Pith. sign in

REVIEW 3 major objections 4 minor 36 references

This paper claims that commonsense knowledge in empathetic dialogue should be coordinated dynamically — adapted when recognizing the user's emotion, when filtering which commonsense relations matter, and at every token of the reply — rather

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

2026-08-01 05:39 UTC pith:74X25CWN

load-bearing objection An honest CEM extension whose diversity gains are confounded by a changed FACE loss weight; the LLM-based quality gains are the more credible evidence. the 3 major comments →

arxiv 2607.22136 v1 pith:74X25CWN submitted 2026-07-24 cs.CL

Dynamic Commonsense Coordination for Empathetic Response Generation

Zhengyu Qi This is my paper
classification cs.CL
keywords empathetic response generationcommonsense reasoningemotion recognitionresponse diversitydynamic commonsense coordinationiterative decodingattention-residual fusionEmpathetic-Dialogues
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.

Empathetic chatbots need to do two different things with commonsense: understand what a user is feeling and why, and then craft a response that uses that understanding. The paper argues that prior systems build commonsense once and reuse the same static representation for both jobs, letting irrelevant knowledge add noise. To fix this, it proposes a Dynamic Commonsense Coordination framework with three modules: one that blends commonsense derived from the dialogue and from the situation description, one that down-weights low-relevance commonsense relations before decoding, and one that re-retrieves commonsense at every decoding step. On Empathetic-Dialogues, the framework raises emotion classification accuracy from 39.11% to 46.09% and lifts lexical diversity (Dist-1 from 0.66 to 1.03, Dist-2 from 2.99 to 4.93) over the CEM baseline while keeping perplexity essentially unchanged (36.13 vs. 36.11). A blind LLM comparison also favors the new framework on relevance, coherence, and informativeness, though not significantly on empathy or overall quality.

Core claim

The central claim is that commonsense about a dialogue and commonsense about its situation are complementary streams that should be coordinated rather than pooled. DCC coordinates them at three levels: SCE-AttnRes uses cross-attention from situation commonsense to dialogue commonsense and then adaptively weights the original two sources and their fusion; AGCF scores each of the five COMET relations (xIntent, xNeed, xWant, xEffect, xReact) against a global dialogue representation and rescales them so low-relevance relations are suppressed; ICAD uses the evolving decoder hidden state as a query to retrieve from these filtered relation memories at each generation step. The paper reports that th

What carries the argument

The load-bearing object is the DCC framework: a pipeline in which COMET-generated commonsense is split into dialogue-derived and situation-derived memory streams, then coordinated by three modules. SCE-AttnRes (residual-based commonsense interaction) fuses the two streams through cross-attention and a learned weighted average with the originals, producing representations for emotion classification and dialogue refinement. AGCF (Association-Guided Commonsense Filtering) treats the five relation vectors as a memory, scores each relation's association with the dialogue, and rescales softmax weights by the number of relations to keep their average scale. ICAD (Iterative Commonsense-Aware Decodin

Load-bearing premise

The load-bearing premise is that the reported accuracy and diversity gains reflect the three new modules rather than single-run training variation, given that the paper states its results lack statistical testing and repeated-seed confirmation.

What would settle it

Re-run the main comparison and the ablated variants over multiple random seeds (for example, five or more) and check whether the roughly seven-point accuracy gain over CEM and the Dist-1 doubling reproduce consistently; if the gains collapse or fall within seed-to-seed noise, the central claim is falsified.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the framework works as claimed, stage-aware commonsense coordination can raise emotion classification accuracy by roughly seven points over the CEM baseline without hurting perplexity.
  • Lexical diversity of generated empathetic responses can be more than doubled in Dist-1 and raised from about 3 to about 5 in Dist-2, as a side effect of coordinated retrieval rather than a separate diversity trick.
  • The three modules have separable effects: cross-source fusion is the main driver of emotion recognition; relation-level filtering mainly protects generation quality; token-level retrieval improves likelihood and emotion prediction while constraining surface diversity.
  • Blind LLM evaluation indicates the main response-level gains are in relevance, coherence, and informativeness, while empathy and overall preference are not significantly different from the baseline.
  • The added cost is modest: about 10% more parameters and 33% more decoding latency, still around nine responses per second on one GPU.

