Pith. sign in

REVIEW 4 major objections 5 minor 4 cited by

CEDAR claims that identical scenarios carry culture-specific emotion ground truths, and that current multilingual LLMs fail to align with them, even when the prompt language matches the culture.

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 · deepseek-v4-flash

2026-08-03 09:38 UTC pith:QBDAFN7U

load-bearing objection Genuinely new cross-cultural affective benchmark with real human labeling, but the LLM-selection control is missing and the dissociation claim is over-stated. the 4 major comments →

arxiv 2601.13024 v2 pith:QBDAFN7U submitted 2026-01-19 cs.CL

Tears or Cheers? Benchmarking LLMs via Culturally Elicited Distinct Affective Responses

Chongyuan Dai , Yaling Shen , Zihan Gao , Jia Li , Yishun Jiang , Yaxiong Wang , Liu Liu , Zongyuan Ge
show 1 more author
Jinpeng Hu
This is my paper
classification cs.CL
keywords culturally elicited affectcross-cultural emotion understandingmultilingual LLM evaluationmultimodal benchmarkaffective alignmentemotion predictioncultural variationRussell's circumplex model
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.

This paper introduces CEDAR, a multilingual and multimodal benchmark built from everyday scenarios in which the same situation is read differently by different cultures: a white-lantern scene can mean mourning in East Asian contexts and calm contentment elsewhere. The central claim is that current LLMs do not internalize culture-specific affective norms, and that prompting a model in the language of the target culture does not reliably improve accuracy — sometimes it makes accuracy worse. Evaluating 17 models across seven languages, the paper reports consistently low accuracy, a systematic bias toward high-arousal emotions, neglect of deactivated states, and larger failures for Asian and low-resource languages. The authors see this as evidence that culturally grounded affective understanding is a distinct weakness that factual cultural-knowledge benchmarks do not capture.

Core claim

The paper's core discovery is a dissociation between language proficiency and cultural alignment: identical semantic content (parallel translations of a narrative, or the same image with a translated question) comes with different ground-truth emotion labels depending on the target culture, and current models generally fail to predict those labels. For example, on Japanese multimodal data, one evaluated model reaches 44.14% accuracy with English prompts but drops to 30.46% with Japanese prompts — a language match hurting rather than helping. The paper also shows that models over-predict salient, high-arousal emotions at the expense of subtle states like contentment and embarrassment, and tha

What carries the argument

The central object is CEDAR itself: 10,962 text and image scenarios across Arabic, Chinese, English, Hindi, Japanese, Spanish, and Swahili, with 14 fine-grained emotion categories and language-specific ground truths. The construction pipeline is the key mechanism: three LLMs generate provisional labels; instances are kept only when the models agree within a language but disagree across languages; Russell's circumplex model selects the clusters with maximal cross-language variation; and native-speaker annotation then fixes the ground-truth labels. This design embodies the paper's operating premise that semantic equivalence across languages does not imply emotional equivalence across cultures.

Load-bearing premise

The whole benchmark rests on the assumption that the cross-language disagreements detected by the three selecting LLMs are genuine cultural differences in emotion rather than translation artifacts, model bias, or label noise — because only those cases are kept and then human-annotated.

What would settle it

Have native-speaker annotators label a sample of the instances that were discarded for having uniform cross-language LLM predictions. If humans show the same cross-cultural disagreement on those discarded items — or if human labels on the retained items do not reproduce the LLM-based cross-language splits — the filtering procedure is manufacturing the claimed cultural variation, and the benchmark's central evidence collapses.

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

If this is right

  • If CEDAR is valid, culturally grounded affective understanding is a distinct capability that current models systematically lack, and it is not captured by factual cultural-knowledge benchmarks.
  • Matching the prompt language to the dataset language does not improve — and can degrade — emotion accuracy, so language fluency and cultural resonance should be evaluated as separate dimensions.
  • Models systematically over-predict high-arousal emotions and under-predict deactivated states, indicating a coarse affective prior that overshadows fine-grained cultural variation.
  • Performance drops markedly for Asian and low-resource languages, suggesting that non-Western affective norms are underrepresented in model training.
  • Multimodal instances are consistently harder than text-only ones, showing that visual-emotional grounding adds difficulty beyond language alone.

