Pith. sign in

REVIEW 5 major objections 5 minor 44 references

Affective-CARA: A Knowledge Graph Driven Framework for Culturally Adaptive Emotional Intelligence in HCI

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that a conversational agent can adapt its emotional responses to the user's culture in real time by combining a hyperbolic knowledge graph of cultural emotion norms with reinforcement learning and a mediator module…

desk verdict A coherent architecture for culture-adaptive affective response whose headline numbers are graph statistics, not response measurements. read the letter →

arxiv 2506.14166 v1 pith:U5G3IRBP submitted 2025-06-17 cs.HC

classification cs.HC
keywords affectivecomputingculturaladaptationknowledgegraphhyperbolicembeddingsreinforcementlearningfromhumanfeedbackhuman-computerinteractionemotionrecognitionbiasmitigation
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 claims that a conversational AI can learn to adjust its emotional responses to the user's culture in real time, rather than relying on a universal emotion model or static cultural profiles. To show this, it builds Affective-CARA, a framework that combines a knowledge graph of culture–emotion norms (enriched with valence, arousal, dominance scores and embedded in hyperbolic space) with a reinforcement learning loop that uses cultural appropriateness, emotional coherence, and user feedback as rewards. The reported results—a cultural semantic density of 9.32/10, a 61% reduction in cultural representation bias (KL divergence 0.28), and a 39% improvement in emotional appropriateness over standard RLHF—would matter because current systems degrade by roughly 20–40% when deployed outside Western contexts. If true, the framework points a way toward inclusive mental health support, education, and customer service where the same words can mean very different things to different users.

What carries the argument

The central object is the Cultural Emotion Knowledge Graph (CEKG): entities for cultures, emotion prototypes, cultural–emotional mappings, and context indicators, connected by relations such as has_emotion and expressed_as, enriched with Valence–Arousal–Dominance (VAD) scores, and embedded in the Lorentz model of hyperbolic geometry so that hierarchical cultural–emotional structure is preserved. This graph feeds a GRPO (gradient-based reward policy optimization) loop whose reward sums cultural similarity, VAD coherence, and explicit user feedback, and a Cultural-Aware Response Mediator (CARM) that decides when to retrieve knowledge, when to use history, and when to generate; together they produce the culturally adapted narrative.

What would settle it

A head-to-head human evaluation with raters from the five cultural clusters, blind to system identity, comparing Affective-CARA's outputs against a baseline that only paraphrases culture-neutral text with surface-level markers: if raters do not reliably prefer Affective-CARA for cultural appropriateness, or if the CSD score does not correlate with rater preference, the metric and the framework's advantage collapse. Additionally, if KL divergence is recomputed against a demographic-representative target (instead of uniform) and the framework no longer shows reduced bias, the debiasing claim is unsupported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that cultural emotional intelligence can be made an explicit, learnable component of a dialogue system: the authors build a Cultural Emotion Knowledge Graph (CEKG) that encodes culture–emotion norms in hyperbolic space with VAD dimensions, and they use a gradient-based reward policy optimization (GRPO) to tune responses so that they are simultaneously culturally aligned, emotionally coherent, and responsive to user feedback. The framework—coordinated by a Cultural-Aware Response Mediator (CARM)—is reported to achieve a cultural semantic density of 9.32/10, cut cultural representation bias by 61% (KL divergence 0.28), improve emotional appropriateness by 39% over standard RLHF, and maintain near-uniform F1 across five cultural clusters on the MERD corpus, outperforming a general large language model and knowledge-enhanced baselines.

Load-bearing premise

The evaluation assumes that the paper's Cultural Semantic Density metric (which counts whether the knowledge graph entails a cultural concept in the response) and the choice of a uniform ideal culture distribution really capture 'cultural nuance' and 'bias'; if either is wrong, the headline CSD and KL numbers do not support the claimed superiority.

Editorial extensions

If this is right

  • If the framework works as claimed, a single agent can tailor its empathy, enthusiasm, and self-disclosure to the user's cultural background without retraining per culture.
  • The hyperbolic embedding and VAD annotations could be reused as a plug-in knowledge layer for any LLM-based dialogue system, not just the authors' generator.
  • The closed-loop feedback (GRPO with cultural rewards) offers a template for debiasing other affective systems: measure representation bias, correct it via reward, and verify by KL divergence.
  • The reported 87% 'culturally appropriate and emotionally resonant' rating in simulated scenarios suggests that real-time cultural adaptation is achievable in textual dialogue, opening a path to multimodal extension.
  • The paper's own ablation studies indicate that hyperbolic embeddings, GRPO, and CARM each contribute independently; removing CARM more than doubles latency, so the orchestration is a practical necessity, not a detail.

