Pith. sign in

REVIEW 4 major objections 6 minor 36 references

M3HG: Multimodal, Multi-scale, and Multi-type Node Heterogeneous Graph for Emotion Cause Triplet Extraction in Conversations

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read M3HG, a multimodal heterogeneous graph with explicit emotion- and cause-specific context nodes, outperforms all seven baselines on both the ECF and new MECAD benchmarks for emotion-cause triplet extraction.

desk verdict A genuinely new dataset and a clearly described graph model, but the headline SOTA numbers are not yet trustworthy because the baselines were not run under matched conditions. read the letter →

arxiv 2508.18740 v1 pith:BSPKBKD3 submitted 2025-08-26 cs.CL cs.AI

classification cs.CLcs.AI
keywords emotioncausetripletextractionmultimodalconversationsheterogeneousgraphattentionnetworkanalysisconversationalrecognitiondatasetMECADM3HG
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that modeling emotion context and cause context as explicit nodes inside a multimodal conversation graph is the key to accurate emotion-cause triplet extraction in conversations, and that this structure is what previous pipeline, graph, and prompt-based models lacked. To support that claim, the authors build MECAD, the first multimodal, multi-scenario Chinese dataset for the task (989 conversations from 56 TV series), reducing the single-show uniformity of the existing ECF dataset. On both benchmarks, their M3HG model reports the highest weighted-average F1 among seven baselines, including the previous state of the art HiLo and a 5-shot LLM baseline. It also claims to be the first MECTEC model that can find cause utterances appearing after the emotion utterance, because every utterance is connected through a conversation-level super-node. The result matters because real conversations often carry the reason for an emotion many turns later and across text, audio, and video, which earlier models were not built to use.

What carries the argument

The central object is M3HG, a multimodal heterogeneous graph attention network whose graph G=(V,E,R) contains three kinds of nodes besides the utterance super-nodes: emotional-context nodes N_e, causal-context nodes N_c, and a conversation super-node N_d, with utterance super-nodes bundling text, audio, and video features. Five relation types (same speaker, different speaker, global, emotion, cause) define the edges, and the network applies node-level and semantic attention over meta-paths to fuse information at intra-utterance and inter-utterance levels; the emotion and cause nodes are the carriers of the explicit context the paper claims earlier models miss.

What would settle it

Score M3HG only on triplets whose emotion and cause utterances are more than K=3 positions apart, a subset MECAD contains with relative positions up to -14 and +13. If M3HG's F1 on this far-apart subset is no better than SHARK's, then the claimed cause-after-emotion capability is not actually carried by the global connection; if it is clearly better, the conversation super-node works as described.

Watch

Extended reading notes

Core claim

The discovery is that constructing a heterogeneous graph with four node types—utterance super-nodes carrying text, audio, and video features, a conversation super-node, emotional-context nodes, and causal-context nodes—and fusing it at two scales is enough to outperform prior approaches. The five edge types (same speaker, different speaker, global, emotion, cause) and the meta-path-based heterogeneous graph attention network allow the model to aggregate context inside an utterance, across utterances, and across the whole conversation. In the authors' framing, the emotional and causal context nodes make the emotion-cause relationship explicit instead of leaving it implicit in a generic uttera

Load-bearing premise

The model relies on a pre-defined set of five edge types and a local window of three preceding utterances being enough to represent every emotion-cause relation in a conversation; if a causal link depends on information outside that window or on a relation the graph does not encode, the model has no way to learn it.

Editorial extensions

If this is right

  • Cause utterances that appear after the emotion utterance are recoverable: the global connection through the conversation super-node gives every utterance a path to every later utterance, and the paper's filtered subset experiments show an F1 gain over SHARK and the LLM baseline on those conversations.
  • Dataset diversity changes what counts as progress: MECAD's multi-scenario, TV-independent setup exposes models to varied speakers, languages, and situations, so methods that overfit a single show will show up as weaker F1 on this benchmark.
  • Both fusion levels are load-bearing: the ablations show that removing the emotion/cause nodes costs the most, and removing intra-utterance fusion costs more than removing inter-utterance fusion, so future models should keep explicit context nodes rather than relying only on generic utterance embeddings.
  • Text-only M3HG already beats all baselines on ECF, which means the graph architecture contributes before multimodal features are added.