Where Pith is reading between the lines

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

  • The paper openly targets high-consensus scenarios validated by strict majority voting; this means the benchmark measures prototypical cultural signals, so the size of the model–human gap could be different in lower-consensus, contested, or personal emotion situations.
  • Because the pipeline selects instances through LLM disagreement, the dataset is partly a product of the selecting models' biases; human labels on the discarded uniform cases would be needed to confirm that the cross-language splits are genuine cultural variation rather than artifacts.
  • The reported dissociation suggests English is a latent reasoning pivot for many models; a testable consequence is that fine-tuning on native-language cultural narratives may improve affective alignment more than translation or prompt-language matching.
  • The 10,962 instances could serve as a training signal for culture-conditioned emotion adapters, with the key question being whether gains transfer to unseen scenarios in the same language-culture groups.

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

4 major / 5 minor

Summary. The paper introduces CEDAR, a multimodal, multilingual benchmark designed to evaluate whether LLMs align with culture-specific affective interpretations of identical scenarios. The construction pipeline first generates English narrative-question pairs, filters them through three LLM judges by retaining only instances where provisional labels agree within a language but disagree across languages, then translates and validates the survivors with native-speaker annotations. The final dataset contains 10,962 items across seven languages and 14 emotion categories, split into 400 multimodal and 1,166 text-only samples per language. The authors evaluate 17 models and report accuracy, emotion-prediction propensity, Russell-quadrant bias, and prompt-language effects. The central claim is that current LLMs show a dissociation between language consistency and cultural alignment: matching the prompt language to the target culture does not reliably improve, and can even hurt, affective accuracy.

Significance. If the benchmark's premise is valid, CEDAR addresses a real gap: most cultural benchmarks test declarative knowledge rather than the subjective, interpretive layer of emotion. The pipeline is detailed and the human-annotation component is a genuine strength, with inter-annotator agreement reported in Table 2. The breadth of evaluation (17 models, 7 languages, two modalities) and the prompt-language heatmaps (Figure 7) are useful empirical contributions. However, the benchmark's entire validity rests on the assumption that the LLM disagreement filter in §2.3 isolates genuine cultural variation. The paper does not validate this assumption against human judgments on the filtered-out instances, and translation equivalence is only checked for fluency, not meaning preservation. These are load-bearing gaps, not cosmetic ones: if the filter selects for translation artifacts or model-specific biases, both the dataset and the headline dissociation result are artifacts of the selection procedure. The contribution is potentially significant, but the current evidence is insufficient to support the central claim.

major comments (4)
  1. [§2.3, §2.5] The selection criterion is unvalidated. The Consistency and Variation Filtering retains instances only when three LLMs agree within a language but disagree across languages; human annotation (§2.5) is applied only to these survivors. There is no human annotation on the rejected pool (roughly 35K of the 42K candidates), so there is no evidence that cross-language human variability is higher in retained than in discarded items. Without this comparison, the benchmark's core claim—that these scenarios carry different culture-specific ground truths—is an operational consequence of the LLM filter, not an empirically established property. Please provide a human baseline on a random sample of the discarded instances (and ideally on re-translated items) showing that retained instances exhibit larger human cross-language divergence than discarded ones. This is essential to establish that CEDAR mea
  2. [§2.5, Figure 9/10] Translation equivalence is not verified semantically. Non-English versions are produced by GPT-4.5 and checked by native speakers for fluency and correctness, but not for meaning preservation. Translation choices can alter connotations (e.g., a differently described gift, a shifted social frame), so cross-language label differences—whether human or model—may reflect stimulus differences rather than cultural interpretation. For example, Table 9 example 9 ('spicy rice and soup dishes with herbs and pork') has ground truth 'disgust' in Arabic but 'contentment' in Chinese and several other languages; without demonstrating semantic equivalence, this could reflect how the dish was translated rather than how Arab versus Chinese respondents culturally interpret the same scenario. Please add a back-translation or human meaning-equivalence check, or report that such checks were conducted.
  3. [§2.3, Table 1] The provisional-label models overlap substantially with the evaluated models. Claude4.5-Sonnet and Gemini2.5-Flash are both used in the selection filter (§2.3) and later appear in the evaluation set (Table 1). GPT-4.5 is used for NQ generation, translation, and the image-necessity filter. Because the benchmark is intentionally built from instances where these particular model families disagree, their low cultural-alignment scores may be inflated by selection pressure. This is not circularity in the human-labeling sense (ground truths are human-derived), but it weakens the dissociation claim. I recommend reporting results on a random, unfiltered subset of the candidate pool, or at least on instances selected by a disjoint set of selector models, to show that the observed pattern is not an artifact of the specific selectors used.
  4. [Abstract, §4.4] The paper frames the low accuracy and the prompt-language dissociation as evidence that 'culturally grounded affective understanding remains a significant challenge.' But the benchmark is adversarially filtered by design: it keeps only scenarios where the selector LLMs disagree across languages. On such a subset, low accuracy is expected even for a culture-competent model, because the items are selected to be the hardest cases. The paper should explicitly contextualize the absolute accuracy numbers against a baseline of naturally occurring cultural variation (e.g., a random sample of the original 42K candidates before filtering). Without this baseline, the 'significant challenge' claim is overstated, though the relative model-level comparisons remain informative.
minor comments (5)
  1. [§4.3] The text refers to 'LSRQB' but Figure 6 labels the metric 'LSB' — please align the notation.
  2. [§2.3] The text says the filtering yields 'approximately 7K NQ pairs,' while the final dataset contains 1,166 text-only samples per language (8,162 text-only instances total) plus 400 multimodal instances per language. Please clarify whether the 7K figure refers to unique English scenarios before translation, and reconcile the numbers explicitly.
  3. [Table 2] The inter-annotator agreement is reported with Krippendorff's alpha and average pairwise F1, but the number of annotators per instance and the annotation instructions (e.g., forced choice among 14 emotions) are only vaguely described in §2.5 and Appendix A.4. Please provide the exact annotation interface and the distribution of majority sizes.
  4. [Figure 7] The heatmaps are dense and the numeric labels are small. Consider using discrete color bins or removing redundant numbers so the language-consistency pattern is visually clearer.
  5. [References] Several references are incomplete or inconsistent (e.g., 'Gemma Team. 2025a. Gemma 3.' lacks full bibliographic details; some system-card citations lack access dates). Please standardize.