Reading between the lines

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

  • An implicit consequence not drawn by the paper is that the knowledge graph could be swapped for a different cultural ontology (e.g., one built from social surveys) and the GRPO reward reweighted, suggesting the framework is a meta-architecture for cultural adaptation rather than a fixed model.
  • The uniform ideal distribution is one specific debiasing choice; recomputing KL divergence against a demographic-representative target would change the headline numbers without changing the underlying system, so the 61% reduction is conditional on that normative assumption.
  • Because the evaluation is text-only, the claimed 'real-time' adaptation might transfer to multimodal inputs (facial expression, tone) but this is not demonstrated; an immediate test is to run the same mediator on audio-visual affective signals.
  • The paper's own limitation list (low-resource cultures, mixed emotions) defines the next benchmarks: the framework should be stressed on cultures absent from the graph to see whether the reported near-uniform performance holds.
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

5 major / 5 minor

Summary. The paper proposes Affective-CARA, an agentic framework for culturally adaptive affective response generation, combining a hyperbolic Cultural Emotion Knowledge Graph (CEKG) built from StereoKG with VAD annotations, a GRPO-based reinforcement learning policy, and a Cultural-Aware Response Mediator (CARM). The authors claim state-of-the-art performance on cultural adaptation, reporting a Cultural Semantic Density (CSD) of 9.32/10, a 61% reduction in cultural representation bias (KL divergence from 0.72 to 0.28), a 39% improvement in emotional appropriateness, and strong cross-cultural F1 scores on MERD. The central claim is that Affective-CARA consistently outperforms baselines by generating culturally nuanced, emotionally appropriate responses in real time.

Significance. If the reported results held, Affective-CARA would represent a meaningful step toward culturally adaptive affective computing, addressing an acknowledged gap in emotionally intelligent HCI systems. The architecture is ambitious and the combination of a cultural knowledge graph, hyperbolic embeddings, and RL-based adaptation is worth investigating. However, the current evaluation does not support the headline claims: the primary metrics are computed on the knowledge graph rather than on system-generated responses, and the bias-reduction result is largely a re-statement of the graph-balancing procedure. The paper also lacks a clear, reproducible human-evaluation protocol. With a revised evaluation that measures properties of the generated outputs, the underlying framework could be valuable, but as written the evidence is insufficient.

major comments (5)
  1. [§3.10, Eq. (12)] The Cultural Semantic Density in Eq. (12) is defined as a sum over knowledge-graph entailments divided by log(|C|+1); it contains no term involving the system's generated response, so it measures knowledge-graph coverage rather than response quality. Consequently, the CSD values in Table 1 (and the headline '9.32/10') cannot support the claim that Affective-CARA generates culturally richer responses; any system retrieving from the same KG would receive identical CSD. This metric mismatch undermines the central superiority claim stated in the abstract and Section 6.
  2. [§3.3.2 and §3.10, Eq. (13)] The KL-bias result is largely circular. Eq. (13) defines bias as KL(P||P_ideal) with P_ideal uniform, while Section 3.3.2 states that the cross-cultural alignment process 'balances cultural representation across the graph' toward uniform coverage. The reported reduction from 0.72 to 0.28 (61%) therefore re-states the balancing procedure rather than measuring the bias of generated responses. Additionally, the paper does not justify why a uniform distribution is the correct unbiased target for cultural representation.
  3. [§4.4 and §3.10] The emotional-appropriateness evaluation is not presented with sufficient rigor. Section 3.10 refers to 'human evaluations' and 'human judgments', but Section 4.4 describes a 'simulated interactive evaluation' that reports percentages (87% vs 52%) without specifying the number of raters, the number of scenarios, the rating protocol, or inter-rater reliability. Table 1 gives an emotional-appropriateness score of 4.82±0.17 without describing how it was measured, and Table 2 omits error bars for several baseline entries. These omissions prevent verification of the claimed +39% improvement.
  4. [§4.1 and Table 2] The statistical significance claim (p<0.001, paired t-test) is stated without details of the samples, number of paired observations, or multiple-comparison corrections. The F1 scores in Table 2 are reported with error bars for only two systems, so the 'near-uniform performance (SD = 0.04)' assertion is not supported by the table as presented.
  5. [§3.5.1 and §3.6] The reward weights α,β,γ in Eq. (6) and the context weights α1..α4 in Eq. (11) are said to be 'dynamically adjusted' but the adjustment mechanism is never defined. Without a specification of how these parameters are set or learned, the reported improvements could simply reflect favorable tuning, and the results are not reproducible.
minor comments (5)
  1. [Throughout] There are numerous grammatical and typographical errors, e.g., 'has experienced rapid advancements' in Section 1, 'theagent’sbehaviour' in the introduction, and a missing space in 'theagents'. A careful proofreading pass is needed.
  2. [References] The reference list contains mismatched citations: Seo et al. (2019) is a galaxy-formation paper but is cited for ATOMIC, and Hemaspaandra et al. (2014) is about scoring rules but is cited for SenticNet. These should be corrected.
  3. [§3.10, Eq. (13)] The KL divergence in Eq. (13) does not specify the logarithm base; this affects the numerical values and should be stated (natural log vs log base 2).
  4. [Table 2 and Table 3] Table 2 should report standard deviations for all entries, not just a few, and the units of the 'Time' column in Table 3 should be stated (presumably milliseconds, but this is not explicit).
  5. [§3.10 and §4.1] The paper initially lists AffectNet, SEMAINE DB, and MERD as evaluation datasets, but Section 4.1 also mentions EmpatheticDialogues without clarifying how it fits into the reported results; the dataset list should be reconciled.