Where Pith is reading between the lines

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

  • Editorial inference: if dynamic commonsense coordination is the active ingredient, the same principle should transfer to LLM-based empathetic generation, where commonsense is often retrieved once and frozen; a testable extension is to gate or re-retrieve knowledge at each generation step.
  • Editorial inference: AGCF's strong down-weighting of xReact during decoding, combined with unchanged or improved emotion classification, suggests the framework is separating emotion recognition from emotion expression; a testable hypothesis is that suppressing affective commonsense before the decoder improves fluency at the cost of expressed warmth.
  • Editorial inference: because the main comparison is single-run, the safest reading is that the effect size is unconfirmed; repeating the same protocol across random seeds and reporting confidence intervals would settle whether the architecture, not the seed, drives the difference.
  • Editorial inference: the paper itself notes it does not beat the strongest recent association-based models on automatic metrics, so the contribution is best read as an architectural diagnosis of where commonsense coordination helps, not as a new state of the art.

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

3 major / 4 minor

Summary. The paper proposes DCC, an extension of the CEM commonsense-aware empathetic response generation model. DCC adds three mechanisms: SCE-AttnRes, which coordinates COMET-based commonsense generated from the dialogue and the situation descriptions; AGCF, which reweights the five commonsense relations by relevance to the dialogue; and ICAD, which retrieves from the filtered relation memories at each decoding step. The model is trained with the same multi-task objective as CEM, including a frequency-aware cross-entropy diversity loss, but with a larger diversity weight (γ=2.0 vs CEM's 1.5). On Empathetic-Dialogues, the paper reports that DCC improves emotion classification accuracy (39.11→46.09) and response diversity (Dist-1 0.66→1.03, Dist-2 2.99→4.93) while maintaining perplexity (36.11 vs 36.13). Additional ablations, LLM-based pairwise evaluation with double-order consistency, human evaluation, and efficiency measurements are reported in the main text and appendices.

Significance. If the reported effects are robust, DCC would provide a useful analysis of where commonsense coordination matters in empathetic dialogue: at cross-source fusion, relation filtering, and token-level retrieval. The paper is transparent about its scope: it does not claim state-of-the-art results, acknowledges the limitations of COMET and of automatic evaluation, and includes an LLM evaluation with double-order consistency checking and Holm correction, which is above the field norm. The AGCF and ICAD visualizations are also informative, and the paper explicitly states that AGCF scoring does not use the target response and that ICAD attention uses only generated tokens. However, the central empirical claim is currently not cleanly established: the main comparison to CEM is confounded by the different diversity-loss weight, and the automatic results are single-run with no variance or significance testing. These issues are fixable but load-bearing.