Circularity Check

1 steps flagged

No significant circularity: human ground truths break the selection loop, though selector/evaluator model overlap mildly inflates variance claims.

specific steps
  1. other [§2.3 Consistency and Variation Filtering; Table 1 (evaluation of Claude4.5-Sonnet and Gemini2.5-Flash)]
    "we employ state-of-the-art LLMs (i.e., Claude4.5-Sonnet (Anthropic, 2025), Gemini2.5-Flash (Comanici et al., 2025), and GPT-4.5) to generate provisional predictions for each instance across languages. We first impose within-language agreement ... We then enforce cross-language variation by comparing these provisional labels across languages and removing instances with uniform predictions"

    The inclusion criterion for CEDAR is cross-language disagreement among Claude4.5-Sonnet, Gemini2.5-Flash, and GPT-4.5. The same Claude and Gemini models later appear as evaluated systems in Table 1. Hence their cross-language variance and language-specific propensity scores on CEDAR are partially guaranteed by the selection filter: any instance in the benchmark was kept only if these models' provisional labels differed across languages. This makes the paper's variance/stability observations for these two models a restatement of the selection rule rather than an independent empirical discovery. The central result—dissociation between prompt-language consistency and accuracy against human ground truth—is not reduced, because human labels are collected after and independently of the filter, s

full rationale

CEDAR's construction chain is: seed data → GPT-4.5 NQ generation → Llama3.3 refinement → basic filtering → LLM provisional-label consistency/variation filtering → image construction → cultural-variation selection → GPT-4.5 translation → native-speaker human annotation → evaluation. The ground-truth labels used for scoring are exclusively human labels (Section 2.5), not the provisional LLM labels, so the central accuracy numbers and the dissociation result are not reductions of the selection filter. The single partial circularity is that Claude4.5-Sonnet and Gemini2.5-Flash are both selector models (Section 2.3) and evaluated models (Table 1); their reported cross-language variance is partly inherited from the inclusion criterion. This is a validity/selection concern, not a full circularity, because the human labels are independent and the paper's main claim about language consistency vs cultural alignment is scored against those human labels. No self-citation chain is load-bearing (the only author-overlap citation, Hu et al. 2024, is a routine related-work citation), and no fitted parameter is renamed as a prediction. The Skeptic's concern—that LLM disagreement may reflect translation artifacts rather than cultural variation—is an unvalidated premise about construct validity, not a circular derivation; the paper does not define cultural truth in terms of LLM outputs after annotation.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No mathematical derivation is attempted; the contribution is empirical. The main hidden costs are the assumptions that LLM cross-language disagreement marks true cultural divergence, that GPT-4.5 translations preserve scenario semantics, and that Prolific majority labels define culture-specific ground truth. No invented physical or theoretical entities are introduced.