Circularity Check

2 steps flagged · score 7.0 of 10

Headline CSD and KL results are computed on the knowledge graph and its balancing target, not on generated responses, so the central cultural-adaptation numbers reduce to graph construction.

  1. self definitional [Section 3.10, Eq. (12); results in Table 1 and Abstract]
    "Cultural Semantic Density (CSD), a custom metric defined as: CSD = P c∈C 1(KG⊨c) / log(|C|+1), where C represents the set of all cultural concepts and 1(KG⊨c) indicates the knowledge graph's inclusion of concept c, measured how effectively the system's responses reflected cultural nuances."

    Eq. (12) contains no term involving the system's generated response; it counts only whether the knowledge graph entails each cultural concept. CSD is thus a property of KG content, so the 9.32/10 figure in Table 1 and the Abstract is determined by which concepts were inserted into the CEKG during construction, not by whether any generated response reflects cultural nuance. Any retriever drawing on a more complete KG scores higher by construction. The paper's own wording equates 'responses reflected cultural nuances' with 'knowledge graph's inclusion of concept c,' making the response-level claim definitionally identical to the graph-construction input.

  2. fitted input called prediction [Sections 3.3.2 and 3.10, Eq. (13); Abstract's '61% (KL-Divergence: 0.28)']
    "This alignment process reduced the KL-Divergence (a measure of cultural representation bias) from 0.72 to 0.28, representing a 61% reduction in cultural bias. ... KLbias = Σ_c P(c) log(P(c)/P_ideal(c)), where P(c) represents the observed proportion of culture c in KG responses and P_ideal(c) denotes a uniform target distribution."

    Bias is defined as KL divergence against a uniform P_ideal, and the KG-construction step in Section 3.3.2 explicitly 'balances cultural representation across the graph' toward that uniform target. The reported reduction from 0.72 to 0.28 is therefore a direct outcome of the balancing operation, restated as an evaluation result, rather than an independent measurement of generated-response bias. The Abstract and contributions present this 61% reduction as a framework performance result, but it is a fitted consequence of the chosen debiasing objective. As with Eq. (13)'s reference to 'KG responses,' the metric contains no generated-response term, so it cannot support the claimed superiority in culturally adaptive response generation.

full rationale

The two headline cultural-adaptation metrics reduce by construction. CSD (Eq. 12) is defined over knowledge-graph entailment, not over system outputs, so the 9.32 score measures what was put into the CEKG. KL bias (Eq. 13) is measured against a uniform target that Section 3.3.2's alignment process is explicitly designed to enforce, so the 0.28/61% figure is a restatement of the balancing procedure. These numbers appear in the Abstract, contributions, conclusion, and Table 1 as evidence of generation quality, making the central claim of 'consistently outperforms baselines in cultural adaptation' partially circular. Other reported results—MERD F1 by cultural subgroup, BLEU/ROUGE, human-rated emotional appropriateness, and latency—are externally meaningful and do not reduce to the graph, so the framework is not entirely circular. There is no load-bearing self-citation chain; the circularity is metric-definitional and construction-based.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central results depend on an unvalidated evaluation metric, a uniform-bias ideal, and unspecified reward and context weights. None of these is justified externally, so the evidentiary burden falls on paper-specific constructs that are not independently grounded.

