Pith. sign in

REVIEW 3 major objections 6 minor 45 references

NTDH: Complex Reasoning for Comprehensive Affective Analysis

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that treating affective analysis as a complex-reasoning problem, with reasoning traces verified against task-specific tolerances and refined by domain-aware strategies, lets an 8B model beat much larger instruction-tuned…

desk verdict Serious, credit-worthy two-stage RL recipe with a clean internal SFT-to-RL comparison, but the headline '14x fewer data' and 'strongest EI-reg' claims are undercut by the Qwen3-8B base-model confound and verbatim reuse of the authors' own EmoLLM baselines. read the letter →

arxiv 2608.06425 v1 pith:JFDKA6GS submitted 2026-08-05 cs.CL cs.LG

Tianlei Zhu , Zhiwei Liu , Yuyan Wang , Xiao-Yang Liu , Sophia Ananiadou This is my paper · ORCID
classification cs.CLcs.LG
keywords affectivecomputingemotionanalysisreasoningpathsreinforcementlearningGRPOreasoning-datasynthesisSemEval-2018complex
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

NTDH claims that the main obstacle to comprehensive affective analysis is not model capacity or data volume but the quality of synthetic reasoning data. The paper turns each raw SemEval-2018 label into a natural-language gold sentence, verifies candidate reasoning chains against each subtask's own scoring tolerance, refines failures with affect-science strategies, and steers stalled chains with hints that report only the direction of the error. Using 16,302 training records, about fourteen times fewer than a comparable instruction-tuned corpus, this pipeline trains an 8B model whose final policy beats its own SFT checkpoint on five of six official-test metrics and achieves the strongest emotion-intensity regression result among compared systems, at a Pearson correlation of 0.862. If right, it shows that verifiable rewards and carefully controlled reasoning data can substitute for large amounts of curated instruction data.

What carries the argument

The load-bearing machinery is the NTDH synthesis-and-training pipeline. Naturalisation rewrites each gold label as a scale-aware sentence so that the generator, the verifier, and the final supervised target share one form. A Tolerance-aware gate replaces a fallible LLM judge with a deterministic check: absolute error at most 0.05 for the regression subtasks, exact class match for ordinal valence, and F1 at least 0.7 for multi-label emotion construction. Domain-aware refinement samples four strategies, Backtrack, Explore, Verify, and Correct, grounded in valence shifters, irony, the circumplex, and emotion co-occurrence, while Directional Hints report only the type and direction of an error, never the gold value. Verified traces feed SFT; unconverged hard cases and the untouched half of the data feed GRPO, which optimises a binary reward defined by the same tolerance. This closes the loop between data synthesis and policy optimisation.

What would settle it

Run the strongest baselines, for example EmoOPT, EmoLLaMA-chat-13B, and SeerNet, through NTDH's exact evaluation harness, answer parsing, and official test split; if any of them reaches or exceeds 0.862 Pearson on EI-reg, the claim of being the strongest system is not supported, while the internal SFT-to-RL gain can still be verified independently.

Watch

Extended reading notes

Core claim

The paper's central claim is that four heterogeneous affective tasks, namely continuous emotion-intensity regression, valence regression, ordinal valence classification, and multi-label emotion classification, can be treated as one complex-reasoning problem, provided the training data satisfy four quality conditions: gold-consistent answer targets, evaluation-aware verification, domain-informed refinement, and label-free corrective hints. Applying these conditions through NTDH, the authors train Qwen3-8B with SFT on 5,388 verified reasoning traces and GRPO on 10,914 answer-only instances, using the same task tolerance for both verification and reward. On the official 9,201-instance test set, the RL-final policy outperforms its SFT initialization on five of six metrics and reports the strongest EI-reg Pearson correlation, 0.862, among the compared systems, while remaining competitive on the other subtasks.

