

    2023 , address=

    Kiesel, Johannes and Alshomary, Milad and Mirzakhmedova, Nailia and Heinrich, Maximilian and Handke, Nicolas and Wachsmuth, Henning and Stein, Benno , booktitle=. 2023 , address=

    2023

  5. [5]

    Mirzakhmedova, Nailia and Kiesel, Johannes and Alshomary, Milad and Heinrich, Maximilian and Handke, Nicolas and Cai, Xiaoni and Barriere, Valentin and Wachsmuth, Henning and Stein, Benno , booktitle=. The. 2024 , address=

    2024

  6. [6]

    Machine Learning , volume=

    Classifier chains for multi-label classification , author=. Machine Learning , volume=. 2011 , publisher=

    2011

  7. [7]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    A review on multi-label learning algorithms , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2014 , doi=

    2014

  8. [8]

    Applied Sciences , volume=

    Application of label correlation in multi-label classification: A survey , author=. Applied Sciences , volume=. 2024 , doi=

    2024

  9. [9]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    Asymmetric loss for multi-label classification , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

  10. [10]

    2023 , url=

    He, Pengcheng and Gao, Jianfeng and Chen, Weizhu , booktitle=. 2023 , url=

    2023

  11. [11]

    Human values in a single sentence: Moral presence, hierarchies, and transformer ensembles on the

    Yeste, Victor and Rosso, Paolo , journal=. Human values in a single sentence: Moral presence, hierarchies, and transformer ensembles on the. 2026 , eprint=

    2026

  12. [12]

    Yeste, Victor and Rosso, Paolo , journal=. Do. 2026 , eprint=

    2026

  13. [13]

    More context, larger models, or moral knowledge? A systematic study of

    Yeste, Victor and Rosso, Paolo , journal=. More context, larger models, or moral knowledge? A systematic study of. 2026 , eprint=

    2026

  14. [14]

    arXiv preprint arXiv:2403.09720 , year=

    Fine-tuning vs prompting, can language models understand human values? , author=. arXiv preprint arXiv:2403.09720 , year=. 2403.09720 , archivePrefix=

    Pith/arXiv arXiv

  15. [15]

    2023 , publisher=

    Ma, Long and Sun, Zeye and Jiang, Jiawei and Li, Xuan , booktitle=. 2023 , publisher=

    2023

  16. [16]

    , booktitle=

    Molazadeh Oskuee, Milad and Rahgouy, Mostafa and Babaei Giglou, Hamed and Seals, Cheryl D. , booktitle=. 2023 , publisher=

    2023

  17. [17]

    Tenzin-Gyatso at

    Kandru, Pavan and Singh, Bhavyajeet and Maity, Ankita and Aditya Hari, Kancharla and Varma, Vasudeva , booktitle=. Tenzin-Gyatso at. 2023 , publisher=

    2023

  18. [18]

    Hitachi at

    Tsunokake, Masaya and Yamaguchi, Atsuki and Koreeda, Yuta and Ozaki, Hiroaki and Sogawa, Yasuhiro , booktitle=. Hitachi at. 2023 , publisher=

    2023

  19. [19]

    Edward Said at Touch

    Aydin, Aisha Nur and Shaar, Shaden and Cardie, Claire , booktitle=. Edward Said at Touch

  20. [20]

    arXiv preprint arXiv:2401.16549 , year=

    Deep learning for multi-label learning: A comprehensive survey , author=. arXiv preprint arXiv:2401.16549 , year=

    Pith/arXiv arXiv

  21. [21]

    PAKT: Perspectivized Argumentation Knowledge Graph and Tool for Deliberation Analysis

    Plenz, Moritz and Heinisch, Philipp and Frank, Anette and Cimiano, Philipp. PAKT: Perspectivized Argumentation Knowledge Graph and Tool for Deliberation Analysis. Robust Argumentation Machines. 2024

    2024

  22. [22]

    Journal of Cross-Cultural Psychology , volume=

    Applying the refined values theory to past data: What can researchers gain? , author=. Journal of Cross-Cultural Psychology , volume=. 2013 , doi=

    2013

  23. [23]

    European Journal of Social Psychology , volume=

    Value tradeoffs and behavior in five countries: Validating 19 refined values , author=. European Journal of Social Psychology , volume=. 2017 , doi=