free parameters (4)
  • Reward weights alpha, beta, gamma = not reported
    Eq. (6) defines total reward as alpha * R_cultural + beta * R_emotional + gamma * R_feedback and says the weights are dynamically adjusted based on the cultural context, but no values or adjustment rule are given.
  • Context weights alpha1 through alpha4 = not reported
    Eq. (11) computes context as a weighted combination of prior context, new information, user context, and retrieved knowledge; the adaptive coefficients are never specified.
  • Regeneration threshold tau = not reported
    Section 3.8 regenerates a response when min(comp_k, cohe) is below tau, but the threshold value and its calibration are not described.
  • Ideal culture distribution P_ideal = uniform
    Eq. (13) uses a uniform target distribution, and Section 3.3.2 balances the graph toward it; the reported bias reduction is therefore relative to a chosen ideal rather than an external ground truth.
assumptions (5)
  • ad hoc to paper Uniform cultural distribution is the correct unbiased target for cultural representation
    Eq. (13) and Section 3.3.2 define the bias target as uniform and balance the graph toward it; if real-world cultural prevalence is nonuniform, this equates debiasing with equal representation.
  • domain assumption StereoKG stereotypes are a valid representation of cultural emotional norms
    Section 3.3.2 uses StereoKG's 4,722 entries as the foundation of the culture emotion graph and treats stereotypes as real-world perceptions to be modeled.
  • domain assumption Hyperbolic Lorentz embeddings preserve culture emotion hierarchy well enough to improve cultural adaptation
    Section 3.4 assumes Euclidean embeddings inadequately capture hierarchy and that the Lorentz model does; the embedding step is imported from prior work.
  • domain assumption Valence, Arousal, and Dominance dimensions are sufficient to mediate cultural emotional appropriateness
    Section 3.3.2 assigns each emotion a VAD triple and Section 3.7 uses VAD cues to guide response tone; cultural emotional meaning may not collapse into these three dimensions.
  • domain assumption The similarity and coherence functions used in the rewards are meaningful and computable
    Eqs. (7) and (8) rely on sim(at, KG_cultural(st)) and VAD_coherence without specifying the models behind them; the reward signal is undefined until these functions are.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Affective-CARA: A Knowledge Graph Driven Framework for Culturally Adaptive Emotional Intelligence in HCI." pith.science (2026). https://pith.science/paper/U5G3IRBP

@misc{pith2026250614166,
  author       = {Pith},
  title        = {Pith review of: Affective-CARA: A Knowledge Graph Driven Framework for Culturally Adaptive Emotional Intelligence in HCI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U5G3IRBP}},
  note         = {Machine review of arXiv:2506.14166}
}
read the original abstract

Culturally adaptive emotional responses remain a critical challenge in affective computing. This paper introduces Affective-CARA, an agentic framework designed to enhance user-agent interactions by integrating a Cultural Emotion Knowledge Graph (derived from StereoKG) with Valence, Arousal, and Dominance annotations, culture-specific data, and cross-cultural checks to minimize bias. A Gradient-Based Reward Policy Optimization mechanism further refines responses according to cultural alignment, affective appropriateness, and iterative user feedback. A Cultural-Aware Response Mediator coordinates knowledge retrieval, reinforcement learning updates, and historical data fusion. By merging real-time user input with past emotional states and cultural insights, Affective-CARA delivers narratives that are deeply personalized and sensitive to diverse cultural norms. Evaluations on AffectNet, SEMAINE DB, and MERD confirm that the framework consistently outperforms baseline models in sentiment alignment, cultural adaptation, and narrative quality. Affective-CARA achieved a Cultural Semantic Density of 9.32 out of 10 and lowered cultural representation bias by 61% (KL-Divergence: 0.28), demonstrating robust performance in generating ethical, adaptive responses. These findings suggest the potential for more inclusive and empathetic interactions, making Affective-CARA an avenue for fostering culturally grounded user experiences across domains such as cross-cultural communication, mental health support, and education.