Load-bearing premise

The headline comparison to prior systems assumes that scores copied verbatim from a related paper are directly comparable to NTDH's own evaluation, even though the systems used different backbones, training data, and inference procedures.

Editorial extensions

If this is right

  • A single 8B policy can cover four different affective output spaces, removing the need for task-specific prediction heads and per-task instruction corpora.
  • Under the strict tolerance-based reward, RL improves EI-reg from 0.800 to 0.862, V-reg from 0.785 to 0.840, V-oc from 0.785 to 0.831, and E-c Jaccard from 0.557 to 0.579, with macro-F1 the only metric that declines.
  • The same tolerance that gates data construction can serve as the GRPO reward, so synthesis and optimisation are aligned by design.
  • Because every training instance is used exactly once, converged traces for SFT and hard cases plus the second half for RL, the recipe avoids throwing away non-converged reasoning attempts.
  • The dominant data-quality lever is the gold-consistent answer target: only 18.4% of initial model conclusions fall within gold tolerance, so naturalisation and the tolerance gate carry most of the benefit.

Reading between the lines

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

  • Editorial inference: the NTDH data-quality recipe should transfer to any task with a checkable answer and a defined scoring tolerance, such as graded relevance, ordinal stance, or multi-label tagging outside affect; the paper names this direction but does not demonstrate it.
  • Editorial inference: the unused gold, silver, and bronze quality tiers suggest a direct follow-up, weighting the SFT loss by tier, that could sharpen the policy on hard-but-converged traces without extra data.
  • Editorial inference: the reported roughly 77% zero-variance GRPO groups imply that, in sparse-reward settings, adding more data is less valuable than selecting prompts that straddle the decision boundary; a targeted experiment could test this selection rule directly.
  • Editorial inference: because the E-c construction gate accepts traces at F1 at least 0.7 while the reward requires exact set match, a label-wise reasoning-to-answer entailment check is a natural next verifier; the paper identifies the mismatch as a limitation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes NTDH, a quality-aware reasoning-data synthesis pipeline for comprehensive affective analysis across four SemEval-2018 Task 1 subtasks (EI-reg, V-reg, V-oc, E-c). The pipeline converts raw labels into naturalised gold-consistent targets, verifies generated reasoning traces with a deterministic tolerance-aware gate, refines failures using domain-aware strategies and directional hints, and routes converged traces to SFT and unconverged/half-B samples to GRPO. The authors train Qwen3-8B with SFT followed by GRPO with a binary tolerance-matched reward, and report that with 16,302 training records (about 14x fewer than EmoLLM's 234K instruction records), the RL-final policy improves over its SFT initialisation on five of six official-test metrics and achieves the strongest EI-reg result among compared systems (Pearson 0.862), while remaining competitive on the other subtasks. The paper includes component ablations, qualitative error analysis, and a discussion of limitations.

Significance. If the headline claims hold, NTDH offers a data-efficient recipe for building verifiable-reward reasoning data in affective computing, and it demonstrates that an 8B open model can be competitive with much larger instruction-tuned systems. The paper has several genuine strengths: the internal SFT-init-to-RL-final comparison is a clean within-model comparison with no checkpoint selection; the tolerance-aware verifier and the explicit routing of every training instance to either SFT or RL are transparent and reproducible; and the authors honestly acknowledge key limitations, including single-run ablations, the E-c reasoning-answer mismatch under the permissive F1>=0.7 gate, and the proxy-based estimate of zero-variance GRPO groups. However, the central data-efficiency and 'strongest EI-reg' claims are currently undermined by two confounds: the comparison against EmoLLM mixes the NTDH recipe with the much more capable Qwen3-8B base model, and the baseline numbers are reproduced verbatim from the authors' own EmoLLM paper without re-evaluation under identical conditions.