Reading between the lines

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

  • Editorial extension: because MECAD additionally labels cause type (event, opinion, emotional influence, self-reflection, greeting) and which modality supported each annotation, a natural next analysis is whether M3HG's errors concentrate in self-reflection or in audio/video-only cues.
  • Editorial extension: the K=3 peak in the reported sensitivity sweep suggests the local window is tuned to these datasets; scoring only far-apart triplets (MECAD contains relative positions up to -14 and +13) would show whether the conversation super-node genuinely handles cause-after-emotion or mainly compensates for a short local window.
  • The paper's own Limitations section names three boundaries: no external knowledge is integrated, input length is capped by the language model, and multimodal fusion can propagate errors when modalities conflict; the reported gains should be read within those boundaries.
  • Editorial extension: since text-only M3HG already beats all baselines on ECF, a controlled comparison with a variant that removes emotion/cause nodes at matched parameter count would isolate whether the graph structure, not the encoders, carries the improvement.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces MECAD, a new Chinese multimodal, multi-scenario MECTEC dataset with 989 conversations from 56 TV series, annotated for emotion categories, cause utterances, cause types, and modal cues. It also proposes M3HG, a heterogeneous-graph model with emotional/causal context nodes, utterance/conversation super-nodes, five edge types, and two-level (intra-utterance and inter-utterance) semantic fusion. The paper reports that M3HG outperforms seven baselines on the ECF dataset (6 Avg F1 40.07 vs. 33.04 for HiLo) and on MECAD (32.82 vs. 27.58 for SHARK), and claims to be the first MECTEC model that can handle cause utterances appearing after emotion utterances. Code and dataset are released.

Significance. If the empirical claims hold, the contribution is substantial: MECAD is a genuinely new resource that addresses the scenario diversity gap of the existing ECF dataset, with a reasonable annotation process (Fleiss's kappa 0.6932) and rich annotations beyond simple triplets. M3HG is a plausible architecture that explicitly models emotion/cause contexts and multi-scale fusion, and the reported ablations in Table 6 show consistent drops when modules are removed. The released code and dataset, if usable, would be valuable to the community. However, the central 'state-of-the-art' claim currently rests on an uncontrolled comparison with borrowed baseline numbers and no variance estimates, so the magnitude and even the direction of the architectural advantage are not yet established.