Figures

Figures reproduced from arXiv: 2506.14166 by the authors.

Figure 1
Figure 1. Affective-CARA Architecture: The system includes the CARM meta-controller, [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 27 canonical work pages

  1. [1]

    write newline

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

  2. [2]

    Hyperbolic hierarchical knowledge graph embeddings for biological entities

    Li, N., Yang, Z., Yang, Y., Wang, J., Lin, H., 2023. Hyperbolic hierarchical knowledge graph embeddings for biological entities. Journal of Biomedical Informatics, 147, 104503. https://doi.org/10.1016/j.jbi.2023.104503

  3. [3]

    Yang, K., Zhang, T., and Ananiadou, S. (2024). Deep emotion recognition in textual conversations: A survey. Artificial Intelligence Review, 15(2), 508–518

  4. [4]

    Zhang, Y., Wang, M., Tiwari, P., et al. (2022). Emotion Recognition With Knowledge Graph Based on Physiological Signals. Frontiers in Neuroscience, 15, 911767

  5. [5]

    Dann, C., Mansour, Y., and Mohri, M. (2023). Reinforcement Learning Can Be More Efficient with Multiple Rewards. Proceedings of Machine Learning Research, 202:6948--6967. Presented at the 40th International Conference on Machine Learning (ICML 2023), Honolulu, United States

  6. [6]

    Zhang, X., Li, Y., and Wang, M. (2024). Implementation of deep reinforcement learning models for emotion detection and learning personalization in hybrid educational environments. Frontiers in Artificial Intelligence, 3, 1458230

  7. [7]

    Yao, B., Shi, W., and Zhang, T. (2024). Leveraging sensory knowledge into Text-to-Text Transfer Transformer for emotion recognition. Information Processing & Management, 60(3), 103235

  8. [8]

    Zhang, Y., Wang, M., Tiwari, P., et al. (2023). A Reinforcement Learning-Based Approach for Promoting Mental Health Using Personalized Interventions. Future Internet, 15(2), 22

Show all 44 references
  1. [9]

    Liang, Q., Wang, W., Bao, F., and Gao, G. (2024). Fully Hyperbolic Rotation for Knowledge Graph Embedding. Proceedings of the 26th European Conference on Artificial Intelligence (ECAI)

  2. [10]

    Charles, R. (2023). Human-Guided Reinforcement Learning: Exploring Techniques and Ethical Considerations

  3. [11]

    Casper, S., Davies, X., Shi, C., et al. (2023). Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback. arXiv preprint arXiv:2307.15217

  4. [12]

    Chen, J., Li, X., and Wang, M. (2023). Mixed Emotion Recognition: A Comprehensive Review. IEEE Transactions on Affective Computing (Early Access), doi:10.1109/TAFFC.2023.1234567

  5. [13]

    Lian, Z., Sun, H., Sun, L., Chen, L., Chen, H., Gu, H., Wen, Z., Chen, S., Zhang, S., Yao, H., Xu, M., Chen, K., Liu, B., Liu, R., Liang, S., Li, Y., Yi, J., and Tao, J. (2024). Open-vocabulary Multimodal Emotion Recognition: Dataset, Metric, and Benchmark. arXiv preprint arXi...

  6. [14]

    McKeown, G., Valstar, M., Cowie, R., Pantic, M., and Schröder, M. (2012). The SEMAINE Database: Annotated Multimodal Records of Emotionally Colored Conversations between a Person and a Limited Agent. IEEE Transactions on Affective Computing, 3(1), 5–17. doi:10.1109/T-AFFC.2011.20

  7. [15]

    Li, Q., Li, P., Ren, Z., Ren, P., and Chen, Z. (2022). Knowledge Bridging for Empathetic Dialogue Generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10993–11001. :contentReference[oaicite:1] index=1

  8. [16]

    Zhao, H., Zha, X., and Zhang, Z. (2024). EmoTransKG: An Innovative Emotion Knowledge Graph to Reveal Emotion Transformation. Findings of the Association for Computational Linguistics: ACL 2024, 12098–12110. DOI:10.18653/v1/2024.findings-acl.720. :contentReference[oaicite:0] index=0

  9. [17]

    Wang, Z., Liu, X., Hu, M., Ying, R., Jiang, M., Wu, J., Xie, Y., Gao, H., and Cheng, R. (2024). ECoK: Emotional Commonsense Knowledge Graph for Mining Emotional Gold. Findings of the Association for Computational Linguistics: ACL 2024, 8055–8074. DOI:10.18653/v1/2024.findings-...