major comments (3)
  1. [§5.2 and Abstract] The headline claim that NTDH 'achieves the strongest EI-reg result among the compared systems at a Pearson correlation of 0.862' and the 'about 14x fewer records' data-efficiency claim are confounded by the base-model choice. NTDH is trained on Qwen3-8B (2025), whereas the EmoLLM family is built on BART, T5, OPT, BLOOM, and LLaMA-2-chat, which are 2023-era models with substantially different pretraining scale and instruction-following ability. The paper provides no control: it does not train an EmoLLM-style SFT pipeline on Qwen3-8B, nor run NTDH on a LLaMA-2-class base, nor report a zero-shot Qwen3-8B baseline in Table 7. To support the data-efficiency interpretation, the authors should add at least one such control experiment (or an equivalent matched-capability comparison), or explicitly reframe the claim as a within-model demonstration rather than a comparison against EmoLLM.
  2. [§5.1 and Table 7] Section 5.1 states 'All baseline numbers are reproduced verbatim from [14]' and Table 7 presents these alongside the two NTDH checkpoints. Because [14] is a paper with overlapping authors and the baselines use different backbones, training data, and inference procedures, the evaluation conditions are not identical for the baselines and the NTDH models. As a result, the 'strongest EI-reg result among the compared systems' claim is not established by the current evidence. The authors should either re-run the baselines under their own evaluation harness (including the same parsing and metric code) and report those numbers, or clearly restrict all claims to 'as reported in [14]' throughout the abstract and body.
  3. [§5.2 and §5.3] The central within-paper comparison (RL-final vs SFT init) is based on a single training run with no error bars, confidence intervals, or significance tests. Tables 7 and the text report only point estimates, and the paper elsewhere acknowledges that some ablations are single-run (e.g., the D ablation in §5.3 is explicitly a single-run estimate). Given the emphasis on the five-of-six improvement, the authors should report variance across at least a few seeds, or provide a paired significance test on the official test set, or explicitly state that the results are single-run observations and temper the corresponding claims.
minor comments (6)
  1. [Abstract and §1] The claim 'to our knowledge, this is the first such treatment covering both sentiment and emotion' is very strong; I suggest softening it to a more precise statement about being the first verifiable-reward reasoning pipeline for this unified task set, and adding a more explicit comparison with THOR, EmoRationale, and InstructERC which also involve reasoning or instruction tuning for affective tasks.
  2. [§3.3] The sentence 'The judge LLM is retained only to parse heterogeneous natural-language conclusions and for logging' is slightly confusing because the T ablation replaces the deterministic gate with the original LLM judge as the verifier. Please clarify the distinction between the judge's parsing role and the verification role, and how the ablation actually works.
  3. [§5.1 and Table 5] Table 5 reports 8,150 half-A CoT generations with a Gold/Silver/Bronze/Unconv sum of 8,150, but the text says 16 empty- or invalid-chain cases are routed to RL, leaving 5,388 SFT trajectories. Please clarify how these 16 cases are counted in the tier distribution and in the 'No sample required label leakage' statement.
  4. [Table 6] In the GRPO column, 'Max length 4096 / 8192 (prompt / completion)' is ambiguous; please clarify whether the prompt length includes the reasoning trace and the answer block, and whether the completion length is the maximum for the generated output only.
  5. [§6] The 'controlled estimate' of zero-variance groups (n=70 prompts, k=8 samples per prompt) lacks details about the proxy generator, the sampling temperature, and the prompts used. Please provide more information or a reference so that the estimate can be assessed.
  6. [§3.2] The note that the E-c index map 'is fixed in the released code and is not the alphabetical listing of §3.2' is helpful but the code is not yet public; since the paper says the repository is to be released, please include the actual index map in the paper or supplementary material to make the evaluation fully self-contained.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity; one ablation metric is definitional, and the self-cited baseline table is an auditability concern rather than a circular reduction.

  1. self definitional [Section 5.3, 'Naturalisation (N)' paragraph; also §3.1 Stage 2 and Contribution 2]
    "With naturalisation, the answer source is instead the naturalised gold label, so the <answer> field of every retained SFT sample is correct by construction. Accordingly, 100% here denotes answer-target correctness."