major comments (3)
  1. [§4.5, Eq. (32) and Table 1] The diversity comparison to CEM is confounded. DCC uses γ=2.0 for the FACE loss, while CEM uses γ=1.5. Since L_div directly penalizes frequent tokens, increasing γ by 33% is itself expected to raise Dist-1/Dist-2 regardless of the proposed modules. The abstract and Section 5.2 attribute the Dist-1/2 gains (0.66→1.03 and 2.99→4.93) to the DCC framework, but the ablations in Table 2 all keep γ=2.0, so they cannot separate module effects from the loss-weight change. Please add matched-γ experiments—e.g., DCC with γ=1.5 and/or CEM with γ=2.0—and report Acc and PPL alongside Dist-1/Dist-2. Without this control, the headline diversity claim is not supported.
  2. [§5.2, Table 1 and §5.3, Table 2] All automatic results are single-run. The Limitations section acknowledges this for the ablations, but the same limitation applies to the main DCC-vs-CEM comparison. In Table 2, removing AGCF raises Acc from 46.09 to 46.89, and removing ICAD raises Dist-2 from 4.93 to 5.59, so several module effects are smaller than plausible seed-to-seed variation. Please report means and standard deviations over at least 3–5 seeds, with paired or bootstrap significance tests for the DCC-vs-CEM comparison and the ablations. This is necessary to support the causal claims made for each module.
  3. [§5.2, Table 1 and Appendix H] The baseline row for CEM in Table 1 is marked with †, indicating that the number is taken from the original CEM paper, whereas DCC is the authors' own run. The LLM evaluation in Appendix H correctly compares DCC against a reproduced CEM under the same greedy decoding protocol; the automatic table should do the same. Please include a same-protocol reproduced CEM row in Table 1 and clarify which baselines are from original papers and which are locally reproduced. Without this, the claimed gains over CEM could be affected by decoding setting, checkpoint selection, or evaluation code differences.
minor comments (4)
  1. [§4.3, Eq. (24)] The rescaling by |R|=5 is a fixed design choice, but no sensitivity analysis or justification is provided beyond preventing uniform shrinkage. AGCF's test-set mean weights show very strong suppression of xReact (≈0.05–0.11); a sentence explaining whether this is intended and whether the rescale factor affects conclusions would help.
  2. [§5.4 and Appendix I] The human evaluation shows non-significant preferences (e.g., empathy p=0.278, κ=0.064), yet the conclusion says 'consistent preference toward DCC, particularly for empathy.' Consider softening this language to match the statistical evidence, e.g., 'numerically preferred but not statistically significant.'
  3. [Appendix G.2] In Case 250 the dialogue and situation are identical, and the different ICAD attention patterns are attributed to separate AGCF scoring heads and ICAD attention heads. This is a useful illustration of source-specific processing, but it is not evidence of utility; the surrounding text acknowledges this, which is good.
  4. [References] Several references use non-standard author fields (e.g., 'Team, 2026', 'DeepSeek-AI, 2025'). Please check the venue metadata and provide full author lists where available, and ensure all arXiv identifiers are consistent with the submission date.

Circularity Check

0 steps flagged

No circularity: DCC's claims rest on held-out evaluations and disclosed hyperparameters; the gamma/FACE mismatch is a statistical confound, not a self-referential derivation.

full rationale

None of the paper's load-bearing claims is obtained by definition or by self-citation. DCC's improvements are empirical results from the held-out Empathetic-Dialogues test set (Table 1), a blind LLM-based pairwise evaluation (Appendix H), and a human preference study (Appendix I), all compared against an externally implemented CEM baseline. The mechanism analyses are not fitted to the target: Appendix F states that AGCF weights 'are computed from the dialogue, situation, and their commonsense memories only; the target response does not influence AGCF scoring,' and Appendix G states that ICAD attention 'is computed from the preceding decoder state; neither the target response nor future generated tokens are used.' The single self-adjacent citation, to 'Team 2026' on attention residuals, is an external design inspiration rather than the author's own prior work, and it is not used as a uniqueness theorem or as a substitute for evidence. The acknowledged limitations are statistical rather than circular: the Limitations section says 'the ablation results are single-run and lack statistical testing,' and Section 4.5 discloses 'we use γ=2.0 in our experiments, while the original CEM uses 1.5.' The γ mismatch is a genuine confound for the Dist-1/2 comparisons and should be addressed with a matched-γ control, but raising a disclosed diversity-loss weight is not a derivation that reduces the reported diversity metric to its own input. No equation is shown equal to another by construction, and no fitted parameter is relabeled as a prediction. The central comparison remains externally anchored to test data, so there is no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No new physical or conceptual entities are postulated. SCE-AttnRes, AGCF, and ICAD are architectural modules, not invented entities. The central claim rests mainly on benchmark evaluation and the domain assumption that fixed COMET relations adequately represent commonsense.

free parameters (3)
  • diversity loss weight γ = 2.0
    Set in Section 4.5; CEM uses 1.5. Directly affects the Dist-1/2 and PPL trade-off in the main claim, with no sensitivity analysis reported.
  • AGCF softmax rescale factor |R| = 5
    Equation 24 multiplies softmax weights by 5 to preserve scale; fixed by design but governs filtering strength.
  • training hyperparameters (LR 1e-4, batch 16, beam 5, seed 42, 1-layer transformer) = as listed in Appendix C
    Inherited from CEM, not tuned for DCC, but still model choices that affect the measured metrics.
axioms (4)
  • domain assumption COMET's five relation types (xIntent, xNeed, xWant, xEffect, xReact) are sufficient carriers of commonsense for empathetic reasoning
    Section 3.1 builds all reasoning on these relations; the Limitations section admits outputs can be noisy or biased and knowledge outside this fixed set is inaccessible.
  • domain assumption Empathetic-Dialogues situation descriptions and dialogue context provide enough signal for emotion recognition and response quality
    Task formulation in Section 3.1; the Limitations section notes generalization to other languages/domains is unexamined.
  • domain assumption LLM-based blind evaluation with DeepSeek-V3.2 approximates response quality
    Appendix H uses one evaluator model and cautions that judgments are not equivalent to human evaluation.
  • standard math Standard softmax attention and cross-attention operations preserve useful information
    Used in Equations 12, 16-17, 24, 26-28; standard neural machinery.

