acceptodds
Under review as a conference paper at ICLR 2027

Learning What to Reject: Counterfactual Preference Alignment for Emotion–Cause Pair Extraction

Abstract

Emotion–cause pair extraction (ECPE) identifies emotion clauses together with the clauses that explain them. A persistent challenge is that ECPE benchmarks are strongly locality-skewed: annotated causes often occur near their emotions, making relative position a highly predictive shortcut for cause selection. Models can therefore favor nearby plausible clauses without learning the distinction that determines the annotation-supported cause. Existing methods mainly strengthen candidate representation or reasoning, but standard positive supervision provides little targeted contrast against shortcut-compatible alternatives. We introduce CPA-DPO, a counterfactual preference-alignment framework that turns such alternatives into explicit competitors. CPA-DPO constructs reliable, challenging, and controlled preferences: verifier-defined eligibility filters counterfactuals, hardest-valid selection chooses the most competitive eligible alternative, and position-balanced formation prevents position from deterministically revealing the preference label. Standard DPO then optimizes these task-specific comparisons. On the matched Qwen2.5-32B setting, CPA-DPO improves ECPE-CN pair F1 from 83.62 to 85.43 over Vanilla-DPO (). Its advantage grows to F1 on LONG pairs and on Rebalanced-CN, while independent hard-negative accuracy rises from 87.92 to 91.35.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Emotion-Cause Pair Extraction, Shortcut Learning, Positional Bias, Counterfactual Preference Alignment, Direct Preference Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_KE55OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on a meaningful problem in ECPE that the model tends to select ones that are close to the emotion clauses, which is also potentially beneficial to tasks with sequential input and sequential labeling. The proposed method aims to construct counterfactuals that are challenging to the predictor that that the model can better learn meaningful emotion-cause relationships rather than easily fall into clause proximity.

The authors use GPT-4o to generate three alternative cause clauses for each gold cause. A verifier is used to filter out low quality couterfactuals. Highest residual emotion-support score lead to the hardest-valid selection, which will be used for the subsequent DPO training.

Critical strengths
  1. The paper identifies a concrete weakness that models can leverage positional proximity as a shortcut, and may overlook genuinely meaningful emotion-support clauses.
  2. The paper reported comprehensive evaluations from diversive perspectives.
  3. The paper includes long-distance and position-rebalanced results, distractor relocation, independent hard negatives, ablations, threshold sensitivity, and cross-backbone results. The 7B, 14B, and Llama results are a useful check that the gains are not confined to one model size.
Critical weaknesses
  1. My major concern is, the proposed hardest-valid selection does not establish that the selected negative is genuinely difficult for the target model. The paper defines counterfactual eligibility using a frozen verifier, with threshold and margin as 0.8 and 0.3. In other words, the hardest negative is defined as the candidate that retains the greatest emotion-support score according to the verifier. My problem is that this verifier uncertainty or residual support is not necessarily equivalent to the difficulty for the trained cause extractor. Therefore, the proposed hardest-valid selection effectively chooses the candidate with the smallest verifier-measured support reduction among candidates that pass the threshold. The central claim is that the method learns from especially challenging counterfactuals. Without validating the hardness measure against the target task, the improvement could instead arise from another property of the selected examples.

  2. Another issue is that the human audit. The authors argue that their counterfactuals are challenging because they are plausible alternatives that compete with the annotated cause. However, according to Appendix D.4, the primary evaluation criteria is contextually plausibility rather than causally plausibility. The audit further shows that 94% of retained counterfactuals were not judged plausible as causes, even though 91.1% were contextually plausible. This raises concerns about whether these are genuinely challenging causal alternatives or merely fluent distractors. The authors might reasonably hypothesize that a negative can be difficult for a model even when humans recognize it as incorrect. That is true. The human audit alone does not prove the negatives are easy for models.

  3. The verifier performs two essential functions: determining whether a candidate is eligible and selecting the hardest candidate. Thus, the quality of the training supervision depends directly on its ability to assess emotion support. Appendix B.3 describes a verifier that calculates a binary probability from the logits corresponding to the Chinese tokens. The paper specifies its quantization and inference configuration but does not provide a clearly identifiable public checkpoint or a direct evaluation establishing its competence at the intended verification task. More importantly, the reported manual audit does not substitute for a systematic assessment of the verifier's discriminative capability.

  4. The experimental design does not fully disentangle the proposed mechanism from the benefits of additional synthetic supervision. Although the matched comparisons control the backbone, initialization, reference policy, and optimization configuration, CPA-DPO additionally benefits from GPT-4o-generated counterfactuals, verifier-based filtering, and augmented training documents. Consequently, the reported improvements may partly arise from additional semantic supervision or favorable training-data selection rather than the proposed preference-construction mechanism. A stronger comparison would include models trained with the same synthetic counterfactuals and comparable training exposure, without the proposed preference-selection procedure.

Questions for authors

See Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to KE55 (1/2): Target-Policy Hardness and Counterfactual Plausibility

Response to KE55 (1/2): Target-Policy Hardness and Counterfactual Plausibility

Dear Reviewer KE55,

Thank you for identifying the key distinction between verifier support and target-model difficulty. We have added direct target-policy diagnostics, a verifier competence test, and controls using the same synthetic resources to the revision. We address your first two concerns here and the remaining two in the next comment.

  1. Does residual verifier support identify difficult negatives for the target policy?

We agree that the two quantities need not coincide. Hardest-valid has the operational definition

j^*=\arg\max_{j\in V}A_{\mathrm{cf}}^{(j)}.$$ Ties follow the original candidate order. This is verifier-relative selection within the eligible pool. The revised text makes this scope explicit. We now evaluate difficulty using each fold's frozen ECE-SFT cause selector before DPO. On the actual augmented inputs, we score the complete structured chosen/rejected responses: $$m_j=L_{\mathrm{SFT}}(y^+\mid x_j^+)-L_{\mathrm{SFT}}(y_j^-\mid x_j^+).$$ $L$ sums completion-token log-probabilities, excluding the prompt and including EOS; lower $m_j$ means stronger competition from the rejected response. Random-valid and hardest-valid are matched by source, role orientation, within-block order, dummy, and input protocol, with candidate text allowed to differ. | Selection | Mean margin | Rejected response outscores chosen | |:--|--:|--:| | Uniform random eligible | 0.43 | 1,161/6,178 (18.79%) | | Highest eligible residual support | 0.31 | 1,372/6,178 (22.21%) | The direction also holds for the 5,994 singleton-gold records: margin 0.435 -> 0.315 and rejected-outscoring rate 18.70% -> 22.14%. Thus the result is not confined to multi-gold cases. These are pre-DPO completion-competition measurements; they are not free-generation error rates. Total log-probabilities can retain length effects. We additionally compare candidate rankings within the same source. The 4,328 sources with at least two eligible candidates contain 10,414 candidates and 7,844 candidate pairs; the remaining 1,850 sources offer no selection choice. Verifier residual-support rankings agree with target-policy difficulty rankings on 5,144/7,844 pairs (65.58%). The verifier selects the policy-hardest candidate in 2,501/4,328 sources (57.79%), compared with a 43.23% theoretical expectation for uniform selection, accounting for the mixture of two- and three-candidate pools. These results support residual support as an imperfect proxy that selects more competitive negatives on average. They do not establish that every selected candidate is the target policy's hardest negative. The downstream Random-valid comparison in the next comment separately tests whether this selection improves task performance. 2. Contextual plausibility versus causal plausibility. We agree that the original audit alone did not demonstrate model difficulty. Its 91.1% contextual plausibility and 6% causally plausible traps measure different properties: a coherent, event-related clause can remain an incorrect cause. Contextual naturalness supports the quality of the intervention; difficulty requires target-policy evidence such as the measurements above. We also audited final augmented preference inputs, separately from the original 180 replacement documents: 286/300 clearly valid, 14 ambiguous, and 0 clearly invalid. This evaluates sampled preference validity, not model difficulty. The human-audit and policy-diagnostic cohorts differ, so their aggregate rates do not establish how often a human-rejected trap also confuses the model. Independent hard-negative evaluation provides a further task-level check. It uses 2,000 held-out instances and 10,000 gold-negative comparisons. Negatives are original held-out clauses assessed for non-causality independently of the training teacher, verifier, and CPA candidate pool. The evaluated policy's yes/no log-odds rank the gold above the negative in 87.92% of comparisons for Vanilla-DPO and 91.35% for CPA-DPO. This supports improved cause discrimination beyond the generated training traps.
AuthorsOpenReview ↗