    The claimed data-quality effect 'answer-source gold consistency 100%' is entailed by the definition of Naturalisation, which overwrites the candidate answer with the naturalised gold label. The 100% is therefore a tautology, not an empirical finding; the contrast with 18.4% measures the raw generator's success rate, not an effect of N. The paper itself states that the answer is 'correct by construction', confirming that this particular ablation result is definitional. This does not invalidate the independent SFT-to-GRPO evaluation, but it does mean the N component's headline data-side number provides no empirical evidence beyond the construction rule.

full rationale

The central derivation chain—NTDH synthesis, SFT on verified traces, GRPO with a binary tolerance-matched reward, and evaluation on the official SemEval-2018 test split—is self-contained. The reported headline results (EI-reg Pearson 0.862, the SFT-init to RL-final improvements, and the E-c Jaccard/micro-F1 changes) are measured on the held-out official test set, and no parameter is fitted to that test set and then renamed as a prediction. The shared tolerance between the construction verifier and the GRPO reward is disclosed and is a consistency property, not a circular reduction: the official metrics (Pearson, Jaccard, micro/macro-F1) are distinct from the binary tolerance reward, so the training objective and the reported evaluation are not the same quantity by construction. The comparison to EmoLLM baselines does rely on numbers 'reproduced verbatim from [14]', a paper with overlapping authors; however, [14] is a published peer-reviewed benchmark, and the baseline values are externally checkable. The base-model confound (Qwen3-8B versus older EmoLLM backbones) is a validity and comparability concern, not a circularity. The one genuine definitional element is the Naturalisation ablation, where 100% answer-source gold consistency is true by definition because the answer is replaced with the gold label; the paper is transparent about this ('correct by construction'), and the result is not load-bearing for the core SFT/GRPO comparison. Overall score 2 reflects this minor self-definitional ablation, not a circular central claim.

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

No new physical or conceptual entities are postulated; NTDH is a pipeline, not a new entity. The free parameters are the hand-chosen tolerances, gates, hint bands, and split seed that define the data construction and reward. The axioms are the dataset, domain-theory, and optimization assumptions that the pipeline rests on.

free parameters (4)
  • tolerance tau = 0.05 = 0.05 absolute error for regression; exact match for V-oc; exact set match for E-c
    Hand-chosen construction and reward threshold, not derived from the official SemEval metrics (Pearson, Jaccard, F1). It defines which traces are kept for SFT and which rewards are given in GRPO, so it directly shapes the reported results.
  • E-c construction gate F1 >= 0.7 = 0.7
    Hand-chosen permissive threshold for multi-label traces; accepted traces may not justify the full gold label set, a limitation the paper acknowledges in sections 3.3 and 6.
  • hint band cut-offs = not specified in the paper, deferred to released code
    Directional hint categories (slightly high, high, close, and similar bands) depend on hand-set error-magnitude thresholds; these affect refinement behavior and are not reported.
  • 50/50 split seed = 42
    Arbitrary fixed seed determines half-A and half-B composition, and therefore the SFT and RL routing. A different seed may change results.
assumptions (4)
  • domain assumption SemEval-2018 Task 1 gold labels are reliable ground truth for supervised training and evaluation.
    The pipeline and all metrics treat the benchmark annotations as correct targets.
  • domain assumption Affective-science heuristics (valence shifters, irony reversal, emotion co-occurrence, circumplex) are valid guides for refining reasoning traces.
    Domain-aware refinement strategies in section 3.4 rely on psychological theories to steer CoT revision; if these heuristics are wrong for the data, refined traces may be biased.
  • domain assumption Reward optimization with a binary tolerance-matched reward transfers to the official continuous and multi-label metrics.
    GRPO optimizes accuracy within tau, while final evaluation uses Pearson, Jaccard, and F1. The paper acknowledges this decoupling in section 6.
  • domain assumption LLM-generated CoT traces, with gold answers attached, are valid supervision for teaching reasoning.
    The SFT stage assumes the synthesized traces improve the policy's reasoning rather than teaching spurious patterns.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NTDH: Complex Reasoning for Comprehensive Affective Analysis." pith.science (2026). https://pith.science/paper/JFDKA6GS