pith-pipeline@v1.3.0-alltime-deepseek · 20121 in / 10386 out tokens · 108803 ms · 2026-08-01T05:39:07.869985+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Dynamic Commonsense Coordination for Empathetic Response Generation." pith.science (2026). https://pith.science/paper/74X25CWN

@misc{pith2026260722136,
  author       = {Pith},
  title        = {Pith review of: Dynamic Commonsense Coordination for Empathetic Response Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/74X25CWN}},
  note         = {Machine review of arXiv:2607.22136}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Empathetic Response Generation (ERG) requires models to recognize users' emotions and generate empathetic responses. Commonsense knowledge has been shown to support such reasoning, yet existing approaches typically reuse fixed commonsense representations across understanding and generation, limiting their ability to coordinate such knowledge across different stages. We propose DCC, a Dynamic Commonsense Coordination Framework with three complementary modules: residual-based commonsense interaction (SCE-AttnRes) to integrate contextual and situational commonsense representations, Association-Guided Commonsense Filtering (AGCF) to down-weight low-relevance commonsense relations, and Iterative Commonsense-Aware Decoding (ICAD) to dynamically retrieve commonsense memories during generation. Experiments on the Empathetic-Dialogues benchmark show that DCC improves emotion classification accuracy and response diversity over the CEM baseline while maintaining comparable perplexity. An LLM-based blind evaluation further demonstrates that DCC generates responses with better relevance, coherence, and informativeness. The code and implementation details will be publicly available at https://github.com/Hanabi-Q/DCC-ERG.

Figures

Figures reproduced from arXiv: 2607.22136 by Zhengyu Qi.

Figure 1
Figure 1. Figure 1: An example from the Empathetic-Dialogues [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed Dynamic Commonsense Coordination Framework. The framework coordinates [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the SCE-AttnRes module for cross-source commonsense coordination. Given dialogue [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of AGCF. Using the global dialogue representation and relation-level commonsense memories, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of ICAD. At each decoding step, the decoder hidden state retrieves relevant information from [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A dialogue example from the EmpatheticDia [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: COMET-generated commonsense inferences for both the dialogue context and the corresponding situation description. • xIntent: PersonX’s likely intention before the event. • xNeed: PersonX’s likely prerequisite or need before the event. • xWant: PersonX’s likely desire or goal after the event. • xEffect: The likely effect of the event on Per￾sonX. • xReact: PersonX’s likely emotional reaction after the event… view at source ↗
Figure 8
Figure 8. Figure 8: AGCF relation weights for Case 3520 and the corresponding test-set means. Dialogue- and situation [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: AGCF relation weights for Case 1724 and the corresponding test-set means. The dialogue contains both [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Token-level ICAD attention for Case 3550. The upper and lower heatmaps show attention over dialogue [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Token-level ICAD attention for Case 250. Although the dialogue and situation contain the same text, [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

36 extracted references · 10 canonical work pages

  1. [1]

    , author=

    Measuring individual differences in empathy: Evidence for a multidimensional approach. , author=. Journal of personality and social psychology , volume=. 1983 , publisher=

    1983

  2. [2]

    The Psychological Record , volume=

    Cognitive empathy and emotional empathy in human behavior and evolution , author=. The Psychological Record , volume=. 2006 , publisher=

    2006

  3. [3]

    Annals of the new York Academy of Sciences , volume=

    The social neuroscience of empathy , author=. Annals of the new York Academy of Sciences , volume=. 2009 , publisher=

    2009

  4. [4]

    Towards Empathetic Open-domain Conversation Models:

