They Are Not the Same: Direct Causes Are Not Grounded Emotion Explanations

Abstract: Emotion-Cause Pair Extraction (ECPE) was introduced to explain why an emotion occurs, but this goal is now often reduced to binary pair/non-pair prediction. This proxy is useful for direct-cause extraction, yet easy to over-read as evidence grounded emotion explanation. We show that this interpretation is only partially valid. In IEMO-MECP, 90.9% of original positives remain emo-cause and 95.0% of original negatives remain non-pair, confirming that the binary ECPE task is largely preserved. The problem is that direct triggers alone do not constitute a grounded explanation. Emo-context, an utterance that helps interpret a target emotion without directly causing it, appears on both sides of the original boundary and is enriched near binary uncertainty, showing that the binary boundary has no stable place for such discourse evidence. Across evaluated ECPE models, direct triggers are recovered more reliably than contextual support. Under shortcut pressure, this imbalance becomes consequential. Binary-trained models assign higher pair scores to nearby lexically similar non-pair candidates than to evidence supported but structurally harder emo-cause and emo-context pairs. Thus, pair scores can reward convenient attributions over grounded explanations. High binary ECPE performance indicates that a model can identify direct triggers; it does not indicate that the model has explained the emotion. Code is publicly available at https://github.com/panzhzh/ECPExsame.

• The paper demonstrates that binary Emotion-Cause Pair Extraction preserves direct-cause decisions but cannot represent emo-context, an explanatory role that is non-causal yet relevant.
• The paper introduces the IEMO-MECP diagnostic audit, finding 5.0% of candidates are emo-context, with 90.9% of original positives retained as causes and strong boundary uncertainty.
• The paper shows models systematically misclassify emo-context as non-pair and favor local lexical shortcuts, motivating separate cause metrics, role-wise evaluation, and evidence-grounding stress tests.

Emotion-Cause Pair Extraction (ECPE) was introduced to explain why an emotion occurs (Liu et al., 2019), yet its evaluation has largely collapsed into binary pair/non-pair prediction. This paper argues that binary pair scores are a valid proxy for direct-cause extraction but cannot certify evidence-grounded emotion explanation, and it substantiates this claim through a diagnostic re-audit of the conversational ECPE candidate space and systematic stress tests of trained models. The central construct is emo-context: an utterance that helps interpret a target emotion without directly causing it. Because this role is non-causal yet explanatorily relevant, it has no stable position under a two-label boundary, and the paper shows empirically that models trained under binary supervision systematically fail to represent it.

Evaluation-validity framing

The authors formalize binary ECPE as a projection $\pi: \mathcal{R} \rightarrow \{0,1\}$ from a three-role evidential space $$\mathcal{R} = \{emo\text{-}cause, emo\text{-}context, non\text{-}pair\}$$ to the binary label space. Setting $\pi(emo\text{-}cause)=1$ and $\pi(non\text{-}pair)=0$ preserves direct-cause decisions, but emo-context admits no consistent projection: mapping it to 1 over-causalizes contextual support as a pair, while mapping it to 0 discards explanatory relevance as noise. This is a structural label-space bottleneck, not a modeling deficiency, and it motivates five falsifiable predictions (P1–P5) covering direct-cause preservation, boundary crossing, uncertainty enrichment, context-channel weakness, and shortcut–evidence conflict.

IEMO-MECP: a role-audited candidate space

IEMO-MECP extends ConvECPE [li2023ecpec] over IEMOCAP dialogues, retaining the inherited 86,075-pair lower-triangular candidate space and source-binary labels while assigning each pair one of three roles. Construction combines LLM-assisted proposals, targeted human refinement, and human review of high-risk boundary transitions; a blind re-annotation audit over 330 boundary-focused judgments yields 0.821 raw pairwise agreement with mean Cohen's $\kappa$ and Krippendorff's $\alpha$ both at 0.732. Notably, all 4,346 emo-context pairs carry explicit non-default provenance, ruling out default-fill artifacts. The audit is deliberately diagnostic rather than a replacement benchmark, and its scope is limited to IEMOCAP-style dyadic dialogues.

The audit largely preserves the original task: 90.9% of original positives remain emo-cause and 95.0% of original negatives remain non-pair (P1). The validity signal lies in the boundary-crossing pattern (P2): emo-context constitutes 8.8% of original positives and 4.6% of original negatives, appearing on both sides of the inherited pair/non-pair boundary. This two-sided movement rules out both a missing-positive account and a "weak emo-cause" interpretation. Boundary uncertainty diagnostics (P3) reinforce this: emo-context's overall rate of 5.0% rises to 9.8% among the top 5% most uncertain binary-scored pairs, versus 0.9% among the top 5% lowest-score pairs—concentrating precisely where the binary model is least confident.