@misc{pith2026260806425,
  author       = {Pith},
  title        = {Pith review of: NTDH: Complex Reasoning for Comprehensive Affective Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JFDKA6GS}},
  note         = {Machine review of arXiv:2608.06425}
}
read the original abstract

Comprehensive affective analysis is challenging for two reasons: it spans heterogeneous prediction tasks with continuous, ordinal, and multi-label outputs, and affective meaning is context-dependent, requiring conflicting cues to be reconciled rather than mapped directly to labels. Existing methods learn this mapping directly and do not model the reconciliation explicitly. We recast the task as a complex-reasoning problem, which yields one output interface across heterogeneous label spaces and a trajectory over which a verifiable reward can be optimised; to our knowledge, this is the first such treatment covering both sentiment and emotion. The obstacle is on the data side: affective reasoning traces must be synthesised, and generic synthesis is misaligned with the targets, tolerances, and phenomena of affect, and discards or leaks its failure cases. We propose NTDH, which addresses these four failures. Naturalisation sets the training answer to the gold label, so it is correct by construction. A Tolerance-aware gate checks each answer against the task's own scoring margin. Domain-aware strategies refine the reasoning using ideas from affective science. Directional Hints report only the type and direction of an error, without exposing the target. We train Qwen3-8B with SFT and then GRPO under the same tolerance used for verification (up to a more permissive construction gate on the multi-label subtask), and a component ablation quantifies the data-quality effect of each part. Using 16,302 training records, about 14x fewer than comparable instruction-tuned systems, the final policy improves over its SFT checkpoint on five of six official-test metrics and achieves the strongest EI-reg result among the compared systems, at a Pearson correlation of 0.862.

Figures

Figures reproduced from arXiv: 2608.06425 by the authors.

Figure 1
Figure 1. Overview of the proposed approach. NTDH turns raw SemEval-2018 data into verified, reasoning-augmented su￾pervision through four synthesis stages. Naturalisation (N), tolerance-aware verification (T), domain-aware refinement (D) and directional hints (H) are realised in Stages 2–3. Converged traces supervise SFT; unconverged hard cases and half-B supply GRPO, yielding the final affective reasoning policy in one inte… view at source ↗
Figure 2
Figure 2. Inner loop of Stage 3 for a single sample. The refinement cycle (c → b) may iterate up to 𝐷 × 𝐴 times (where 𝐷 is the search depth and 𝐴 the number of restart attempts) before the fallback path (d) is taken. A prediction is counted correct under a task-matched strict criterion 𝜏: absolute error | ̂𝑦 − 𝑦| ≤ 0.05 for the regression subtasks, exact class match for V-oc, and exact set match for E-c ( [PITH_FULL_IMAGE:f… view at source ↗
Figure 3
Figure 3. A worked example of one NTDH synthesis loop (illustrative). An ironic tweet is first read literally as positive. The tolerance-aware verifier (T) computes the task-specific error and rejects the trace; the directional hint (H) reports only that the prediction is too positive, without exposing the gold value; and Backtracking (D) then revisits the literal reading and detects the irony. The corrected trace passes veri… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 29 canonical work pages

  1. [14]

    Z. Liu, K. Yang, Q. Xie, T. Zhang, S. Ananiadou, EmoLLMs: A series of emotional large language models and annotation tools for comprehensive affective analysis, in: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’24), ACM, Barcelona, Spain, 2024, pp. 5487–5496. doi:10.1145/3637528.3671552

    arXiv 2024

  2. [1]