    Hannah Rashkin and Eric Michael Smith and Margaret Li and Y. Towards Empathetic Open-domain Conversation Models:. Proceedings of the 57th Conference of the Association for Computational Linguistics,. 2019 , url =. doi:10.18653/V1/P19-1534 , timestamp =

    doi:10.18653/v1/p19-1534 2019

  5. [5]

    MoEL: Mixture of Empathetic Listeners , booktitle =

    Zhaojiang Lin and Andrea Madotto and Jamin Shin and Peng Xu and Pascale Fung , editor =. MoEL: Mixture of Empathetic Listeners , booktitle =. 2019 , url =. doi:10.18653/V1/D19-1012 , timestamp =

    work page doi:10.18653/v1/d19-1012 2019

  6. [6]

    Gelbukh and Rada Mihalcea and Soujanya Poria , editor =

    Navonil Majumder and Pengfei Hong and Shanshan Peng and Jiankun Lu and Deepanway Ghosal and Alexander F. Gelbukh and Rada Mihalcea and Soujanya Poria , editor =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing,. 2020 , url =. doi:10.18653/V1/2020.EMNLP-MAIN.721 , timestamp =

    doi:10.18653/v1/2020.emnlp-main.721 2020

  7. [7]

    EmpDG: Multi-resolution Interactive Empathetic Dialogue Generation , booktitle =

    Qintong Li and Hongshen Chen and Zhaochun Ren and Pengjie Ren and Zhaopeng Tu and Zhumin Chen , editor =. EmpDG: Multi-resolution Interactive Empathetic Dialogue Generation , booktitle =. 2020 , url =. doi:10.18653/V1/2020.COLING-MAIN.394 , timestamp =

    work page doi:10.18653/v1/2020.coling-main.394 2020

  8. [8]

    Thirty-Fifth

    Peixiang Zhong and Di Wang and Pengfei Li and Chen Zhang and Hao Wang and Chunyan Miao , title =. Thirty-Fifth. 2021 , url =. doi:10.1609/AAAI.V35I16.17713 , timestamp =

    work page doi:10.1609/aaai.v35i16.17713 2021

  9. [9]

    Thirty-Sixth

    Sahand Sabour and Chujie Zheng and Minlie Huang , title =. Thirty-Sixth. 2022 , url =. doi:10.1609/AAAI.V36I10.21373 , timestamp =

    doi:10.1609/aaai.v36i10.21373 2022

  10. [10]

    Thirty-Sixth

    Qintong Li and Piji Li and Zhaochun Ren and Pengjie Ren and Zhumin Chen , title =. Thirty-Sixth. 2022 , url =. doi:10.1609/AAAI.V36I10.21347 , timestamp =

    work page doi:10.1609/aaai.v36i10.21347 2022

  11. [11]

    Empathetic Dialogue Generation via Sensitive Emotion Recognition and Sensible Knowledge Selection , booktitle =

    Lanrui Wang and Jiangnan Li and Zheng Lin and Fandong Meng and Chenxu Yang and Weiping Wang and Jie Zhou , editor =. Empathetic Dialogue Generation via Sensitive Emotion Recognition and Sensible Knowledge Selection , booktitle =. 2022 , url =. doi:10.18653/V1/2022.FINDINGS-EMNLP.340 , timestamp =

    work page doi:10.18653/v1/2022.findings-emnlp.340 2022

  12. [12]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),

    Jinfeng Zhou and Chujie Zheng and Bo Wang and Zheng Zhang and Minlie Huang , editor =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),. 2023 , url =. doi:10.18653/V1/2023.ACL-LONG.457 , timestamp =

    doi:10.18653/v1/2023.acl-long.457 2023

  13. [13]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,

    Fengyi Fu and Lei Zhang and Quan Wang and Zhendong Mao , editor =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,. 2023 , url =. doi:10.18653/V1/2023.EMNLP-MAIN.653 , timestamp =

    work page doi:10.18653/v1/2023.emnlp-main.653 2023

  14. [14]

    Improving Empathetic Dialogue Generation by Dynamically Infusing Commonsense Knowledge , booktitle =