Response to KE55 (2/2): Verifier Competence and Matched Synthetic-Data Controls

Response to KE55 (2/2): Verifier Competence and Matched Synthetic-Data Controls

  1. Public verifier identity and direct competence test.

The verifier is Qwen/Qwen2.5-32B-Instruct, frozen with 0 ECPE-specific training samples, epochs, or updates. One verifier is shared across folds; candidate generation/scoring uses only the corresponding training fold. It is separate from the trainable cause selector. Support is the binary softmax over final-position logits for single-token "是"/"否", with left padding, maximum length 2,048, NF4 double quantization, and BF16 compute. These details are explicit in the revised appendix.

We directly test its intended task: whether the specified emotion has causal support in the current document. From the pre-gate constructible pool, we sample 400 source pairs (40/fold; 100 same-clause, 300 different-clause), uniformly selecting one structurally admissible replacement per source without consulting scores. Original/replacement documents yield 800 views.

Two annotators independently judge Support/Non-support/Ambiguous, blinded to scores and retention status; a third adjudicates. The source quotation identifies the target; causal facts absent from the current document cannot count as support. We use the original verifier scoring procedure and a pre-fixed A≥0.5A\ge0.5 diagnostic threshold, without tuning on these labels.

The replacement views contain 232 Support, 148 Non-support, and 20 Ambiguous cases. On the 380 clear replacements:

Human labelPredicted SupportPredicted Non-support
Support20725
Non-support18130

Accuracy is 88.68%, recall 89.22%, specificity 87.84%, and balanced accuracy 88.53%. Same-/different-clause replacement balanced accuracies are 87.02%/89.07%. Across all views, 770 are clear and 30 ambiguous; balanced accuracy is 90.62%. We emphasize replacement-only results because originals are predominantly positive.

We separately audit whether the actual 0.8/0.30.8/0.3 gate identifies replacements that substantially weaken support while preserving target/referent consistency:

Human paired-intervention judgmentGate passGate fail
Clearly satisfies12525
Clearly fails9195
Ambiguous541

Among all 139 gate passes, 125 satisfy this criterion, 9 fail, and 5 are ambiguous. On clear judgments, precision is 93.28% and recall 83.33%. This evaluates support-changing interventions; it does not identify exact gold-cause membership or the final preference error rate.

  1. Separating preference construction from additional synthetic supervision.

We agree that Vanilla versus CPA alone does not isolate the contribution of synthetic augmentation. We therefore add controls with the same synthetic inputs or candidate resources:

  • C1: CPA's exact augmented inputs and complete gold chosen outputs, trained with chosen-only cross-entropy.
  • C2: the same inputs/chosen outputs, with DPO against a mapped generic original-document non-gold/non-emotion negative. CPA/C2 both use singleton rejected outputs.
  • Random-valid: the same teacher pool, eligibility, sources, role/order variables, and dummy, selecting uniformly within the eligible pool.
  • No candidate Drop gate: the same CPA source cohort and source threshold, removing only the candidate Drop criterion and maximizing residual support over structurally admissible candidates.
  • C3: the same retained records and selected candidates, with gold-original/trap-appended placement fixed.

All use the same 6,178 source identities, 100% source-set overlap, fold-specific SFT starting policy, and 263 additional updates across ten folds. C1 has 6,178 chosen-only examples; DPO controls have 6,178 pairs. These match sources/examples/updates, without establishing equal token exposure or FLOPs.

MethodOverall F1LONG F1HN accuracy (%)
Vanilla-DPO83.6270.4287.92
C1: chosen-only augmented SFT84.1272.2588.98
C2: same-input generic-negative DPO84.4272.9689.84
Random-valid84.4873.4190.24
No candidate Drop gate84.0972.0289.33
C3: fixed placement85.0674.5090.87
CPA-DPO85.4375.4691.35

C2 exceeds Vanilla by +0.80 Overall/+2.54 LONG, consistent with a benefit from augmented inputs. CPA adds +1.31/+3.21 over C1, +1.01/+2.50 over C2, +0.95/+2.05 over uniform eligible selection, and +1.34/+3.44 over removing the candidate Drop gate. These comparisons support gains beyond synthetic-input exposure under the tested controls; they are not an additive decomposition of component effects. C3 changes position and distance jointly.

AuthorsOpenReview ↗

Looking Forward to Further Discussion

Dear Reviewer KE55,

Thank you again for your careful review and constructive suggestions. We sincerely appreciate your recognition of the motivation and experimental scope of our work, as well as the opportunity to strengthen the empirical validation of our proposed method. We have carefully considered your concerns and provided detailed responses with additional experimental evidence in our rebuttal.

Regarding your concern about hardest-valid selection and target-model difficulty, we have added direct target-policy hardness diagnostics using the frozen pre-DPO ECE-SFT model. The results show that hardest-valid selection produces more competitive rejected responses than uniform random selection within the eligible pool. We have also provided candidate-level ranking analyses to further examine the relationship between verifier-based selection and target-policy difficulty.

Regarding counterfactual plausibility and preference validity, we have clarified the distinction between contextual plausibility, causal plausibility, and model difficulty. We have additionally conducted a 300-instance human audit of the final augmented inputs and provided independent hard-negative evaluations to assess the learned cause discrimination beyond the generated training counterfactuals.

Regarding verifier reliability, we have specified the exact public verifier checkpoint and its inference configuration, and conducted a direct human-labeled competence evaluation. These results provide additional evidence supporting the verifier's ability to assess emotion support and the reliability of the eligibility screening procedure.

Regarding the contribution beyond additional synthetic supervision, we have added several controlled experiments, including chosen-only augmented SFT (C1), same-input generic-negative DPO (C2), random-valid selection, and fixed-placement DPO (C3). These comparisons use matched source records and training-update budgets, providing further evidence for the benefits of our targeted preference construction beyond synthetic augmentation alone.

We hope these clarifications and additional experimental results address your main concerns. We would greatly appreciate your feedback on whether the new evidence has adequately addressed these issues, and please let us know if any concerns remain. We would be happy to provide further explanations and continue the discussion.

Thank you again for your time and consideration.

Reviewer_1diZOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CPA-DPO to mitigate positional shortcuts in emotion–cause pair extraction. It constructs preference pairs through counterfactual generation, frozen-verifier filtering, hardest-valid selection, and position-balanced placement, then trains the cause selector using standard DPO. Reported improvements over Vanilla-DPO are larger on LONG and Rebalanced-CN evaluations, suggesting reduced reliance on positional shortcuts.

Critical strengths

The paper addresses a well-motivated problem: locality bias can encourage ECPE models to favor nearby plausible clauses over annotation-supported causes.

The preference-construction pipeline coherently addresses negative reliability, difficulty, and positional label leakage. Generation and verification are performed offline, preserving the original inference pipeline.

The reported gains are promising, including +1.81 Overall F1 and larger improvements on LONG and Rebalanced-CN. Component ablations, human audits, and consistent Overall F1 improvements across ten folds provide complementary evidence.

Critical weaknesses

Core implementation and evaluation details remain insufficiently specified. These include the verifier checkpoint and training provenance, the complete ECE-SFT configuration, Vanilla-DPO negative construction, and the construction and size of independent hard-negative and relocation evaluations. Because the contribution centers on preference construction, these omissions substantially limit reproducibility and assessment of the main comparisons.