    2017

  24. [24]

    Journal of Research in Personality , volume=

    Testing the circular structure and importance hierarchy of value states in real-time behaviors , author=. Journal of Research in Personality , volume=. 2018 , doi=

    2018

  25. [25]

    2025 , isbn =

    Zhu, Wenhao and Xie, Yuhang and Song, Guojie and Zhang, Xin , title =. 2025 , isbn =. doi:10.24963/ijcai.2025/934 , booktitle =

    work page doi:10.24963/ijcai.2025/934 2025

  26. [26]

    Measuring human value expression in social media texts: Calibrated

    Milkova, Maria and Rudnev, Maksim , journal=. Measuring human value expression in social media texts: Calibrated

  27. [27]

    Rethinking the

    Szegedy, Christian and Vanhoucke, Vincent and Ioffe, Sergey and Shlens, Jon and Wojna, Zbigniew , booktitle=. Rethinking the

  28. [28]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Soft labels for ordinal regression , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  29. [29]

    Predicting Structured Data , publisher=

    A tutorial on energy-based learning , author=. Predicting Structured Data , publisher=

  30. [30]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Identifying the human values behind arguments , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2022 , address=

    2022

  31. [31]

    Journal of Personality and Social Psychology , volume=

    Liberals and conservatives rely on different sets of moral foundations , author=. Journal of Personality and Social Psychology , volume=. 2009 , doi=

    2009

  32. [32]

    and Chin, Jenna and Leong, Christian and Leung, Jun Yen and Mirinjian, Arineh and Dehghani, Morteza , journal=

    Hoover, Joe and Portillo-Wightman, Gwenyth and Yeh, Leigh and Havaldar, Shreya and Davani, Aida Mostafazadeh and Lin, Ying and Kennedy, Brendan and Atari, Mohammad and Kamel, Zahra and Mendlen, Madelyn and Moreno, Gabriela and Park, Christina and Chang, Tingyee E. and Chin, Jenna and Leong, Christian and Leung, Jun Yen and Mirinjian, Arineh and Dehghani, ...

    2020

  33. [33]

    Journal of Personality and Social Psychology , volume=

    A circumplex model of affect , author=. Journal of Personality and Social Psychology , volume=. 1980 , doi=

    1980

  34. [34]

    Proceedings of the Eighteenth International Conference on Machine Learning (ICML) , pages=

    Conditional random fields: Probabilistic models for segmenting and labeling sequence data , author=. Proceedings of the Eighteenth International Conference on Machine Learning (ICML) , pages=

  35. [35]

    Proceedings of the 33rd International Conference on Machine Learning (ICML) , pages=

    Structured prediction energy networks , author=. Proceedings of the 33rd International Conference on Machine Learning (ICML) , pages=

  36. [36]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    When does label smoothing help? , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  37. [37]

    International Conference on Learning Representations (ICLR) , year=

    Decoupled weight decay regularization , author=. International Conference on Learning Representations (ICLR) , year=

  38. [38]

    Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    The hitchhiker's guide to testing statistical significance in natural language processing , author=. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2018 , address=

    2018

  39. [39]

    2025 , primaryClass=

    Qwen2.5 Technical Report , author=. 2025 , primaryClass=

    2025

  40. [40]

    Advances in Experimental Social Psychology , editor=

    Universals in the content and structure of values: Theoretical advances and empirical tests in 20 countries , author=. Advances in Experimental Social Psychology , editor=. 1992 , publisher=

    1992

  41. [41]

    Journal of Research in Personality , volume=

    Evaluating the structure of human values with confirmatory factor analysis , author=. Journal of Research in Personality , volume=. 2004 , doi=

    2004

  42. [42]

    Personality and Social Psychology Bulletin , volume=

    Values and behavior: Strength and structure of relations , author=. Personality and Social Psychology Bulletin , volume=. 2003 , doi=

    2003

  43. [43]

    International Journal of Data Warehousing and Mining , volume=

    Multi-label classification: An overview , author=. International Journal of Data Warehousing and Mining , volume=. 2007 , doi=

    2007

  44. [44]

    Proceedings of the 14th ACM International Conference on Information and Knowledge Management (CIKM) , pages=