  10. [18]

    Zhang, Y., Yang, X., Xu, X., Gao, Z., Huang, Y., Mu, S., Feng, S., Wang, D., Zhang, Y., Song, K., and Yu, G. (2024). Affective Computing in the Era of Large Language Models: A Survey from the NLP Perspective. arXiv preprint arXiv:2408.04638v1 [cs.CL], July 30, 2024

  11. [19]

    Pei, G., Li, H., Lu, Y., Wang, Y., Hua, S., and Li, T. (2024). Affective Computing: Recent Advances, Challenges, and Future Trends. Intelligent Computing, 3. doi:10.34133/icomputing.0076

  12. [20]

    Y., Li, Q., and Wang, X

    Hu, H., Wan, Y., Tang, K. Y., Li, Q., and Wang, X. (2025). Affective-Computing-Driven Personalized Display of Cultural Information for Commercial Heritage Architecture. Applied Sciences, 15(7), 3459. https://doi.org/10.3390/app15073459

  13. [21]

    Barthet, M., Khalifa, A., Liapis, A., and Yannakakis, G. N. (2022). Play with Emotion: Affect-Driven Reinforcement Learning. In Proceedings of the 2022 10th International Conference on Affective Computing and Intelligent Interaction (ACII), University of Malta, Msida, Malta

  14. [22]

    Sabour, S., Liu, S., Zhang, Z., Liu, J., Zhou, J., Sunaryo, A., Lee, T., Mihalcea, R., and Huang, M. (2024). EmoBench: Evaluating the Emotional Intelligence of Large Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vo...

  15. [23]

    Pervez, F., Shoukat, M., Usama, M., Sandhu, M., Latif, S., and Qadir, J. (2024). Affective Computing and the Road to an Emotionally Intelligent Metaverse. IEEE Open Journal of the Computer Society, PP, 1–20. doi:10.1109/OJCS.2024.3389462

  16. [24]

    Gilbert, T., Lambert, N., Dean, S., Zick, T., and Snoswell, A. (2023). Reward Reports for Reinforcement Learning. Proceedings, pages 84–130. doi:10.1145/3600211.3604698

  17. [25]

    A., Khan, I

    Afzal, S., Khan, H. A., Khan, I. U., Piran, M. J., and Lee, J. W. (2023). A Comprehensive Survey on Affective Computing; Challenges, Trends, Applications, and Future Directions. arXiv preprint arXiv:2305.07665 [cs.AI]. https://doi.org/10.48550/arXiv.2305.07665

  18. [26]

    Mascarenhas, S., Prada, R., Paiva, A., and Hofstede, G. J. (2013). Social Importance Dynamics: A Model for Culturally-Adaptive Agents. In Proceedings (Vol. 8108, pp. 325–338). doi:10.1007/978-3-642-40415-3\_29

  19. [27]

    Hemaspaandra, E., Hemaspaandra, L., and Schnoor, H. (2014). A Control Dichotomy for Pure Scoring Rules. Proceedings of the AAAI Conference on Artificial Intelligence, 28(1). https://doi.org/10.1609/aaai.v28i1.8800

  20. [28]

    Seo, W.-Y., Kim, W.-T., Kwak, S., Hsieh, P.-Y., Han, C., and Hopkins, P. F. (2019). Effects of Gas on Formation and Evolution of Stellar Bars and Nuclear Rings in Disk Galaxies. arXiv preprint arXiv:1901.02021 [astro-ph.GA]. https://doi.org/10.48550/arXiv.1901.02021

  21. [29]

    and Kiela, D

    Nickel, M. and Kiela, D. (2017). Poincaré Embeddings for Learning Hierarchical Representations. arXiv preprint arXiv:1705.08039 [cs.AI]. https://doi.org/10.48550/arXiv.1705.08039

  22. [30]

    Balažević, I., Allen, C., and Hospedales, T. (2019). Multi-relational Poincaré Graph Embeddings. arXiv preprint arXiv:1905.09791 [cs.LG]. https://doi.org/10.48550/arXiv.1905.09791

  23. [31]

    Steinborn, V., Dufter, P., Jabbar, H., and Schuetze, H. (2022). An Information-Theoretic Approach and Dataset for Probing Gender Stereotypes in Multilingual Masked Language Models. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 921--932, Seattl...