major comments (4)
  1. [§5.2, Table 2, Appendix D] The headline comparison is uncontrolled. Five baseline rows (ECPE-2D, UECA-Prompt, SHARK on ECF; HiLo on ECF) are transcribed from prior papers, and those papers use their own text encoders and feature extractors, whereas M3HG fine-tunes RoBERTa-large on ECF and chinese-roberta-wwm-ext-large on MECAD, plus Wav2Vec2 and DenseNet. The ablation study in Table 6 holds the M3HG backbone fixed, so it cannot separate gains due to the graph/context-node design from gains due to stronger pretrained encoders. The 5–7 point 6 Avg advantage may therefore be an encoder/feature-extraction artifact. To support the central outperformance claim, the authors should re-run the seven baselines under identical feature extractors, data splits, and evaluation protocols, or provide a matched text-only comparison with the same backbone.
  2. [Table 2] No variance or significance testing is reported. All numbers appear to come from single runs. The MECAD 6 Avg gap over SHARK is 5.24 points, which is not large relative to run-to-run variability in F1 on small test sets. The authors should report mean±std over at least 3–5 seeds and a significance test (e.g., paired bootstrap or approximate randomization) for the main comparisons. This is especially important for the MECAD results, where the strongest baseline is text-only and the gap is modest.
  3. [§4.4, Table 7] The claim that M3HG is 'the first MECTEC model capable of handling situations where cause utterances appear after emotion utterances' is not supported by the evidence. Table 7 compares only SHARK and GPT-4o on a filtered subset; HiLo—the other MECTEC E2E baseline—is absent, and the filtering criterion (how 'cause after emotion' is defined, which pairs are included) is not described. The global connection edge to the conversation node is a plausible mechanism, but a claim of being 'first' requires a controlled comparison against a baseline with a full-context mechanism under the same backbone. The manuscript's own Limitations section further notes that M3HG cannot handle excessively long conversations, which is precisely the regime where global cause-after-emotion context matters.
  4. [Limitations and §4.4] The paper's stated limitations partially undercut the robustness claims. The Limitations paragraph acknowledges that M3HG 'cannot handle excessively long conversations' and 'may suffer from error propagation in the multimodal fusion process when emotion labels have uneven information across modalities.' Table 4 indeed shows a drop for conversations with >10 utterances, and the case study in Figure 8 shows M3HG (T+A+V) underperforming M3HG (T) in a modal-conflict example. These are not fatal, but the claims of superiority in long-conversation and cause-after-emotion scenarios should be scoped accordingly, and the case study's negative example should be discussed more explicitly.
minor comments (6)
  1. [Equation (2)] The formula for H^m has a missing closing parenthesis: 'H m = LN (Em + E′m + F F N(E′m)' should end with '))'. Also, the notation 'E′m' is defined only by context; please define it explicitly.
  2. [Table 2] Typo: 'Pipline' should be 'Pipeline' in the first column.
  3. [§5.3] The percentages '21.28%' and '17.17%' are relative improvements, but the text does not say so. Please clarify that they are relative to HiLo's scores, not absolute percentage-point increases.
  4. [Appendix D] The unimodal feature extraction section says 'We selected Speaker-Aware RoBERTa (SA-RoBERTa) as the PLM,' but the implementation appendix states that ECF uses RoBERTa-large. Please clarify whether SA-RoBERTa or RoBERTa-large (or both) is used and why.
  5. [Algorithm 1] Typo in the algorithm title: 'Graph Constru tuon' should be 'Graph Construction.' Also, the algorithm only connects each utterance to K preceding utterances, which is consistent with the local window K=3, but the global connection edge is added only to the conversation super-node. The relationship between local edges and the global node is not shown in the pseudo-code; consider making it more explicit.
  6. [Table 4] The numbers for M3HG (T, A, V) on ECF for conversations with num_utt≤10 show 6 Avg 41.95 and 4 Avg 40.42. Since the 4 Avg excludes two low-frequency categories, it seems counterintuitive that 4 Avg is lower than 6 Avg. Please double-check the calculation or add a footnote explaining the weighting.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: M3HG's benchmark gains are empirical comparisons; no claimed prediction reduces to fitted inputs or self-citation chains.

full rationale

The paper's central claim is an empirical F1 superiority of M3HG over seven baselines on fixed ECF and MECAD test sets (Table 2), supported by ablations (Table 6) and subset analyses (Tables 4, 5, 7). Nothing in the derivation chain is defined in terms of the target triplet labels. The graph nodes (utterance, conversation, N^e, N^c) are initialized from unimodal features H', not from gold emotion/cause annotations; the edge types and meta-paths are hand-specified structural choices, and the emotion/cause classifier MLPs are trained with focal loss on the same supervised labels, which is standard supervised learning rather than a self-fulfilling definition. The hyperparameter K is selected on validation performance (Figure 7), so reporting the best K=3 is normal model selection, not a fitted input relabeled as a prediction. The only apparent in-house citation, 'Following (Chen et al., 2023)' for the PFFN layer in Section 4.5, is a minor architectural detail and not load-bearing: it does not supply the uniqueness of the architecture, the definition of the graph, or the benchmark outcome. Likewise, the claim that M3HG is 'the first MECTEC model capable of handling situations where cause utterances appear after emotion utterances' is an architectural/empirical claim validated in Appendix E.2, not a conclusion imported from a self-citation. Borrowing baseline numbers from prior papers and using larger PLMs in M3HG (Appendix D) is a legitimate comparability/correctness concern about whether the SOTA margin is mainly due to encoder capacity, but that is an experimental-control issue, not a circularity: the reported numbers are not constructed from the baselines' outputs. No circular step meeting the quote-and-reduction bar was found.