    Collective multi-label classification , author=. Proceedings of the 14th ACM International Conference on Information and Knowledge Management (CIKM) , pages=. 2005 , doi=

    2005

  45. [45]

    2018 , url=

    Yang, Pengcheng and Sun, Xu and Li, Wei and Ma, Shuming and Wu, Wei and Wang, Houfeng , booktitle=. 2018 , url=

    2018

  46. [46]

    Proceedings of the 34th International Conference on Machine Learning (ICML) , pages=

    On calibration of modern neural networks , author=. Proceedings of the 34th International Conference on Machine Learning (ICML) , pages=

  47. [47]

    arXiv preprint arXiv:1701.06548 , year=

    Regularizing neural networks by penalizing confident output distributions , author=. arXiv preprint arXiv:1701.06548 , year=. 1701.06548 , archivePrefix=

    Pith/arXiv arXiv

  48. [48]

    2020 , url=

    Demszky, Dorottya and Movshovitz-Attias, Dana and Ko, Jeongwoo and Cowen, Alan and Nemade, Gaurav and Ravi, Sujith , booktitle=. 2020 , url=

    2020

  49. [49]

    Emotion: Theory, Research, and Experience , editor=

    A general psychoevolutionary theory of emotion , author=. Emotion: Theory, Research, and Experience , editor=. 1980 , publisher=

    1980

  50. [50]

    Computational Linguistics , volume=

    Argument mining: A survey , author=. Computational Linguistics , volume=. 2019 , doi=

    2019

  51. [51]

    A survey on moral foundation theory and pre-trained language models: current advances and challenges

    Zangari, Lorenzo and Greco, Candida Maria and Picca, Davide and Tagarelli, Andrea. A survey on moral foundation theory and pre-trained language models: current advances and challenges. AI & SOCIETY. 2025. doi:10.1007/s00146-025-02225-w

    work page doi:10.1007/s00146-025-02225-w 2025

  52. [52]

    Cultural value alignment in large language models: A prompt-based analysis of

    Segerer, Robin , journal=. Cultural value alignment in large language models: A prompt-based analysis of

  53. [53]

    , booktitle=

    Rodrigues, Diego Dimer and Recamonde-Mendoza, Mariana and Moreira, Viviane P. , booktitle=. Beyond single models: Leveraging. 2024 , publisher=

    2024

  54. [54]

    de la Cruz Fern. Value. European Conference on Artificial Intelligence (ECAI) , pages=. 2025 , doi=

    2025

  55. [55]

    Computers in Human Behavior Reports , volume=

    Large language models meet moral values: A comprehensive assessment of moral abilities , author=. Computers in Human Behavior Reports , volume=. 2025 , issn=

    2025

  56. [56]

    arXiv preprint arXiv:2504.15236 , year=

    Values in the wild: Discovering and analyzing values in real-world language model interactions , author=. arXiv preprint arXiv:2504.15236 , year=

    Pith/arXiv arXiv

  57. [57]

    arXiv preprint arXiv:2601.03481 , year=

    Self-explaining hate speech detection with moral rationales , author=. arXiv preprint arXiv:2601.03481 , year=

    arXiv

  58. [58]

    Information Processing & Management , volume=

    Moralization-aware identity fusion for detecting violent radicalization in social media , author=. Information Processing & Management , volume=. 2026 , issn =

    2026

  59. [59]

    International Journal of Computer Vision , volume=

    The earth mover's distance as a metric for image retrieval , author=. International Journal of Computer Vision , volume=. 2000 , doi=

    2000

  60. [60]

    arXiv preprint arXiv:1611.05916 , year=

    Squared Earth Mover's Distance-based Loss for Training Deep Neural Networks , author=. arXiv preprint arXiv:1611.05916 , year=. 1611.05916 , archivePrefix=

    Pith/arXiv arXiv

  61. [61]

    Data Mining and Knowledge Discovery , volume=

    Evaluation measures for hierarchical classification: a unified view and novel approaches , author=. Data Mining and Knowledge Discovery , volume=. 2015 , doi=

    2015

  62. [62]

    Proceedings of the International AAAI Conference on Web and Social Media , volume=

    Whose Values? Measuring the (Subjective) Expression of Basic Human Values in Social Media Posts , author=. Proceedings of the International AAAI Conference on Web and Social Media , volume=. 2026 , doi=

    2026