    Hua Cai and Xuli Shen and Qing Xu and Weilin Shen and Xiaomei Wang and Weifeng Ge and Xiaoqing Zheng and Xiangyang Xue , editor =. Improving Empathetic Dialogue Generation by Dynamically Infusing Commonsense Knowledge , booktitle =. 2023 , url =. doi:10.18653/V1/2023.FINDINGS-ACL.498 , timestamp =

    work page doi:10.18653/v1/2023.findings-acl.498 2023

  15. [15]

    Exploiting Emotion-Semantic Correlations for Empathetic Response Generation , booktitle =

    Zhou Yang and Zhaochun Ren and Yufeng Wang and Xiaofei Zhu and Zhihao Chen and Tiecheng Cai and Yunbing Wu and Yisong Su and Sibo Ju and Xiangwen Liao , editor =. Exploiting Emotion-Semantic Correlations for Empathetic Response Generation , booktitle =. 2023 , url =. doi:10.18653/V1/2023.FINDINGS-EMNLP.320 , timestamp =

    work page doi:10.18653/v1/2023.findings-emnlp.320 2023

  16. [16]

    Zhou Yang and Zhaochun Ren and Yufeng Wang and Haizhou Sun and Xiaofei Zhu and Xiangwen Liao , title =. Inf. Process. Manag. , volume =. 2024 , url =. doi:10.1016/J.IPM.2024.103824 , timestamp =

    arXiv 2024

  17. [17]

    An Iterative Associative Memory Model for Empathetic Response Generation , booktitle =

    Zhou Yang and Zhaochun Ren and Yufeng Wang and Haizhou Sun and Chao Chen and Xiaofei Zhu and Xiangwen Liao , editor =. An Iterative Associative Memory Model for Empathetic Response Generation , booktitle =. 2024 , url =. doi:10.18653/V1/2024.ACL-LONG.170 , timestamp =

    work page doi:10.18653/v1/2024.acl-long.170 2024

  18. [18]

    CoRR , volume =

    Xinhao Chen and Chong Yang and Man Lan and Li Cai and Yang Chen and Tu Hu and Xinlin Zhuang and Aimin Zhou , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2408.11599 , eprinttype =. 2408.11599 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2408.11599 2024

  19. [19]

    Fine-Grained Emotion Recognition via In-Context Learning , booktitle =

    Zhaochun Ren and Zhou Yang and Chenglong Ye and Haizhou Sun and Chao Chen and Xiaofei Zhu and Xiangwen Liao , editor =. Fine-Grained Emotion Recognition via In-Context Learning , booktitle =. 2025 , url =. doi:10.1145/3746252.3761319 , timestamp =

    arXiv 2025

  20. [20]

    ReflectDiffu: Reflect between Emotion-intent Contagion and Mimicry for Empathetic Response Generation via a RL-Diffusion Framework , booktitle =

    Jiahao Yuan and Zixiang Di and Zhiqing Cui and Guisong Yang and Usman Naseem , editor =. ReflectDiffu: Reflect between Emotion-intent Contagion and Mimicry for Empathetic Response Generation via a RL-Diffusion Framework , booktitle =. 2025 , url =. doi:10.18653/V1/2025.ACL-LONG.1235 , timestamp =

    work page doi:10.18653/v1/2025.acl-long.1235 2025

  21. [21]

    CoRR , volume =

    Hongru Ji and Yuyin Fan and Meng Zhao and Xianghua Li and Lianwei Wu and Chao Gao , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2604.07100 , eprinttype =. 2604.07100 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.07100 2026

  22. [22]

    Proceedings of the 57th Conference of the Association for Computational Linguistics,

    Antoine Bosselut and Hannah Rashkin and Maarten Sap and Chaitanya Malaviya and Asli Celikyilmaz and Yejin Choi , editor =. Proceedings of the 57th Conference of the Association for Computational Linguistics,. 2019 , url =. doi:10.18653/V1/P19-1470 , timestamp =

    doi:10.18653/v1/p19-1470 2019

  23. [23]

    Hwang and Chandra Bhagavatula and Ronan Le Bras and Jeff Da and Keisuke Sakaguchi and Antoine Bosselut and Yejin Choi , title =