    Zhang, X

    Y. Zhang, X. Yang, X. Xu, Z. Gao, Y. Huang, S. Mu, S. Feng, D. Wang, Y. Zhang, K. Song, G. Yu, Affective computing in the era of large language models: A survey from the NLP perspective, arXiv preprint arXiv:2408.04638 (2024)

    arXiv 2024

  3. [2]

    Munezero, C

    M. Munezero, C. S. Montero, E. Sutinen, J. Pajunen, Are they different? Affect, feeling, emotion, sentiment, and opinion detection in text, IEEE Transactions on Affective Computing 5 (2014) 101–111

    work page 2014

  4. [3]

    Kennedy, D

    A. Kennedy, D. Inkpen, Sentiment classification of movie reviews using contextual valence shifters, Computational Intelligence 22 (2006) 110–125

    work page 2006

  5. [4]

    Wilson, J

    T. Wilson, J. Wiebe, P. Hoffmann, Recognizing contextual polarity: An exploration of features for phrase-level sentiment analysis, Computational Linguistics 35 (2009) 399–433

    work page 2009

  6. [5]

    Reyes, P

    A. Reyes, P. Rosso, D. Buscaldi, From humor recognition to irony detection: The figurative language of social media, Data & Knowledge Engineering 74 (2012) 1–12

    work page 2012

  7. [6]

    Prochnow, J

    A. Prochnow, J. E. Bendler, C. Lange, F. I. Tzavellos, B. M. Goritzer, M. ten Thij, R. Batista-Navarro, IDEM: The IDioms with EMotions datasetforemotionrecognition,in:Proceedingsofthe2024JointInternationalConferenceonComputationalLinguistics,LanguageResources and Evaluation (LREC-COLING), 2024. URL:https://aclanthology.org/2024.lrec-main.752/

    work page 2024

  8. [7]

    J. A. Russell, A circumplex model of affect., Journal of personality and social psychology 39 (1980) 1161

    1980

Show all 45 references
  1. [8]

    A. Mill, L. Kööts-Ausmees, J. Allik, A. Realo, The role of co-occurring emotions and personality traits in anger expression, Frontiers in psychology 9 (2018) 123

    2018

  2. [9]

    Jabreel, A

    M. Jabreel, A. Moreno, A deep learning-based approach for multi-label emotion classification in tweets, Applied Sciences 9 (2019) 1123

    2019

  3. [10]

    S. M. Mohammad, F. Bravo-Marquez, M. Salameh, S. Kiritchenko, SemEval-2018 task 1: Affect in tweets, in: Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), 2018, pp. 1–17. doi:10.18653/v1/S18-1001

    2018 doi

  4. [11]

    P. D. Turney, Thumbs up or thumbs down? semantic orientation applied to unsupervised classification of reviews, in: Proceedings of ACL, 2002, pp. 417–424

    2002

  5. [12]

    B.Pang,L.Lee,S.Vaithyanathan, Thumbsup?sentimentclassificationusingmachinelearningtechniques, in:ProceedingsofEMNLP,2002, pp. 79–86. doi:10.3115/1118693.1118704

    2002

  6. [13]

    C. J. Hutto, E. Gilbert, VADER: A parsimonious rule-based model for sentiment analysis of social media text, in: Proceedings of the International AAAI Conference on Web and Social Media (ICWSM), volume 8, 2014, pp. 216–225

    2014

  7. [15]

    DeepSeek-AI,DeepSeek-R1:IncentivizingreasoningcapabilityinLLMsviareinforcementlearning,arXivpreprintarXiv:2501.12948(2025)

    2025 arXiv

  8. [16]

    Cobbe, V

    K. Cobbe, V. Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, et al., Training verifiers to solve math word problems, arXiv preprint arXiv:2110.14168 (2021)

    2021 arXiv

  9. [17]