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

The central claim rests on standard supervised-learning assumptions: annotation quality, alignment of source videos, and the sufficiency of a fixed graph topology. The free parameters are the graph window K, focal loss settings, and unspecified fusion depth. The invented entities are model-internal graph nodes with no external evidence, so they carry a moderate burden but are not hypothesized physical entities.

free parameters (4)
  • Local context window K = 3
    Same-speaker and different-speaker edges only reach back K utterances; Figure 7 shows best performance at K=3, so the architecture's locality is tuned to the benchmarks.
  • Focal loss hyperparameters gamma and alpha_beta = not reported
    Equation 10 uses category balancing factors and a modulation parameter, but their values are not stated, leaving the training objective under-specified.
  • Learning rate, batch size, training epochs = 5e-6, 16, 50
    These are chosen without reported search and directly affect the final F1 numbers.
  • Number of HGAT fusion iterations/PFFN layers = not reported
    The depth of multi-scale semantic fusion is not specified, though it controls how far information propagates through the graph.
assumptions (5)
  • domain assumption M3ED conversation timestamps and source videos align correctly
    MECAD is built by collecting video segments using timestamps from M3ED; if alignment is imperfect, audio and video features are noisy.
  • domain assumption Majority-vote annotations are ground truth
    Three psychology students label each conversation with Fleiss kappa 0.6932, and the paper treats the resulting majority labels as correct for training and evaluation.
  • ad hoc to paper Five edge types and the K=3 local window are sufficient to model emotion-cause relationships
    The graph design in Equations 3-8 assumes same-speaker, different-speaker, global, emotion, and cause edges capture the causal context; no learned relational mechanism is included.
  • domain assumption Pretrained encoders provide usable unimodal features
    SA-RoBERTa, Wav2Vec2, and DenseNet are used without domain adaptation, and model performance depends on these representations.
  • domain assumption Ekman's six emotion categories are appropriate for this data
    The task restricts labels to six basic emotions, excluding nuanced or mixed emotions that appear in TV dialogues.
invented entities (4)
  • Emotional context nodes N^e
    purpose: Represent emotion-specific context per utterance and aggregate cross-modal cues before classification.
    These are internal graph node types with no falsifiable prediction outside the model; they are optimized end-to-end on benchmarks.
  • Causal context nodes N^c
    purpose: Represent cause-specific context per utterance and aggregate cross-modal cues before cause prediction.
    Same as emotional context nodes: model-internal constructs with no independent external evidence.
  • Conversation super-node SN^d
    purpose: Summarize global conversation state to propagate information across all utterances.
    An architectural device; its effectiveness is shown only through ablations on the same benchmarks used to train the model.
  • Utterance super-node SN^u containing modality nodes
    purpose: Encapsulate the three modality nodes of one utterance so that intra-utterance fusion can occur.
    A graph design choice with no independent falsifiable handle outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of M3HG: Multimodal, Multi-scale, and Multi-type Node Heterogeneous Graph for Emotion Cause Triplet Extraction in Conversations." pith.science (2026). https://pith.science/paper/BSPKBKD3

@misc{pith2026250818740,
  author       = {Pith},
  title        = {Pith review of: M3HG: Multimodal, Multi-scale, and Multi-type Node Heterogeneous Graph for Emotion Cause Triplet Extraction in Conversations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BSPKBKD3}},
  note         = {Machine review of arXiv:2508.18740}
}
read the original abstract

Emotion Cause Triplet Extraction in Multimodal Conversations (MECTEC) has recently gained significant attention in social media analysis, aiming to extract emotion utterances, cause utterances, and emotion categories simultaneously. However, the scarcity of related datasets, with only one published dataset featuring highly uniform dialogue scenarios, hinders model development in this field. To address this, we introduce MECAD, the first multimodal, multi-scenario MECTEC dataset, comprising 989 conversations from 56 TV series spanning a wide range of dialogue contexts. In addition, existing MECTEC methods fail to explicitly model emotional and causal contexts and neglect the fusion of semantic information at different levels, leading to performance degradation. In this paper, we propose M3HG, a novel model that explicitly captures emotional and causal contexts and effectively fuses contextual information at both inter- and intra-utterance levels via a multimodal heterogeneous graph. Extensive experiments demonstrate the effectiveness of M3HG compared with existing state-of-the-art methods. The codes and dataset are available at https://github.com/redifinition/M3HG.