A compatibility check addresses a natural objection: after remapping three-role predictions to the binary task, role-aware supervision matches or improves Pair F1 relative to direct binary training (e.g., RoBERTa improves from 37.60 to 40.87 when predicted emo-context is dropped). Gains come from separating direct-cause from non-triggering evidence during training, not from expanding the positive class—mapping predicted emo-context to the positive class generally hurts text-grounded settings.

Models preserve causes, not context

Across 18 model-modality settings (RoBERTa, WavLM, CLIP baselines plus HiLo, MECPE-2step, and M3HG-style systems) over three seeds, emo-context remains the limiting role in every family. Even the strongest setting (M3HG, T+A) achieves only 26.84 emo-context F1 against 56.73 emo-cause F1, and correctly predicts just 34.0% of gold emo-context pairs while assigning 55.0% to non-pair. The error direction is diagnostic: if emo-context were merely weak emo-cause, misses would drift toward emo-cause; instead they collapse into non-pair, a pattern consistent across families. Under binary supervision, gold emo-context scores closer to non-pair than to emo-cause in mean $p(\text{pair})$, confirming the absence of a context channel (P4). Three-class supervision exposes a learnable signal—candidate removal produces the largest emo-context-probability drop for gold emo-context pairs—but it remains fragile (only 21.3% context recall for three-class RWC-Fusion). The implication is that stronger contextual or multimodal modeling alone does not resolve a label-space validity problem.

Shortcut pressure: pair scores favor convenient attributions

The shortcut-versus-evidence stress test (P5) uses naturally occurring, unedited candidates: local, lexically overlapping non-pair controls versus long-distance emo-cause pairs and original-negative emo-context pairs. The ordering is stable across all binary-source models: local lexical non-pair controls receive the highest mean pair scores (e.g., 0.302 for RoBERTa) versus 0.139 for long-distance emo-cause and 0.105 for original-negative emo-context. Binary pair scores thus reward shortcut-compatible locality and lexical overlap over evidence-supported but structurally harder pairs.

Matched controls sharpen this result. When non-pair controls are matched to gold emo-context pairs on distance, source label, speaker relation, emotion, length, lexical overlap, and dialogue type, a positive but narrow separation remains: binary $p(\text{pair})$ gaps of 0.020–0.023 and three-class context-probability gaps of 0.040–0.050. Emo-context is therefore learnable but weakly represented once structural cues are controlled. Cross-validated probes using only observable structural and lexical features (no text or multimodal content) recover binary-source hard-pair decisions with AUC between 0.833 and 0.855 and score rankings with Spearman $\rho$ between 0.462 and 0.605. A substantial portion of binary-source model behavior is thus reachable from shortcut-compatible structure alone. Distance is treated as a risk factor rather than a role definition: 68.4% of emo-context pairs occur at distance $\geq 3$, but 1,375 occur at shorter distances.

Limitations and open questions

The paper is explicit that IEMO-MECP is a diagnostic audit of an inherited candidate space, not a universal ontology of emotion explanation; the emo-context criterion is intentionally strict to prevent it from becoming a catch-all label. The audit rests on IEMOCAP-style dyadic dialogues with author-team annotators rather than external crowdworkers, and the LLM-assisted construction stages were not blind to source-binary metadata—only the 330-judgment re-annotation audit was. Whether the boundary-compression pattern generalizes to multi-party, spontaneous, or culturally diverse interactions, and whether richer pragmatic and discourse features can stabilize the context channel, remain open. It is also unresolved whether background information interpretable as part of a causal chain can be reliably separated from non-triggering support at larger annotation scale.

Conclusion

This paper establishes that high binary ECPE performance certifies direct-trigger extraction, not grounded emotion explanation. Emo-context crosses both sides of the inherited binary boundary, concentrates near binary uncertainty, collapses into non-pair across model families, and loses to shortcut-compatible local candidates under stress. The recommended remedy is procedural rather than architectural: evaluations should report direct-cause metrics (Pair F1) separately from explanation diagnostics, including role-wise metrics, boundary uncertainty analysis, and shortcut-versus-evidence stress tests. Binary pair scores answer a narrower question than explanation-oriented systems often ask, and should not be exposed as faithful emotion explanations without role-level verification.