  24. [32]

    Zhu, H., Wang, X., Wang, Z., and Xv, K. (2024). An emotion-sensitive dialogue policy for task-oriented dialogue system. Scientific Reports, 14. doi:10.1038/s41598-024-70463-x

  25. [33]

    A., and Choi, Y

    Sap, M., LeBras, R., Allaway, E., Bhagavatula, C., Lourie, N., Rashkin, H., Roof, B., Smith, N. A., and Choi, Y. (2019). ATOMIC: An Atlas of Machine Commonsense for If-Then Reasoning. arXiv preprint arXiv:1811.00146 [cs.CL]. https://doi.org/10.48550/arXiv.1811.00146

  26. [34]

    D., Ahmed, A

    Belay, T. D., Ahmed, A. H., Grissom II, A., Ameer, I., Sidorov, G., Kolesnikova, O., and Yimam, S. M. (2025). CULEMO: Cultural Lenses on Emotion -- Benchmarking LLMs for Cross-Cultural Emotion Understanding. arXiv preprint arXiv:2503.10688 [cs.CL]. https://doi.org/10.48550/arX...

  27. [35]

    Deas, N., Turcan, E., Pérez Mejía, I., and McKeown, K. (2024). MASIVE: Open-Ended Affective State Identification in English and Spanish. Columbia University, Department of Computer Science

  28. [36]

    Li, C., Teney, D., Yang, L., Wen, Q., Xie, X., and Wang, J. (2024). CulturePark: Boosting Cross-cultural Understanding in Large Language Models. Institute of Software, CAS; Idiap Research Institute; Westlake University; Squirrel AI; Microsoft Research; William & Mary

  29. [37]

    C., Gurevych, I., and Korhonen, A

    Liu, C. C., Gurevych, I., and Korhonen, A. (2025). Culturally Aware and Adapted NLP: A Taxonomy and a Survey of the State of the Art. arXiv preprint arXiv:2406.03930 [cs.CL]. https://doi.org/10.48550/arXiv.2406.03930

  30. [38]

    R., Jurafsky, D., and King, S

    Hofmann, V., Kalluri, P. R., Jurafsky, D., and King, S. (2024). AI generates covertly racist decisions about people based on their dialect. Nature, 633(8028), 147–154. doi:10.1038/s41586-024-07856-5

  31. [39]

    A., Kleiner, S., Patton, D

    Deas, N., Grieser, J. A., Kleiner, S., Patton, D. U., Turcan, E., and McKeown, K. (2023). Evaluation of African American Language Bias in Natural Language Generation. In Proceedings of EMNLP 2023

  32. [40]

    Cao, Y., Chen, M., and Hershcovich, D. (2024). Bridging Cultural Nuances in Dialogue Agents through Cultural Value Surveys. Huazhong University of Science and Technology; University of Copenhagen; South China University of Technology

  33. [41]

    Arora, A., Kaffee, L.-A., and Augenstein, I. (2023). Probing Pre-Trained Language Models for Cross-Cultural Differences in Values. In Proceedings of the First Workshop on Cross-Cultural Considerations in NLP (C3NLP), pages 114--130, Dubrovnik, Croatia. Association for Computat...

  34. [42]

    K., and Choudhury, M

    Saha, S., Pandey, S. K., and Choudhury, M. (2025). Meta-Cultural Competence: Climbing the Right Hill of Cultural Awareness. arXiv preprint arXiv:2502.09637 [cs.CY]. https://doi.org/10.48550/arXiv.2502.09637

  35. [43]

    H., Soon, L.-K., and Su, T

    Ng, S. H., Soon, L.-K., and Su, T. T. (2023). Emotion-aware chatbot with cultural adaptation for mitigating work-related stress. In E. Sari, M. Ghazali, & A. B. Tedjasaputra (Eds.), Asian HCI Symposium'23 Proceedings (pp. 41--50). Association for Computing Machinery (ACM). doi...

  36. [44]

    Friedrich, F., Stammer, W., Schramowski, P., & Kersting, K. (2023). Revision Transformers: Instructing Language Models to Change their Values. In Proceedings of the 26th European Conference on Artificial Intelligence (ECAI 2023) (pp. 756--763). IOS Press. doi:10.3233/FAIA230341

Pith tools

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