Although the backbone and inference protocols are controlled, matching of source records and supervision amounts between Vanilla-DPO and CPA-DPO is not established. Consequently, the reported gains cannot be cleanly separated from source-selection and training-budget effects.

The leakage probes support the intended construction control of position balancing, but do not establish its downstream contribution. A controlled ablation that changes only the placement policy is missing.

Questions for authors

Could the authors specify the verifier identity and training data, the complete ECE-SFT setup, and the construction and scoring protocols for the main baseline and diagnostic evaluations?

How much improvement remains with matched source records and training budgets? What is the effect of removing position balancing while keeping the retained records and counterfactual candidates fixed?

How are same-clause emotion–cause cases handled when position balancing moves the gold clause but retains the original emotion index (Appendix B.6)? Please clarify any exclusion or special processing and the corresponding label-validity checks.

3 comments
AuthorsOpenReview ↗

Response to 1diZ (1/2): Implementation and Diagnostic Evaluation

Response to 1diZ (1/2): Implementation and Diagnostic Evaluation

Dear Reviewer 1diZ,

Thank you for the careful review. The revised manuscript expands the implementation and evaluation details. We address Q1 here; matched comparisons and same-clause cases follow in the next comment.

Q1. Verifier, training, baseline, and diagnostic protocols.

Verifier identity and provenance. We use Qwen/Qwen2.5-32B-Instruct, frozen with 0 ECPE-specific training samples, epochs, or updates. One verifier is shared across folds; generation/verification use only the corresponding training fold. It is separate from the fold-specific cause selector.

Support is the binary softmax of final-input-position logits for the single-token verbalizers "是"/"否":

A(D,e)=exp⁡(ℓ是)exp⁡(ℓ是)+exp⁡(ℓ否).A(D,e)=\frac{\exp(\ell_{\text{是}})}{\exp(\ell_{\text{是}})+\exp(\ell_{\text{否}})}.

Scoring is deterministic: left padding, maximum length 2,048, 4-bit NF4 double quantization, BF16. Eligibility requires Aorig≥0.8A_{\mathrm{orig}}\ge0.8 and Aorig−Acf≥0.3A_{\mathrm{orig}}-A_{\mathrm{cf}}\ge0.3; hardest-valid selects the eligible candidate with highest residual AcfA_{\mathrm{cf}}. Support change is an operational screening criterion.

ECE-SFT and DPO. Each fold's cause selector starts from Qwen2.5-32B-Instruct. ECE-SFT uses completion-only cross-entropy; LoRA rank 64, alpha 16, dropout 0.05 on q/k/v/o and gate/up/down projections; LR 10−410^{-4}; 2 epochs; batch/accumulation 2/8; maximum length 2,048; paged_adamw_32bit; cosine scheduling; warmup 0.03; weight decay 0; gradient clipping 1.0; NF4/BF16; gradient checkpointing; seed 42+f42+f. Prompt tokens are masked from the loss.

DPO initializes policy and frozen reference from the same fold's ECE-SFT checkpoint and updates only cause-selector adapters, keeping the emotion extractor fixed: β=0.1\beta=0.1, LR 3×10−63\times10^{-6}, 1 epoch, batch/accumulation 1/24, sequence/prompt limits 512/448. Final-epoch checkpoints are used without test-based selection. The revised training appendix consolidates these settings.

Vanilla-DPO and headline evaluation. Vanilla uses the original document, the complete annotated gold-cause set as chosen, and a uniformly sampled singleton rejected cause outside the complete gold set and emotion clause. SFT, Vanilla, and CPA share fold-wise emotion predictions and direct JSON cause-list generation with the same parser and Pair-F1 evaluator. Teacher generation and verifier scoring occur only offline.

Independent hard negatives. The benchmark has 2,000 held-out query instances (200/fold), from 1,941 distinct parent-emotion targets and 1,814 documents. Each instance contains one gold and five negatives, yielding 10,000 gold-negative comparisons. The 59 additional instances reuse targets with different negative sets. For multi-gold targets, one gold is fixed by stable index order.

Negatives are original held-out clauses, excluding all annotated gold causes, the target emotion clause, and gold-text duplicates. Two annotators independently assess non-causality; a third adjudicates. Potential causes and ambiguous candidates are excluded; five eligible negatives are sampled without replacement per instance. Queries have no same-fold training-parent overlap. Teacher/verifier outputs, CPA candidates, and evaluated-policy scores do not select negatives.

The evaluated policy scores

sθ(c)=log⁡Pθ(是∣x,c)−log⁡Pθ(否∣x,c).s_\theta(c)=\log P_\theta(\text{是}\mid x,c)-\log P_\theta(\text{否}\mid x,c).

Accuracy counts sθ(cg)>sθ(cn)s_\theta(c_g)>s_\theta(c_n); ties fail. Vanilla/CPA obtain 8,792/9,135 correct comparisons, or 87.92%/91.35%. This is pairwise accuracy; comparisons sharing a target are correlated.

Rebalanced-CN retains all held-out documents with LONG pairs (∣e−k∣≥2|e-k|\ge2) and samples purely local documents with probability 0.30 (Random(42+f), document-ID order). Full documents/labels are kept; same-fold training-text matches are removed without replenishment, leaving 817 documents/940 pairs.

Relocation. From 1,943 eligible held-out parents, we select 1,800 bases. One existing distractor, excluding every annotated emotion/gold clause, is moved immediately before/after the target emotion or to the document start/end. No text is rewritten or duplicated; other clauses retain relative order. Labels and fixed upstream predictions are remapped by the same permutation.

The four operations yield 7,200 views, including 523 retained no-ops and 6,677 changed views. A separate 200-view audit gives 194 valid/6 ambiguous/0 clearly invalid. Using the same clean/perturbed bases,

W=110∑f=110max⁡r∈R(Ff,clean−Ff,r),W=\frac1{10}\sum_{f=1}^{10}\max_{r\in\mathcal R}(F_{f,\mathrm{clean}}-F_{f,r}),

where FF is Pair-F1 and R\mathcal R contains the four operations. Worst Drop is computed within each fold before averaging: 4.27 for Vanilla and 2.72 for CPA, in F1 points.

AuthorsOpenReview ↗

Response to 1diZ (2/2): Matched Comparisons and Same-Clause Label Validity

Response to 1diZ (2/2): Matched Comparisons and Same-Clause Label Validity

Q2. Matched sources, supervision amounts, and placement ablation.

Vanilla-DPO, CPA-DPO, and C1/C2/C3 use exactly the same 6,178 source identities, defined before augmentation as (fold,parent document,emotion,targeted gold cause). Their source-set overlap is 100%, covering 5,995 fold-parent combinations. Per-fold counts are 478/619/639/602/640/629/612/627/673/659.

All start from the same fold-specific SFT policy and use one additional epoch, batch 1, accumulation 24, one process, and a final partial update. Per-fold additional updates are 20/26/27/26/27/27/26/27/29/28, totaling 263 across ten folds. C1 has 6,178 chosen-only examples; the DPO methods have 6,178 preference pairs. This matches sources, examples, and updates; it does not establish equal tokens or FLOPs.

MethodOverall F1LONG F1Rebalanced F1HN accuracy (%)
Vanilla-DPO83.6270.4274.6587.92
C1: chosen-only augmented SFT84.1272.2576.5588.98
C2: same-input generic-negative DPO84.4272.9677.3189.84
C3: fixed placement85.0674.5078.7290.87
CPA-DPO85.4375.4679.9591.35

Thus CPA's gains over matched Vanilla remain +1.81 Overall, +5.04 LONG, +5.30 Rebalanced, and +3.43 HN. C1 uses CPA's augmented inputs and complete gold chosen outputs with chosen-only cross-entropy. C2 uses the same inputs/chosen outputs but replaces the trap with a mapped generic original-document non-gold/non-emotion negative. CPA and C2 both use singleton rejected outputs. CPA exceeds C1 by +1.31 Overall/+3.21 LONG and C2 by +1.01/+2.50, providing additional controls for augmentation and negative construction.