    J. Chen, Z. Cai, K. Ji, X. Wang, W. Liu, R. Wang, J. Hou, B. Wang, HuatuoGPT-o1: Towards medical complex reasoning with LLMs, arXiv preprint arXiv:2412.18925 (2024)

    2024 arXiv

  10. [18]

    E.Zelikman,Y.Wu,J.Mu,N.Goodman, Star:Bootstrappingreasoningwithreasoning, AdvancesinNeuralInformationProcessingSystems 35 (2022) 15476–15488

    2022

  11. [19]

    Hsieh, C.-L

    C.-Y. Hsieh, C.-L. Li, C.-K. Yeh, H. Nakhost, Y. Fujii, A. Ratner, R. Krishna, C.-Y. Lee, T. Pfister, Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes, in: Findings of the Association for Computational Linguistics: A...

    2023

  12. [20]

    Yang, et al., Qwen3 technical report, arXiv preprint arXiv:2505.09388 (2025)

    A. Yang, et al., Qwen3 technical report, arXiv preprint arXiv:2505.09388 (2025)

    2025 arXiv

  13. [21]

    Z.Shao,P.Wang,Q.Zhu,R.Xu,J.Song,X.Bi,H.Zhang,M.Zhang,Y.Li,Y.Wu,D.Guo,DeepSeekMath:Pushingthelimitsofmathematical reasoning in open language models, arXiv preprint arXiv:2402.03300 (2024)

    2024 arXiv

  14. [22]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, in: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, v...

    2019

  15. [23]

    Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, V. Stoyanov, Roberta: A robustly optimized bert pretraining approach, arXiv preprint arXiv:1907.11692 (2019). Page 15 of 16 NTDH for Comprehensive Affective Analysis

    2019 arXiv

  16. [24]

    D. Q. Nguyen, T. Vu, A.-T. Nguyen, Bertweet: A pre-trained language model for english tweets, in: Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, 2020, pp. 9–14

    2020

  17. [25]

    3695–3706

    D.Yin,T.Meng,K.-W.Chang,Sentibert:Atransferabletransformer-basedarchitectureforcompositionalsentimentsemantics,in:Proceedings of the 58th annual meeting of the association for computational linguistics, 2020, pp. 3695–3706

    2020

  18. [26]

    V.Duppada,R.Jain,S.Hiray,SeerNetatSemEval-2018task1:Domainadaptationforaffectintweets,in:Proceedingsofthe12thInternational Workshop on Semantic Evaluation (SemEval-2018), 2018, pp. 18–23. doi:10.18653/v1/S18-1002

    2018 doi

  19. [27]

    Baziotis, N

    C. Baziotis, N. Athanasiou, A. Chronopoulou, A. Kolovou, G. Paraskevopoulos, N. Ellinas, S. Narayanan, A. Potamianos, NTUA-SLP at SemEval-2018 task 1: Predicting affective content in tweets with deep attentive RNNs and transfer learning, in: Proceedings of the 12th Internation...

    2018 doi

  20. [28]

    Zhang, Y

    W. Zhang, Y. Deng, B. Liu, S. J. Pan, L. Bing, Sentiment analysis in the era of large language models: A reality check, in: Findings of the Association for Computational Linguistics: NAACL, 2024

    2024

  21. [29]

    H. Fei, B. Li, Q. Liu, L. Bing, F. Li, T.-S. Chua, Reasoning implicit sentiment with chain-of-thought prompting, in: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 2023, pp. 1171–1182

    2023

  22. [30]

    S.Lei,G.Dong,X.Wang,K.Wang,R.Qiao,S.Wang, InstructERC:Reformingemotionrecognitioninconversationwithmulti-taskretrieval- augmented large language models, arXiv preprint arXiv:2309.11911 (2023)

    2023 arXiv

  23. [31]