    Jena D. Hwang and Chandra Bhagavatula and Ronan Le Bras and Jeff Da and Keisuke Sakaguchi and Antoine Bosselut and Yejin Choi , title =. Thirty-Fifth. 2021 , url =. doi:10.1609/AAAI.V35I7.16792 , timestamp =

    doi:10.1609/aaai.v35i7.16792 2021

  24. [24]

    Smith and Yejin Choi , title =

    Maarten Sap and Ronan Le Bras and Emily Allaway and Chandra Bhagavatula and Nicholas Lourie and Hannah Rashkin and Brendan Roof and Noah A. Smith and Yejin Choi , title =. The Thirty-Third. 2019 , url =. doi:10.1609/AAAI.V33I01.33013027 , timestamp =

    doi:10.1609/aaai.v33i01.33013027 2019

  25. [25]

    ConceptNet 5.5: An Open Multilingual Graph of General Knowledge , booktitle =

    Robyn Speer and Joshua Chin and Catherine Havasi , editor =. ConceptNet 5.5: An Open Multilingual Graph of General Knowledge , booktitle =. 2017 , url =. doi:10.1609/AAAI.V31I1.11164 , timestamp =

    doi:10.1609/aaai.v31i1.11164 2017

  26. [26]

    Mohammad , editor =

    Saif M. Mohammad , editor =. Obtaining Reliable Human Ratings of Valence, Arousal, and Dominance for 20, 000 English Words , booktitle =. 2018 , url =. doi:10.18653/V1/P18-1017 , timestamp =

    doi:10.18653/v1/p18-1017 2018

  27. [27]

    Gomez and Lukasz Kaiser and Illia Polosukhin , editor =

    Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin , editor =. Attention is All you Need , booktitle =. 2017 , url =

    2017

  28. [28]

    Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,

    Jacob Devlin and Ming. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,. 2019 , url =. doi:10.18653/V1/N19-1423 , timestamp =

    doi:10.18653/v1/n19-1423 2019

  29. [29]

    CoRR , volume =

    Kimi Team , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2603.15031 , eprinttype =. 2603.15031 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2603.15031 2026

  30. [30]

    DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models , journal =

    DeepSeek. DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models , journal =. 2025 , url =. doi:10.48550/ARXIV.2512.02556 , eprinttype =. 2512.02556 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2512.02556 2025

  31. [31]

    A Diversity-Promoting Objective Function for Neural Conversation Models , booktitle =

    Jiwei Li and Michel Galley and Chris Brockett and Jianfeng Gao and Bill Dolan , editor =. A Diversity-Promoting Objective Function for Neural Conversation Models , booktitle =. 2016 , url =. doi:10.18653/V1/N16-1014 , timestamp =

    doi:10.18653/v1/n16-1014 2016

  32. [32]

    Improving Neural Response Diversity with Frequency-Aware Cross-Entropy Loss , booktitle =

    Shaojie Jiang and Pengjie Ren and Christof Monz and Maarten de Rijke , editor =. Improving Neural Response Diversity with Frequency-Aware Cross-Entropy Loss , booktitle =. 2019 , url =. doi:10.1145/3308558.3313415 , timestamp =

    arXiv 2019

  33. [33]

    Kingma and Jimmy Ba , editor =

    Diederik P. Kingma and Jimmy Ba , editor =. Adam:. 3rd International Conference on Learning Representations,. 2015 , url =

    2015

  34. [34]

    Manning , editor =

    Jeffrey Pennington and Richard Socher and Christopher D. Manning , editor =. Glove: Global Vectors for Word Representation , booktitle =. 2014 , url =. doi:10.3115/V1/D14-1162 , timestamp =

    doi:10.3115/v1/d14-1162 2014

  35. [35]

    Miller and Adam Fisch and Jesse Dodge and Amir

    Alexander H. Miller and Adam Fisch and Jesse Dodge and Amir. Key-Value Memory Networks for Directly Reading Documents , booktitle =. 2016 , url =. doi:10.18653/V1/D16-1147 , timestamp =

    doi:10.18653/v1/d16-1147 2016

  36. [36]

    Educational and psychological measurement , volume=

    The equivalence of weighted kappa and the intraclass correlation coefficient as measures of reliability , author=. Educational and psychological measurement , volume=. 1973 , publisher=

    1973