C3 directly addresses the requested placement ablation: it fixes the retained source records, selected gold/trap, dummy, within-block order, initialization, and training budget, setting every orientation to gold-original/trap-appended. Restoring role balancing improves Overall/LONG/Rebalanced/HN by +0.37/+0.96/+1.23/+0.48. This is downstream evidence beyond the metadata probes. The comparison measures the placement policy's joint effect on role, position, and distance; distance itself is not held fixed.

Q3. Same-clause targets when the gold clause moves.

Same-clause records are retained: 3,241/6,178, including 1,620 in the swapped orientation. There is no blanket exclusion or separate objective for these cases. The revised Appendix B.6 distinguishes source target identity from the emotion-bearing clause's current position.

The augmented training input explicitly supplies the source emotion ID and quotation. emotion_idx returns this fixed source-target ID; cause_idxs refer to current augmented-document positions. We separately track the relocated emotion-bearing clause. The parameterized input template, in English translation, is:

Augmented document (current clause numbers): {augmented_document}
Source emotion target ID: {source_emotion_id}
Source emotion quotation: "{source_emotion_quote}"
The ID identifies the source target; text at that position may differ after augmentation.
For this target, select all causes from the current document.
Return JSON only: retain the source ID in emotion_idx;
use current document positions in cause_idxs.

For illustration, if source esrc=ksrc=2e_{\mathrm{src}}=k_{\mathrm{src}}=2 and its gold/emotion-bearing clause moves to position 5, chosen is {"emotion_idx":2,"cause_idxs":[5]} and rejected is {"emotion_idx":2,"cause_idxs":[2]}. Both concern the same specified target; clause 2 now contains the trap. Other gold causes are retained and remapped in the complete chosen set. On unaugmented test documents, source IDs and current emotion-clause positions coincide.

We audited the actual final augmented inputs, rather than relying on the original 180-item replacement-stage audit:

Final-input groupNValidAmbiguousClearly invalid
Same-clause, gold original1009640
Same-clause, gold relocated1009280
Different-clause, both orientations1009820
Total300286140

The earlier 47-case swapped same-clause check is included in the 100 relocated cases. Target/referent consistency, output indices, and visibility after actual tokenization/truncation pass in 47/47; semantic preferences are 43 valid/4 ambiguous/0 clearly invalid. All 6,178 records pass remapping and necessary-span visibility checks. Structural checks and sampled semantic validity are distinct; neither establishes zero population error. Target binding also does not rule out shortcuts from quotation-gold text overlap.

AuthorsOpenReview ↗

Looking Forward to Further Discussion

Dear Reviewer 1diZ,

Thank you again for your careful review and constructive suggestions. We sincerely appreciate your recognition of the motivation and potential value of our work, as well as the opportunity to improve its clarity, reproducibility, and experimental validation. We have carefully considered your concerns and provided detailed responses in our rebuttal.

For Q1 regarding implementation and evaluation reproducibility, we have clarified the verifier identity and configuration, the complete ECE-SFT and DPO training settings, the Vanilla-DPO negative construction procedure, and the detailed protocols for Independent HN, Rebalanced-CN, and distractor-relocation evaluations.

For Q2 regarding controlled comparisons and the contribution of position balancing, we have verified identical source coverage (6,178 source records) and matched training-update budgets across methods. We have also added three controlled experiments: chosen-only augmented SFT (C1), same-input generic-negative DPO (C2), and fixed-placement DPO (C3). These results provide further evidence that the improvements cannot be explained solely by positional augmentation or generic negative supervision. In particular, the comparison with C3 provides direct downstream evidence for the benefit of position balancing.

For Q3 regarding same-clause emotion–cause cases, we have clarified the distinction between source emotion identity and augmented-document clause positions, including the handling of relocated gold clauses. We have also conducted a 300-instance human audit of the final augmented inputs and additional structural validity checks to evaluate label consistency.

We hope these clarifications and additional experimental results address your main concerns. We would greatly appreciate your feedback on whether the new evidence has adequately addressed these issues, and please let us know if any concerns remain. We would be happy to provide further explanations and continue the discussion.

Thank you again for your time and consideration.

Reviewer_XbAvOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes CPA-DPO, a method for constructing preference pairs for Direct Preference Optimisation (DPO) in Emotion–Cause Pair Extraction (ECPE), with the aim of reducing reliance on locality shortcuts (i.e., the tendency of causes to appear close to their emotion clause).

For each training example, GPT-4o generates several “decoy” rewrites of the true cause clause; a frozen verifier filters them, and the eligible decoy with the highest residual verifier score is selected. The training input is then augmented by randomly placing the true cause and the decoy at either the original cause position or an appended position (together with a filler clause), and standard DPO is used to prefer the true cause over the decoy.

The main comparison uses Qwen2.5-32B against four baselines (4-shot prompting, DECC-style 4-shot prompting, SFT, and Vanilla-DPO). Results for earlier ECPE methods are taken from the literature, where they were obtained with older models.

Critical strengths
  • The motivation is relevant: reducing ECPE models’ reliance on positional shortcuts is a well-known and practically important problem.
  • The idea of explicitly constructing counterfactual preference pairs, so that position alone cannot determine the preferred output, is interesting.
Critical weaknesses
  • Presentation. The paper is difficult to read in several places, and parts of it appear to have been heavily machine-generated without careful revision. Among others, for example:

    • \citet is used where \citep is required (e.g., "…which clauses explain them Xia & Ding (2019)").
    • Cross-references are broken ("Eq. equation 12").
    • "Gold-E ECE" is only glossed in the main text as "isolating cause selection"; that it relies on gold emotion annotations must be inferred from Appendix E.1.
    • The method is inconsistently named "CPA-DPO", "Balanced CPA-DPO", and "Balanced CPA" across the paper and appendix.
    • The FF symbols in Equations 14–15 are never defined.
    • Several appendix passages read as automatically generated descriptions of the code rather than of the method. For instance: "The implementation identifies the prompt as counterfactual_generation_v2. The API payload contains a single user message and does not add a separate system message. The implementation does not explicitly set top_p, max_tokens, a generation seed, or a JSON response-format option." Similarly, Appendix C is titled "Confirmed Training Configuration", and it states that the SFT configuration "is intentionally not reconstructed here from incomplete provenance".
  • Missing related work. Highly relevant prior work is not discussed:

    • Bao et al., Multi-Granularity Semantic Aware Graph Model for Reducing Position Bias in Emotion-Cause Pair Extraction, Findings of ACL 2022.
    • Ding & Kejriwal, An Experimental Study of the Effects of Position Bias on Emotion Cause Extraction, arXiv 2020, and Yan et al., Position Bias Mitigation: A Knowledge-Aware Graph Model for Emotion Cause Extraction, ACL 2021.
  • Choice of backbone. The main results use only Qwen2.5-32B, a mid-sized model released in 2024. No recent or frontier model is evaluated, either fine-tuned or prompted.

  • Comparison with prior work. The results for previous methods (Table 1a) are copied from the literature and were not re-run with the same backbone. It therefore remains unclear to what extent the reported gains are due to the stronger backbone rather than the proposed method.

  • Undefined baselines. The main baseline, Vanilla-DPO, is never defined; in particular, it is not described how its rejected outputs are obtained. The other baselines are not described in sufficient detail either, and the prompt for the 4-shot baseline is not provided (the Reproducibility Statement refers to supplementary material containing the complete prompts, but none was submitted).

  • Human evaluation. The human evaluation is not described in sufficient detail to be interpreted (annotators, expertise, guidelines, blinding, sample sizes for the filtered-out set).

  • Ablation. It is unclear how the ablation in Table 4 was performed: the "Random negative" and "No eligibility" variants are not defined, and the number of training pairs per variant is not reported.