Figures

Figures reproduced from arXiv: 2508.18740 by the authors.

Figure 1
Figure 1. An example of the MECTEC task. Each utterance contains three different modalities - text, audio, and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The framework of proposed M3HG. It consists of three main components: unimodal feature extraction, graph construction and multi-scale semantic fusion, and emotion-cause classification. Nodes. To explicitly model emotion and cause￾related contexts in conversations, we model them as emotional context nodes Ne and causal con￾text nodes Nc , respectively. To enable G to accu￾rately perceive the conversation information,… view at source ↗
Figure 5
Figure 5. Percentage of five cause types in the MECAD [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Schematic representation of Cohen’s Kappa [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 6
Figure 6. Figure 6: Super-Node-based edges and relations con [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Results of M3HG with various K values. Effect of the hyperparameter K. The hyperpa￾rameter K is closely related to the spatio-temporal complexity of the M3HG’s graph construction. We vary the size of K (ranging from 1 to 6) to test its effect, and the result of the M3H…
Figure 8
Figure 8. Figure 8: Comparison of utter-cause-emotion triplet on two test samples. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 28 canonical work pages

  1. [1]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449--12460

  2. [2]

    Tiantian Chen, Ying Shen, Xuri Chen, Lin Zhang, and Shengjie Zhao. 2023. Mpeg: A multi-perspective enhanced graph attention network for causal emotion entailment in conversations. IEEE Transactions on Affective Computing

  3. [3]

    Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37--46

  4. [4]

    Zixiang Ding, Rui Xia, and Jianfei Yu. 2020 a . Ecpe-2d: Emotion-cause pair extraction based on joint two-dimensional representation, interaction and prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3161--3170

  5. [5]

    Zixiang Ding, Rui Xia, and Jianfei Yu. 2020 b . End-to-end emotion-cause pair extraction based on sliding window multi-label learning. In Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), pages 3574--3583

  6. [6]

    Ren \'e Dirven. 1997. Emotions as cause and the cause of emotions. The language of emotions: Conceptualization, expression, and theoretical foundation, pages 55--83

  7. [7]

    Paul Ekman. 1992. An argument for basic emotions. Cognition & emotion, 6(3-4):169--200

  8. [8]

    Jia-Chen Gu, Tianda Li, Quan Liu, Zhen-Hua Ling, Zhiming Su, Si Wei, and Xiaodan Zhu. 2020. Speaker-aware bert for multi-turn response selection in retrieval-based chatbots. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, pages 2041--2044

Show all 36 references
  1. [9]

    Xiaojie Gu, Renze Lou, Lin Sun, and Shangxin Li. 2023. Page: A position-aware graph-based model for emotion cause entailment in conversation. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1--5. IEEE

  2. [10]

    Guimin Hu, Zhihong Zhu, Daniel Hershcovich, Hasti Seifi, and Jiayuan Xie. 2024. Unimeec: Towards unified multimodal emotion recognition and emotion cause. arXiv preprint arXiv:2404.00403

  3. [11]

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. 2017. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700--4708

  4. [12]

    JR Landis. 1977. The measurement of observer agreement for categorical data. Biometrics

  5. [13]

    Bobo Li, Hao Fei, Fei Li, Tat-seng Chua, and Donghong Ji. 2024 a . Multimodal emotion-cause pair extraction with holistic interaction and label constraint. ACM Transactions on Multimedia Computing, Communications and Applications

  6. [14]

    Jiang Li, Xiaoping Wang, Yingjian Liu, and Zhigang Zeng. 2024 b . Cfn-esa: A cross-modal fusion network with emotion-shift awareness for dialogue emotion recognition. IEEE Transactions on Affective Computing

  7. [15]

    Jiangnan Li, Fandong Meng, Zheng Lin, Rui Liu, Peng Fu, Yanan Cao, Weiping Wang, and Jie Zhou. 2022 a . Neutral utterances are also causes: Enhancing conversational causal emotion entailment with social commonsense knowledge. arXiv preprint arXiv:2205.00759