free parameters (4)
  • Emotion-to-quadrant mapping = Q1:{amusement,happiness,surprise};Q2:{anger,disgust,fear,pain};Q3:{embarrassment,sadness};Q4:{awe,contentment,desire,rel
    Hand-chosen mapping (Fig. 8) drives both instance selection (§2.5) and the RQB bias metric (§3.3, §4.3); a different mapping would change which instances are retained and what bias looks like.
  • LLM disagreement selection thresholds = within-language majority; cross-language non-uniform; top-quadrant-disagreement per cluster
    Thresholds in §2.3 and §2.5 determine benchmark composition and difficulty; they are author-selected and not derived from data or released as a configuration.
  • Text length bounds = 50–200 characters
    Applied in Basic Filtering (§2.3), removes ~2.7% of candidates; arbitrary bounds affect which scenarios survive.
  • Annotator majority rule = at least 5; +2 if no majority
    Human annotation protocol (§2.5, Appendix A.4); Krippendorff's alpha ranges 0.69–0.93 across languages/modalities, so label uncertainty remains.
axioms (5)
  • domain assumption The 14 emotion categories adapted from Ekman (1992) and Cordaro et al. (2016) are a valid universal rubric for cross-cultural emotion comparison.
    Used as the fixed label set in all seven languages; if the categories themselves are culturally biased, all ground-truth labels are constrained by that frame.
  • domain assumption GPT-4.5 translation preserves semantic equivalence across the seven languages.
    All non-English versions are produced by translating English candidates (§2.5); native-speaker verification checks fluency and correctness, not emotional equivalence.
  • domain assumption Cross-language disagreement in LLM provisional labels marks genuine cultural variation.
    The selection filter in §2.3 retains only instances where three LLMs agree within a language but differ across languages; human labels are collected only after this filter.
  • domain assumption Native-speaker majority labels are valid culture-specific ground truth.
    At least five Prolific annotators per instance (§2.5, Appendix A.4); inter-annotator agreement is moderate (α 0.69–0.93), not perfect.
  • domain assumption Russell's circumplex quadrants capture meaningful affective dimensions.
    The four-quadrant model is used for data selection (§2.5) and for the RQB analysis (§3.3, §4.3).

pith-pipeline@v1.3.0-alltime-deepseek · 26521 in / 12464 out tokens · 131213 ms · 2026-08-03T09:38:11.517776+00:00 · methodology

0 comments
read the original abstract

Culture serves as a fundamental determinant of human affective processing and profoundly shapes how individuals perceive and interpret emotional stimuli. Despite this intrinsic link extant evaluations regarding cultural alignment within Large Language Models primarily prioritize declarative knowledge such as geographical facts or established societal customs. These benchmarks remain insufficient to capture the subjective interpretative variance inherent to diverse sociocultural lenses. To address this limitation, we introduce CEDAR, a multimodal benchmark constructed entirely from scenarios capturing Culturally \underline{\textsc{E}}licited \underline{\textsc{D}}istinct \underline{\textsc{A}}ffective \underline{\textsc{R}}esponses. To construct CEDAR, we implement a novel pipeline that leverages LLM-generated provisional labels to isolate instances yielding cross-cultural emotional distinctions, and subsequently derives reliable ground-truth annotations through rigorous human evaluation. The resulting benchmark comprises 10,962 instances across seven languages and 14 fine-grained emotion categories, with each language including 400 multimodal and 1,166 text-only samples. Comprehensive evaluations of 17 representative multilingual models reveal a dissociation between language consistency and cultural alignment, demonstrating that culturally grounded affective understanding remains a significant challenge for current models.

Figures

Figures reproduced from arXiv: 2601.13024 by Chongyuan Dai, Jia Li, Jinpeng Hu, Liu Liu, Yaling Shen, Yaxiong Wang, Yishun Jiang, Zihan Gao, Zongyuan Ge.