Questions for authors
  1. Could the authors discuss the relationship between their work and the related works listed above?
  2. Could the authors report results with a more recent LLM?
  3. Could the authors re-run some of the prior methods (or prompting-based baselines) on the same backbone, to disentangle the contribution of the backbone from that of the proposed method?
  4. Why are the true cause and the decoy only ever placed at the original cause position or appended at the end of the document? Were other positions considered?
4 comments
AuthorsOpenReview ↗

Response to XbAv (1/2): Related Work, Backbones, and Controlled Comparisons

Response to XbAv (1/2): Related Work, Backbones, and Controlled Comparisons

Dear Reviewer XbAv,

Thank you for the careful review. We address Q1-Q4 here; methodological details and presentation follow in the next comment.

Q1. Relationship to prior work.

Position bias and its mitigation precede our work. Ding and Kejriwal (2020) diagnosed its effects on ECE; Yan et al. (2021) introduced adversarial examples to diagnose position bias and a knowledge-aware graph model to mitigate it; Bao et al. (2022) developed a multi-granularity semantic graph for ECPE. We have added these references and clarified their relationship to CPA-DPO in the revised Related Work section.

CPA-DPO contributes preference construction for a pretrained generative extractor: verifier-filtered counterfactual candidates, hardest-valid selection, and role-balanced placement under standard DPO. Our claim concerns improved cause discrimination and reduced positional sensitivity in the tested settings.

Q2. More recent backbones.

During rebuttal, we evaluated five newer backbones. Cells report Vanilla-DPO -> CPA-DPO F1:

Backbone repositoryOverallLONG
google/gemma-4-31B-it84.72 -> 85.8872.86 -> 76.42
zai-org/GLM-4.7-Flash84.51 -> 85.6272.34 -> 75.71
mistralai/Mistral-Small-4-119B-260385.28 -> 86.3174.12 -> 77.35
zai-org/GLM-5.3-Flash-BF1686.07 -> 86.9475.81 -> 78.64
deepseek-ai/DeepSeek-V4.1-Flash-101786.32 -> 87.1376.19 -> 78.85

Each backbone uses its native tokenizer/chat template. Within each backbone, both methods share the fold-specific SFT initialization/reference, upstream emotion predictions, source cohort, teacher/verifier, evaluation, and checkpoint selection. Gains are +0.81 to +1.16 Overall and +2.66 to +3.56 LONG.

Q3. Separating backbone and method contributions.

Table 1a provides published-system context, without isolating CPA's contribution. The same-Qwen2.5-32B results are:

MethodOverallLONG
Direct 4-shot64.2054.10
DECC-style 4-shot adaptation72.1060.80
SFT (ICLR Table 1b)73.0746.39
Vanilla-DPO83.6270.42
C1: chosen-only augmented SFT84.1272.25
C2: generic-negative DPO84.4272.96
C3: fixed placement85.0674.50
CPA-DPO85.4375.46

Method attribution rests on CPA versus Vanilla: +1.81 Overall / +5.04 LONG, with the SFT initialization, frozen emotion extractor, upstream predictions, DPO reference, optimizer, and inference fixed. We do not attribute the full SFT-to-CPA gain to preference construction. C1/C2 share CPA's augmented inputs and complete gold chosen outputs; C3 fixes the candidates and gold-original placement. CPA exceeds C1/C2/C3 by +1.31/+1.01/+0.37 Overall, addressing augmentation, negative selection, and placement, respectively.

Direct 4-shot uses four demonstrations from the training fold, excluding held-out parents: fixed source order, independent random.Random(42+fold), greedy decoding, and max_new_tokens=128. We parse the last complete JSON, deduplicate valid indices, and count content/format failures as errors without answer repair; transport retries are separate.

English translation of the instruction, with four numbered training documents and gold-pair outputs preceding the query:

Extract all emotion-cause clause pairs from the numbered document.
Allow same-clause pairs and multiple causes for one emotion.
Use document evidence to identify causes.
Return JSON only: {"pairs":[[emotion_idx,cause_idx],...]}.
[Four training demonstrations]
Document: {numbered_test_document}
Answer:

Test gold emotions are not supplied. DECC-style passes outputs through Recognizing -> Locating -> Analysing -> Summarizing, using fold-training demonstrations and clause-index scoring. Its demonstration selection and scoring differ from original DECC. The template specifies the interface; the revised appendix includes the actual fold-specific demonstrations and complete prompts for each stage.

Q4. Other insertion positions.

Original/end placement preserves the order of other clauses and supplies a second slot for controlled gold/trap role swaps. We also tested interior insertion: the same semantic-clause/dummy block is inserted once into a uniformly sampled internal gap between adjacent original clauses, retaining source, selected gold/trap, role/order variables, and dummy; all indices are remapped. Interior placement gives 85.31 Overall / 75.62 LONG, versus 85.43 / 75.46 for original/end placement. These observations show no consistent advantage for either policy; they do not establish equivalence or end-placement optimality. Placement changes distance and surrounding context jointly.

AuthorsOpenReview ↗

Response to XbAv (2/2): Baselines, Human Evaluation, and Manuscript Revisions

Response to XbAv (2/2): Baseline Definitions, Human Evaluation, and Presentation

We clarify the baseline definitions, audits, and presentation below.

  1. Baselines, ablations, and training budgets.

Vanilla-DPO uses the original document, the complete annotated gold-cause set as chosen, and a uniformly sampled singleton rejected cause outside the complete gold set and emotion clause. CPA also uses a singleton rejected output; other gold causes remain in the document and chosen output.

Let HruleH_{\mathrm{rule}} be the structurally admissible teacher-candidate pool and define

V={j∈Hrule:Aorig≥0.8, Aorig−Acf(j)≥0.3}.V=\{j\in H_{\mathrm{rule}}:A_{\mathrm{orig}}\ge0.8,\ A_{\mathrm{orig}}-A_{\mathrm{cf}}^{(j)}\ge0.3\}.

CPA selects arg⁡max⁡j∈VAcf(j)\arg\max_{j\in V}A_{\mathrm{cf}}^{(j)}: highest residual verifier support among eligible candidates, with ties resolved by original candidate order.

  • No eligibility removes only the candidate-level Drop gate: it retains the CPA source cohort and source threshold, then maximizes AcfA_{\mathrm{cf}} over HruleH_{\mathrm{rule}}.
  • Random negative samples uniformly from HruleH_{\mathrm{rule}} without the candidate Drop gate. Random-valid instead samples uniformly from VV; these are distinct variants.
  • C1 applies chosen-only cross-entropy to CPA inputs/outputs. C2 replaces CPA's rejected candidate with a generic original-document non-gold/non-emotion clause mapped to its augmented position. C3 fixes gold-original/trap-appended placement.

All use the same 6,178 source identities (fold,parent,emotion,targeted gold cause); C1 uses 6,178 examples. Interior insertion retains this cohort. Per-fold counts are 478/619/639/602/640/629/612/627/673/659. With one epoch, batch 1, accumulation 24, one process, and the final partial accumulation retained, updates are 20/26/27/26/27/27/26/27/29/28: 263 across ten folds. Sources/examples/updates are matched; tokens/FLOPs are not established as equal.

No candidate Drop gate obtains 84.09 Overall / 72.02 LONG; Random negative 83.77/70.91; Random-valid 84.48/73.41; CPA 85.43/75.46.

ECE-SFT uses LoRA rank/alpha/dropout 64/16/0.05, completion-only loss, LR 10−410^{-4}, 2 epochs, batch/accumulation 2/8, maximum length 2,048, paged AdamW, cosine scheduling, warmup 0.03, and seed 42+f42+f. DPO uses the fold's SFT policy/frozen reference, β=0.1\beta=0.1, LR 3×10−63\times10^{-6}, one epoch, and sequence/prompt limits 512/448. The verifier is frozen Qwen2.5-32B-Instruct with no ECPE-specific training.