  8. [16]

    Wei Li, Yang Li, Vlad Pandelea, Mengshi Ge, Luyao Zhu, and Erik Cambria. 2022 b . Ecpec: Emotion-cause pair extraction in conversations. IEEE Transactions on Affective Computing, 14(3):1754--1765

  9. [17]

    I Loshchilov. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101

  10. [18]

    Mary L McHugh. 2012. Interrater reliability: the kappa statistic. Biochemia medica, 22(3):276--282

  11. [19]

    Soujanya Poria, Navonil Majumder, Devamanyu Hazarika, Deepanway Ghosal, Rishabh Bhardwaj, Samson Yu Bai Jian, Pengfei Hong, Romila Ghosh, Abhinaba Roy, Niyati Chhaya, et al. 2021. Recognizing emotion cause in conversations. Cognitive Computation, 13:1317--1332

  12. [20]

    T-YLPG Ross and GKHP Doll \'a r. 2017. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980--2988

  13. [21]

    Weizhou Shen, Siyue Wu, Yunyi Yang, and Xiaojun Quan. 2021. Directed acyclic graph network for conversational emotion recognition. arXiv preprint arXiv:2105.12907

  14. [22]

    Mohammad Soleymani, David Garcia, Brendan Jou, Bj \"o rn Schuller, Shih-Fu Chang, and Maja Pantic. 2017. A survey of multimodal sentiment analysis. Image and Vision Computing, 65:3--14

  15. [23]

    Andrew Steptoe and Lena Brydon. 2009. Emotional triggering of cardiac events. Neuroscience & Biobehavioral Reviews, 33(2):63--70

  16. [24]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems

  17. [25]

    Fanfan Wang, Zixiang Ding, Rui Xia, Zhaoyu Li, and Jianfei Yu. 2022. Multimodal emotion-cause pair extraction in conversations. IEEE Transactions on Affective Computing, 14(3):1832--1844

  18. [26]

    Fanfan Wang, Jianfei Yu, and Rui Xia. 2023. Generative emotion cause triplet extraction in conversations with commonsense knowledge. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 3952--3963

  19. [27]

    Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. 2019. Heterogeneous graph attention network. In The world wide web conference, pages 2022--2032

  20. [28]

    Penghui Wei, Jiahao Zhao, and Wenji Mao. 2020. Effective inter-clause modeling for end-to-end emotion-cause pair extraction. In Proceedings of the 58th annual meeting of the association for computational linguistics, pages 3171--3181

  21. [29]

    Bernard Weiner. 1985. An attributional theory of achievement motivation and emotion. Psychological review, 92(4):548

  22. [30]

    Duzhen Zhang, Zhen Yang, Fandong Meng, Xiuyi Chen, and Jie Zhou. 2022. Tsam: A two-stream attention model for causal emotion entailment. arXiv preprint arXiv:2203.00819

  23. [31]

    Jinming Zhao, Tenggan Zhang, Jingwen Hu, Yuchen Liu, Qin Jin, Xinchao Wang, and Haizhou Li. 2022. M3ed: Multi-modal multi-scene multi-label emotional dialogue database. arXiv preprint arXiv:2205.10237

  24. [32]

    Weixiang Zhao, Yanyan Zhao, Zhuojun Li, and Bing Qin. 2023. Knowledge-bridged causal interaction network for causal emotion entailment. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 14020--14028

  25. [33]

    Li Zheng, Donghong Ji, Fei Li, Hao Fei, Shengqiong Wu, Jingye Li, Bobo Li, and Chong Teng. 2023. Ecqed: emotion-cause quadruple extraction in dialogs. arXiv preprint arXiv:2306.03969

  26. [34]

    Xiaopeng Zheng, Zhiyue Liu, Zizhen Zhang, Zhaoyang Wang, and Jiahai Wang. 2022. Ueca-prompt: Universal prompt for emotion cause analysis. In Proceedings of the 29th International Conference on Computational Linguistics, pages 7031--7041

  27. [35]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  28. [36]

    write newline

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

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.