Figure 1
Figure 1. Figure 1: A representative example of culturally distinct [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Statistics of CEDAR. PolyGuard (Kumar et al., 2025) which is a special￾ized multilingual safety moderation framework to identify and eliminate toxic content including hate speech or harmful stereotypes. By applying these rigorous quality control measures we establish a curated corpus that is both task relevant and safe for broad research applications. Consistency and Variation Filtering. To identify candid… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of GEPP across multimodal and text-only subsets. The scatter points represent the individual [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: LSEPP for four illustrative models on multimodal and text-only subsets. Each axis represents the LSEPP [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Model comparison based on the GRQB. The chart displays the GRQB of model predictions across the four quadrants for multimodal and text-only subsets. ciated with low arousal. In the multimodal sub￾set, models exhibit a strong and consistent incli￾nation towards high-arousal emotions situated in Quadrant I and Quadrant II, while systematically under-predicting deactivated emotions associated with Quadrants I… view at source ↗
Figure 6
Figure 6. Figure 6: LSRQB values for all evaluated models across [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance heatmaps analyzing the im￾pact of prompt language (y-axis) versus dataset lan￾guage (x-axis) on standard accuracy. Results are shown for Gemma3-27B-It (top) and Qwen3-VL-8B (bottom) across both multimodal and text-only subsets. on the multimodal set demonstrate that the combi￾nation of a Japanese prompt with Japanese data yields an accuracy of only 30.46% which repre￾sents the lowest score amon… view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of the 14 emotion categories on [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Evaluation prompts for the text-only subset. [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ShredBench: Evaluating the Semantic Reasoning Capabilities of Multimodal LLMs in Document Reconstruction

    cs.CV 2026-04 unverdicted novelty 7.0

    ShredBench shows state-of-the-art MLLMs perform well on intact documents but suffer sharp drops in restoration accuracy as fragmentation increases to 8-16 pieces, indicating insufficient cross-modal semantic reasoning...

  2. IntervenSim: Intervention-Aware Social Network Simulation for Opinion Dynamics

    cs.SI 2026-04 unverdicted novelty 7.0

    IntervenSim is an intervention-aware social network simulation that couples source interventions with crowd interactions in a feedback loop, improving MAPE by 41.6% and DTW by 66.9% over prior static frameworks on rea...

  3. Expressing Social Emotions: Misalignment Between LLMs and Human Cultural Emotion Norms

    cs.CL 2026-04 unverdicted novelty 6.0

    Frontier LLMs over-express engaging emotions relative to disengaging ones and generate deterministic responses that fail to match the cultural and individual diversity observed in human social emotion expression.

  4. CogEvolution: A Human-like Generative Educational Agent to Simulate Student's Cognitive Evolution

    cs.AI 2026-04 unverdicted novelty 4.0

    CogEvolution combines ICAP cognitive taxonomy, IRT memory retrieval, and evolutionary algorithms into a generative agent that simulates dynamic student cognitive evolution and outperforms baselines in fidelity and lea...

Reference graph

Works this paper leans on

19 extracted references · 5 linked inside Pith · cited by 4 Pith papers

  1. [1]

    Grounding, which describes the environment or setting where the story takes place

  2. [2]

    InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 32854–32883, Suzhou, China

    CARE: Multilingual human preference learn- ing for cultural awareness. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 32854–32883, Suzhou, China. Association for Computational Linguistics. Abdullah Hashmat, Muhammad Arham Mirza, and Agha Ali Raza. 2025. PakBBQ: A culturally adapted bias benchmark for QA. In...

    2025

  3. [3]

    a proud Italian

    Action, which depicts interactions between the character(s) and the environment, or between characters themselves. NOTE: DO NOT include any specific location or nationality information in the [scenario] or [narrative] (e.g., avoid phrases like "a proud Italian", "in Chicago’s South Side", or "in the vibrant city of Seville"). Finally, provide a[question] ...

  4. [4]

    InPro- ceedings of the 2025 Conference on Empirical Meth- ods in Natural Language Processing, pages 24–51, Suzhou, China

    Break the checkbox: Challenging closed-style evaluations of cultural alignment in LLMs. InPro- ceedings of the 2025 Conference on Empirical Meth- ods in Natural Language Processing, pages 24–51, Suzhou, China. Association for Computational Lin- guistics. Shinobu Kitayama and Dov Cohen. 2010. Handbook of cultural psychology. Priyanshu Kumar, Devansh Jain, ...