  1. Human evaluation: procedure and denominators.

Two annotators proficient in the document language and trained on ECPE guidelines independently judge each item; a third adjudicates disagreements. Method names, verifier scores, and gate status are hidden, and candidate A/B order is randomized. Contextual naturalness, causal plausibility, and preference direction are judged separately. The three preference labels are gold clearly preferred, ambiguous, and trap equally/better supported; unresolved implicit causality or multi-cause granularity remains ambiguous.

The original 180-item audit examined replacement documents (94 same-clause, 86 different-clause): 91.1% contextual plausibility, 93% gold preference, 6% causally plausible traps, and κ=0.78\kappa=0.78 for that audit.

Final augmented-input audit:

GroupNValidAmbiguousClearly invalid
Same-clause, gold original1009640
Same-clause, gold relocated1009280
Different-clause, both orientations1009820
Total300286140

The earlier 47-case final-input check is included in these 300. Source target identity is fixed by ID/quotation; cause indices refer to the augmented document. The sampled proportions do not establish zero population error.

A paired audit samples 180 of the 2,236 sources where CPA and NoGate select different replacements: 360 views. Gold is clearly preferred in 166/180 versus 135/180; traps remain causally plausible in 11/180 versus 43/180; naturalness is 169/180 versus 174/180. This concerns switched sources, not the entire filtered-out candidate pool. We withdraw the earlier filtered-out percentages because their sampling denominator is not documented.

For interior placement, a separate 47-case audit gives 42 valid/5 ambiguous/0 clearly invalid, with target/index/visibility checks passing in 47/47.

  1. Presentation corrections.

The revised manuscript explicitly defines Gold-E ECE as cause extraction conditioned on gold emotion annotations; end-to-end ECPE uses predicted emotions. In Eqs. 14-15, we define Fs,fmF^m_{s,f} as Pair-F1 (0-100) for method mm, slice ss, and fold f=1,…,10f=1,\ldots,10. We have standardized the name CPA-DPO, corrected citation formatting and malformed equation references, replaced implementation-log-style passages with methodological explanations, and consolidated the ECE-SFT/DPO settings in the training appendix.

After reading the rebuttal, I will maintain my score.

AuthorsOpenReview ↗

Follow-up on Remaining Concerns After Rebuttal

Dear Reviewer XbAv,

Thank you for taking the time to read our rebuttal and reconsider our submission.

We respect your decision to maintain your score. However, we would greatly appreciate some clarification regarding which specific concerns remain unresolved.

In our rebuttal, we have addressed your four questions through additional experiments and analyses, including evaluations on five newer backbones, controlled comparisons under the same backbone, and experiments with alternative insertion positions. We have also clarified the baseline definitions, human evaluation procedures, ablations, and presentation issues.

Given that these additions directly address the major concerns raised in your original review, we would appreciate understanding whether there are remaining technical or methodological issues that we have not adequately resolved, or whether your assessment is primarily based on other considerations. Any specific feedback would be very helpful for us to understand the basis for the maintained rating.

Thank you again for your time and consideration.

Reviewer_S3bsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CPA-DPO for emotion–cause pair extraction, targeting the tendency to select nearby clauses as causes. A teacher generates counterfactual alternatives, a frozen verifier filters candidates based on their reduction of document-level emotion support, and the hardest eligible alternative becomes the rejected answer. Gold and counterfactual clauses are balanced across original and appended positions before standard DPO training.

On Qwen2.5-32B, CPA-DPO improves ECPE-CN pair F1 from 83.62 to 85.43 over Vanilla-DPO, with larger gains on long-distance and position-rebalanced evaluations. The paper also reports hard-negative tests, human audits, component ablations and supplementary results across backbones and datasets. Inference retains the original two-stage extraction pipeline.

Critical strengths
  1. The reported improvements are consistent and concentrated in relevant settings. CPA-DPO gains 1.81 F1 overall, 5.04 on LONG pairs and 5.30 on Rebalanced-CN over Vanilla-DPO. The overall improvement occurs in all ten folds, and supplementary results show the same direction across four backbones.

  2. The preference-construction design addresses concrete problems: ambiguous negatives, insufficiently challenging comparisons and position-based label leakage. Eligibility and hardest-valid selection have separate ablations, and a human audit of 180 retained records finds the gold cause clearly preferable in 93% of comparisons. Holding the upstream emotion predictions and inference procedure fixed also strengthens the comparison between cause selectors.

Critical weaknesses
  1. Label preservation is unresolved when the emotion and cause occupy the same clause. In the swapped orientation, the construction replaces the original cause clause and appends the gold clause elsewhere, while Appendix B.6 explicitly keeps the emotion index unchanged. If the cause clause is also the emotion clause, this can relocate the target emotion without updating its index. Same-clause pairs constitute 23.58% of the benchmark. Unless these records are excluded or specially handled, the construction can produce incorrectly conditioned preferences. Valid-index checks alone do not resolve this semantic issue.

  2. Essential training provenance is missing. Appendix B.3 identifies the verifier only through a local checkpoint path and explicitly declines to specify its model identity. Its training data and fold-wise isolation are therefore unclear. The ECE-SFT configuration is also incomplete, while Vanilla-DPO’s preference-generation procedure, data size and source coverage are insufficiently specified. Matching the backbone and optimization settings does not resolve these differences. Since the contribution is preference-data construction, the provenance of those data and their comparison baseline is particularly important.

  3. The strongest diagnostic evaluations are underdefined. Independent hard negatives are described as separately constructed, but their sampling or generation procedure, validation and sample counts are not provided. The exact construction of Rebalanced-CN and distractor relocation is also unclear. These evaluations carry much of the evidence for reduced positional shortcut use, so their semantic validity and independence from training artifacts need to be assessable. The English and RECCON experiments likewise need a clearer account of training versus transfer.

  4. The contribution of position balancing is not isolated from positional augmentation. Moving the gold cause to an appended position directly changes its training distance from the emotion. Consequently, the source-distance statistics in Tables 16–17 do not establish the distance distribution of the actual augmented training inputs. Near-chance metadata probes demonstrate a construction control, but do not show which intervention produces the downstream gains. A matched balancing ablation and a chosen-output-only training control would help distinguish the proposed preference mechanism from ordinary augmentation.

Questions for authors
  1. How are records where the emotion and cause share a clause handled in the swapped orientation? How many such records survive filtering, and does the human audit verify the target emotion and preferred cause in the final augmented document?

  2. What exact verifier checkpoint and training data were used, and was it isolated from each fold’s held-out documents? Please also specify the ECE-SFT configuration and how Vanilla-DPO preferences were generated, including counts and source coverage.

3.Please provide the construction procedures, sample counts and validation criteria for Independent HN, Rebalanced-CN and distractor relocation. For ECPE-EN and RECCON, which datasets enter training, validation and testing?

  1. Validation of the proposed mechanism: What is the gold cause’s distance distribution after augmentation? Do the gains persist against chosen-output-only training on the same augmented inputs, and how much changes when position balancing is removed while candidates and training exposure are held fixed?
5 comments
AuthorsOpenReview ↗

Response to S3bs (1/2): Label Validity and Training Provenance

Response to S3bs (1/2): Label Validity and Training Provenance

Dear Reviewer S3bs,

Thank you for the careful review. We address label preservation and training provenance here, followed by evaluation protocols and matched controls in the next comment.

Q1. Same-clause targets and final-label validity.

Same-clause records are retained: 3,241 of the 6,178 training records, including 1,620 in the swapped orientation. The source target stays fixed while its emotion-bearing clause can move.

Replacement screening identifies the target through its source quotation and emotion label. Final cause-selection training supplies the source ID and quotation. The parameterized input template is shown below in English translation:

Augmented document (current clause numbers):
{augmented_document}
Source emotion target ID: {source_emotion_id}
Source emotion quotation: "{source_emotion_quote}"
The ID identifies the source target; text at that position may differ after augmentation.
For this same target, select all causes from the current document.
Return JSON only: emotion_idx must retain the source ID; cause_idxs must use current document positions.