    S. Feng, G. Sun, N. Lubis, W. Wu, C. Zhang, M. Gašić, Affect recognition in conversations using large language models, in: Proceedings of the 25th Annual Meeting of the Special Interest Group on Discourse and Dialogue, 2024, pp. 259–273

    2024

  24. [32]

    J. Wu, Y. Shen, Z. Zhang, L. Cai, Enhancing large language model with decomposed reasoning for emotion cause pair extraction, arXiv preprint arXiv:2401.17716 (2024)

    2024 arXiv

  25. [33]

    Cheng, Q

    L. Cheng, Q. Shao, C. Zhao, S. Bi, G.-A. Levow, TEII: Think, explain, interact and iterate with large language models to solve cross-lingual emotiondetection,in:Proceedingsofthe14thWorkshoponComputationalApproachestoSubjectivity,Sentiment,andSocialMediaAnalysis, 2024, pp. 495–504

    2024

  26. [34]

    Y. Chen, S. Yan, S. Liu, Y. Li, Y. Xiao, EmotionQueen: A benchmark for evaluating empathy of large language models, in: Findings of the Association for Computational Linguistics: ACL 2024, 2024, pp. 2149–2176

    2024

  27. [35]

    20467–20485

    N.Deas,E.Turcan,I.E.PerezMejia,K.McKeown,MASIVE:Open-endedaffectivestateidentificationinenglishandspanish,in:Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024, pp. 20467–20485

    2024

  28. [36]

    Saeedi, A

    D. Saeedi, A. Kheirandish, S. Saeedi, H. Sahour, A. Panahi, I. Naeeni, GT-NLP at SemEval-2025 task 11: EmoRationale, evidence-based emotiondetectionviaretrieval-augmentedgeneration, in:Proceedingsofthe19thInternationalWorkshoponSemanticEvaluation(SemEval- 2025), 2025, pp. 640–650

    2025

  29. [37]

    Bonard, G

    C. Bonard, G. Cortal, Improving language models for emotion analysis: Insights from cognitive science, in: Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics, 2024, pp. 264–277

    2024

  30. [38]

    J. Wei,X. Wang,D. Schuurmans,M. Bosma, B.Ichter, F.Xia, E. H.Chi, Q.V. Le,D. Zhou, Chain-of-thought promptingelicits reasoningin large language models, in: Advances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  31. [39]

    Kojima, S

    T. Kojima, S. S. Gu, M. Reid, Y. Matsuo, Y. Iwasawa, Large language models are zero-shot reasoners, in: Advances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  32. [40]

    X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdhery, D. Zhou, Self-consistency improves chain of thought reasoning in language models, in: International Conference on Learning Representations (ICLR), 2023

    2023

  33. [41]

    A.Jaech,A.Kalai,A.Lerer,A.Richardson,A.El-Kishky,A.Low,A.Helyar,A.Madry,A.Beutel,A.Carney,etal., OpenAIo1systemcard, arXiv preprint arXiv:2412.16720 (2024)

    2024 arXiv

  34. [42]

    Ouyang, J

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, et al., Training language models to follow instructions with human feedback, in: Advances in Neural Information Processing Systems, volume 35, Curran Associates, Inc., 2022, pp. 27730–27744. URL:https: //proceedings.neu...

    2022

  35. [43]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, O. Klimov, Proximal policy optimization algorithms, arXiv preprint arXiv:1707.06347 (2017)

    2017 arXiv

  36. [44]

    URL:https://github.com/huggingface/open-r1

    Hugging Face, Open R1: A fully open reproduction of DeepSeek-R1, 2025. URL:https://github.com/huggingface/open-r1

    2025

  37. [45]

    von Werra, Y

    L. von Werra, Y. Belkada, L. Tunstall, E. Beeching, T. Thrush, N. Lambert, S. Huang, K. Rasul, Q. Gallouédec, TRL: Transformers reinforcement learning, 2020. URL:https://github.com/huggingface/trl. Page 16 of 16

    2020

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.