    Pith/arXiv arXiv 2025

  5. [5]

    InProceedings of the 2025 Conference on Empiri- cal Methods in Natural Language Processing, pages 23939–23967, Suzhou, China

    Scalable and culturally specific stereotype dataset construction via human-LLM collaboration. InProceedings of the 2025 Conference on Empiri- cal Methods in Natural Language Processing, pages 23939–23967, Suzhou, China. Association for Com- putational Linguistics. Mistral AI Team. 2025. Mistral small 3.1. https: //mistral.ai/news/mistral-small-3-1 . Ac- c...

    2025

  6. [6]

    InProceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 16366–16393, Bangkok, Thailand

    Having beer after prayer? measuring cultural bias in large language models. InProceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 16366–16393, Bangkok, Thailand. Association for Computational Linguistics. Shravan Nayak, Mehar Bhatia, Xiaofeng Zhang, Verena Reiser, Lisa Anne Hendricks, Sj...

    2025

  7. [7]

    CROPE: Evaluating in-context adaptation of vision and language models to culture-specific con- cepts. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7917– 7936, Albuquerque, New Mexico. Association for Computational Lin...

    Pith/arXiv arXiv 2025

  8. [8]

    Qwen Team

    Kimi k2: Open agentic intelligence.arXiv preprint arXiv:2507.20534. Qwen Team. 2025b. Qwen3 technical report.Preprint, arXiv:2505.09388. Ahmet Üstün, Viraat Aryabumi, Zheng Yong, Wei-Yin Ko, Daniel D’souza, Gbemileke Onilude, Neel Bhan- dari, Shivalika Singh, Hui-Lee Ooi, Amr Kayid, Fred- die Vargus, Phil Blunsom, Shayne Longpre, Niklas Muennighoff, Marzi...

    Pith/arXiv arXiv 2024

  9. [10]

    Background context, which describes what happened before the story begins

  10. [12]

    Identify the main character(s) in the[narrative]

  11. [13]

    Sofia" ->

    For the main character, convert all third-person references to second-person equivalents while maintaining sentence fluency and coherence (e.g., "Sofia" -> "you", "the friends" -> "you and your friends", "a young man’s phone" -> "your phone", "the Kannadiga students" -> "you and your compatriots", "a Chinese student, Wei" -> "you")

  12. [14]

    DO NOT modify any plot details or descriptions

    Preserve all plot points and descriptions from the original [narrative] and[question]. DO NOT modify any plot details or descriptions

  13. [15]

    refined_narrative

    Adjust verb forms and grammar as needed to ensure grammatical correctness in second-person narration. Here are the given narrative and question:[narrative]: {narrative}[question]: {question} Provide your response in the following format. DO NOT include any explanation: ```JSON { "refined_narrative": "...", "refined_question": "..." } 20 Prompts for Contex...

  14. [16]

    Identify the target action: Understand the specific action or event in the[question] that requires emotion prediction

  15. [17]

    Locate relevant content: Find that action and its related descriptions in the[narrative]

  16. [18]

    fostering a sense of unity and pride,

    Remove emotional descriptions: Delete all words and phrases that explicitly express emotions (e.g., "fostering a sense of unity and pride," "with a mix of amusement and concern", "sparking excitement and hesitation"), while preserving objective factual descriptions

  17. [19]

    No need to modify

    If the [narrative] does NOT contain explicit emotional descriptions, return exactly: "No need to modify." NOTE: DO NOT modify any plot, action, or event; only remove emotional description words and phrases; maintain the coherence and readability of the narrative after removal. Here are some examples: # Example 1: [narrative]: Under the shade of old oak tr...

  18. [2024]

    Yuchen Huang, Zhiyuan Fan, Zhitao He, Sandeep Polisetty, Wenyan Li, and Yi R

    Psycollm: Enhancing llm for psychological understanding and evaluation.IEEE Transactions on Computational Social Systems. Yuchen Huang, Zhiyuan Fan, Zhitao He, Sandeep Polisetty, Wenyan Li, and Yi R. Fung. 2025. Cul- tureCLIP: Empowering CLIP with cultural awareness through synthetic images and contextualized captions. InSecond Conference on Language Mode...

    Pith/arXiv arXiv 2025

  19. [2025]

    In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 10978–11002, Suzhou, China

    NileChat: Towards linguistically diverse and culturally aware LLMs for local communities. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 10978–11002, Suzhou, China. Association for Com- putational Linguistics. Shangbin Feng, Taylor Sorensen, Yuhan Liu, Jillian Fisher, Chan Young Park, Yejin Choi, and Yulia...

    Pith/arXiv arXiv 2025