Thus emotion_idx identifies the source target, while cause_idxs locate causes in the augmented document. We separately track the emotion-bearing clause's current position. Appendix B.6 omitted this distinction. On unaugmented test documents, source IDs and current emotion-clause positions coincide.

For illustration, if source esrc=ksrc=2e_{\mathrm{src}}=k_{\mathrm{src}}=2 and its gold/emotion-bearing clause moves to position 5, chosen is {"emotion_idx":2,"cause_idxs":[5]} and rejected is {"emotion_idx":2,"cause_idxs":[2]}. Both answer for the same target; position 2 contains the trap. Chosen retains the complete remapped gold set.

The original 180-record audit examined replacement documents, including 94 same-clause records. We additionally audited final augmented training inputs:

Final-input groupNValidAmbiguousClearly invalid
Same-clause, gold original1009640
Same-clause, gold relocated1009280
Different-clause, both orientations1009820
Total300286140

The original 47-case final-input re-audit is included in the 100 swapped same-clause cases. Target/referent consistency, gold/trap output indices, and visibility after tokenization/truncation passed in 47/47; preference validity was 43/4/0. All 6,178 records pass remapping and necessary-span visibility checks. These checks and sampled semantic judgments are distinct; the 300-case results do not estimate target-index errors or prove zero population errors. The source quote may contain causal facts, so target binding does not rule out policy shortcuts from quote–gold text overlap.

Q2. Verifier, ECE-SFT, and Vanilla-DPO provenance.

The verifier is Qwen/Qwen2.5-32B-Instruct, frozen without ECPE-specific training (0 task-training samples/epochs/updates). One verifier is shared across folds; candidate generation/scoring use only the corresponding fold's training documents. It is separate from the fold-specific cause selector. Scoring uses binary softmax over final-position logits for single-token “是”/“否”, with left padding, maximum length 2,048, NF4 double quantization, and BF16.

ECE-SFT initializes Qwen2.5-32B-Instruct separately per fold and uses LoRA rank 64, alpha 16, dropout 0.05 on attention/MLP projections; completion-only cross-entropy; 2 epochs; LR 10−410^{-4}; batch/accumulation 2/8; maximum length 2,048; paged_adamw_32bit; cosine scheduling; warmup 0.03; seed 42+f42+f.

DPO initializes policy/frozen reference from that fold's ECE-SFT and updates the same cause-selector adapters, keeping the emotion extractor fixed: β=0.1\beta=0.1, LR 3×10−63\times10^{-6}, 1 epoch, batch/accumulation 1/24, maximum length 512/prompt limit 448. Final-epoch checkpoints are used; teacher/verifier operations are offline.

Vanilla-DPO uses the original document, all annotated gold causes as chosen, and a uniformly sampled singleton rejected cause outside the complete gold set and emotion clause. CPA and C2 also use singleton rejected outputs: other gold causes remain in the input and chosen output, not in rejected. CPA and C2 therefore match in output-set sizes, although multi-gold chosen outputs can still be longer than rejected outputs. Vanilla, CPA, and the matched controls use the same 6,178 (fold, parent document, emotion, targeted gold cause) identities (5,995 fold–parent combinations).

Per-fold records are 478/619/639/602/640/629/612/627/673/659; additional updates are 20/26/27/26/27/27/26/27/29/28, totaling 263 across ten folds. This establishes source/example/update matching, not equal tokens/FLOPs.

AuthorsOpenReview ↗

Response to S3bs (2/2): Evaluation Protocols and Matched Mechanism Controls

Response to S3bs (2/2): Evaluation Protocols and Matched Mechanism Controls

Q3. Diagnostic evaluations and training versus transfer.

Independent HN uses 2,000 held-out instances (200/fold), 1,941 targets/1,814 documents, and 10,000 gold–negative comparisons. Each instance has one gold and five negatives; multiple-gold targets use a fixed-index gold. The 59 repeated-target instances use different negative sets. Negatives are original held-out clauses, excluding all gold/emotion clauses and gold-text duplicates. Two annotators independently assess non-causality; a third adjudicates. Potential causes/ambiguous candidates are excluded, and five eligible negatives are sampled without replacement per instance. There is no within-fold training-parent overlap; teacher/verifier outputs, the CPA pool, and policy scores do not select negatives.

The evaluated policy scores

sθ(c)=log⁡Pθ(是∣x,c)−log⁡Pθ(否∣x,c).s_\theta(c)=\log P_\theta(\text{是}\mid x,c)-\log P_\theta(\text{否}\mid x,c).

Accuracy is the fraction with sθ(cg)>sθ(cn)s_\theta(c_g)>s_\theta(c_n): 87.92%→91.35% for Vanilla→CPA. Comparisons sharing targets/documents are not independent samples.

Rebalanced-CN retains all held-out documents containing a LONG pair and samples purely local documents with probability 0.30 using Random(42+f) in document-ID order. Full documents/all labels are retained; normalized-text matches to that fold's training documents are removed without replenishment. The view has 817 documents/940 gold pairs, 314 LONG (33.40%), with zero within-fold training parent/text overlap. Vanilla/CPA F1 is 74.65/79.95.

Relocation moves one existing distractor, excluding every annotated gold/emotion clause, before/after the target emotion or to the document start/end. Other clauses retain relative order; all labels are remapped, without rewriting/duplication. The 1,800 shared bases yield 7,200 views, including 523 no-ops. A 200-view audit gives 194 valid / 6 ambiguous / 0 clearly invalid. On the same clean/perturbed set per fold,

W=110∑f=110max⁡r∈R(Ff,clean−Ff,r).W=\frac1{10}\sum_{f=1}^{10}\max_{r\in\mathcal R}(F_{f,\mathrm{clean}}-F_{f,r}).

Worst Drop is taken within each fold before averaging: Vanilla 4.27, CPA 2.72 F1 points.

ECPE-EN uses each fold's English training partition for ECE-SFT and preference construction, starting from the official base model. Train/validation/test sizes are 2,275/284/284 per fold; validation is unused for selection/tuning; checkpoints follow a fixed final-epoch rule. Pair-F1 is 67.15→68.12. Of 2,840 test records, 412 exactly match same-fold training text; reevaluating the existing checkpoints on the shared 2,428-record clean subset gives 66.25→67.18 (+0.93). This is an additional subset check, not clean retraining or CN→EN zero-shot transfer.

RECCON uses DD train (834 dialogues) for SFT/preferences, DD validation (47) for common checkpoint/hyperparameter selection, and DD test (225) for in-domain evaluation. The same DD-trained checkpoint is tested on IE (16 test dialogues), with no IE training, adaptation, or selection: DD→IE zero-shot transfer. Vanilla→CPA Macro-F1 is 91.35→92.90 on DD and 84.95→86.60 on IE; this metric differs from ECPE Pair-F1.

Q4. Final gold distances and matched augmentation controls.

We measure distance from the tracked emotion-bearing clause, d=∣pe+−pc+∣d=|p_e^+-p_c^+|, rather than the source-target ID:

DistanceTargeted goldAll chosen gold occurrences
03,2413,273
11,4461,574
2356376
≥3\ge31,1351,157
Total6,1786,380

Targeted-gold LONG increases from 5.10% in source records to 24.13% after augmentation; all-chosen LONG is 24.03%. Augmentation therefore increases long-distance exposure.

C1 uses CPA's starting policy and augmented inputs/chosen outputs for chosen-only cross-entropy. C2 uses the same inputs/chosen outputs, replacing the trap singleton with a remapped Vanilla generic-negative singleton. C3 fixes retained records, candidates, initialization, and updates, using gold-original placement with dummy/order controls. All share the source/update matching in Q2.

ConditionOverall F1LONG F1Rebalanced F1HN accuracy (%)
Vanilla-DPO83.6270.4274.6587.92
C1: chosen-only augmented SFT84.1272.2576.5588.98
C2: same-input generic-negative DPO84.4272.9677.3189.84
C3: fixed placement85.0674.5078.7290.87
CPA-DPO85.4375.4679.9591.35

CPA exceeds chosen-only C1 by 1.31 Overall/3.21 LONG F1 while preserving the same augmentation. Removing balancing with candidates fixed lowers Overall/LONG by 0.37/0.96. C3 measures the placement policy's joint role/position/distance effect, not an effect with distance held fixed. These controls provide downstream evidence beyond metadata probes.

Thanks for the reply. Some of my concerns were addressed but there are still some unresolved ones like:

  1. The training preferences still carry locality and string-match cues: table 17 shows that role balancing does not balance distance to the emotion-bearing clause. In all 3,241 same-clause records (52% of the training data), gold sits at tracked distance 0 and the trap at distance ≥1. Overall, 70.9% of traps are LONG versus 24.1% of gold.
  2. Same-clause records are over-represented, and the verifier signal may be confounded. Same-clause pairs are 23.6% of the benchmark and about 25% of the reconstructed source pool (4,455/18,056), but 52% of the final training set. For these records, replacing the cause also replaces the emotion clause itself. The support drop Δ may then partly reflect removing the emotion expression rather than the cause.
  3. Gold-E and end-to-end results disagree. With identical upstream emotion predictions. The Gold-E gap is only 0.79 for SFT to Vanilla and 0.93 for Vanilla to CPA. If the selectors differ only in cause selection, why does SFT lose 15 points from predicted emotions while DPO loses about 5?
Reviewer_PeU3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses locality shortcuts in emotion–cause pair extraction (ECPE), where models may favor clauses near an emotion rather than distinguish annotation-supported causes from plausible alternatives. CPA-DPO keeps the upstream emotion extractor fixed and changes supervision for the emotion-conditioned cause selector. A teacher first generates counterfactual replacements. A frozen document-level emotion-support verifier then filters these candidates according to the decrease in target-emotion support caused by replacement. Among the eligible candidates, the method selects the one with the highest residual support. Finally, the gold cause and counterfactual trap are balanced across original and appended positions. Standard DPO optimizes the resulting shared-input preferences. Teacher generation and verifier scoring are used only offline. The contribution is therefore task-specific preference construction, not a new optimization objective. Evaluation includes matched model comparisons, long-distance and position-rebalanced tests, distractor relocation, hard-negative discrimination, component ablations, and human auditing. Relative to Vanilla-DPO on Qwen2.5-32B, the paper reports gains of 1.81 ECPE-CN pair-F1 points, 5.04 LONG-F1 points, 5.30 Rebalanced-CN-F1 points, and 3.43 percentage points in Independent HN Accuracy.

Critical strengths
  1. A coherent, task-motivated supervision design.

  2. Consistent matched improvements and complementary behavioral evidence.

Critical weaknesses
  1. The central verifier is insufficiently reproducible, and its task alignment is only partially validated.

Appendix B.3 gives a repository-local path but no stable verifier model/checkpoint identity or adequate account of task-specific training and data scope. Because the verifier determines both eligibility and candidate selection, these omissions prevent faithful reconstruction of the preference data. The incomplete ECE-SFT configuration adds a related reproducibility gap (Appendix C).

Document-level emotion support is not equivalent to clause-level cause membership. With multiple independent causes, removing one annotated cause may leave document-level emotion support high. Conversely, replacing a gold cause with an alternative that remains a valid cause may reduce the support score without yielding a suitable rejected example. Verification uses a replacement document, whereas training includes both gold and trap. The human audit is useful, but its coverage of final augmented inputs, multiple-cause cases, and coincident emotion/cause indices is unclear (Appendix D.4).

Appendix B.5 correctly restricts “hardest-valid” to verifier-relative difficulty. However, the selection ablation establishes utility, not correspondence with the target policy’s confusion. These gaps weaken both reproducibility and the interpretation of the claimed reliable, challenging supervision.

  1. Rejection supervision is not adequately separated from positional augmentation and data-selection effects.

Position balancing removes the deterministic association between appended position and preference label, but it also relocates gold causes and changes their distances from the target emotion (Section 3.2 and Appendix B.6). LONG gains may therefore partly reflect positional augmentation rather than learning to reject misleading causes. Table 17’s source/retained-record locality statistics do not establish the distance distribution of the final augmented inputs. Near-chance metadata probes verify a construction control, not its downstream performance contribution.

The paper lacks chosen-only SFT on the same augmented inputs and correct outputs, and otherwise matched DPO with versus without position balancing. The pairwise-ranking control still uses positive–negative comparisons and does not replace the former (Appendix D.5). Appendix A also does not establish source-ID equality between Vanilla-DPO and CPA-DPO. Negative-construction rules, per-fold counts, source coverage, and update budgets require clearer reporting. Filtering may legitimately change the retained population, but those changes need to be distinguished from the value of the specific comparison design.

The existing ablations provide partial support. I do not require the teacher to contribute nothing or every component to be indispensable. The unresolved question is the incremental value over simpler alternatives using the same teacher resources and positional augmentation. This directly affects the demonstrated value of the central supervision-construction contribution.

Questions for authors

1.Which exact model and checkpoint were used for the frozen document-level verifier? Was the verifier trained or fine-tuned specifically for this task, and if so, what were its training objectives, data sources, and fold-isolation procedures? Can the authors provide a stable checkpoint identifier or sufficient configuration details to reproduce it, together with the key initialization and training settings for the ECE-SFT cause selector?

2.Did the audit of 180 retained preference records examine the replacement documents used for verifier screening or the final position-balanced augmented inputs used for DPO? How many of the audited and retained records involve multiple annotated causes or coincident emotion and cause indices? When the targeted cause and emotion occur in the same clause, how is the target emotion preserved or represented during replacement? Is there direct evidence that the verifier-selected candidates are more confusable for the target SFT policy under the same augmented input?

3.How were negative examples constructed for Vanilla-DPO, No eligibility, and Random-valid, and what were their per-fold preference counts, source coverage, and training-update budgets? What are the emotion–cause distance distributions in the final augmented inputs after gold/trap relocation? Can the authors report chosen-only SFT results using the identical augmented inputs and preferred outputs, as well as otherwise matched DPO results with and without position balancing? These comparisons could help distinguish the contribution of targeted preference supervision from positional augmentation and data-selection effects without requiring additional tasks or datasets.

4.What are the sample size, negative sources, annotation procedure, and scoring rule for the Independent HN evaluation? Does independence refer only to the evaluation examples, or does it also extend to the negative-construction procedures? How is train/test isolation maintained for Rebalanced-CN within each fold? What distractor-relocation operations are permitted, and how is Worst Drop aggregated across perturbations and folds?

3 comments
AuthorsOpenReview ↗

Response to PeU3: Clarifications and New Experimental Evidence

Dear Reviewer PeU3,

Thank you for the careful review. We address Q1–Q2 here and Q3–Q4 in the next comment.

Q1. Verifier and ECE-SFT configuration.

The verifier is Qwen/Qwen2.5-32B-Instruct, frozen with no ECPE-specific training (0 task-training samples/epochs/updates). It is shared across folds and separate from the fold-specific ECE-SFT cause selector. Candidate generation/scoring use only each fold's training documents. Scores use a binary softmax over final-position logits for single-token “是”/“否”, with left padding, maximum length 2,048, NF4 double quantization, and BF16.

ECE-SFT uses per-fold Qwen2.5-32B-Instruct initialization; LoRA rank 64, alpha 16, dropout 0.05 on attention/MLP projections; completion-only cross-entropy; 2 epochs; LR 10−410^{-4}; batch/accumulation 2/8; maximum length 2,048; paged_adamw_32bit; cosine scheduling; warmup 0.03; seed 42+f42+f.

DPO initializes policy/frozen reference from that fold's ECE-SFT, updating the same cause-selector adapters with the emotion extractor fixed: β=0.1\beta=0.1, LR 3×10−63\times10^{-6}, 1 epoch, batch/accumulation 1/24, maximum length 512/prompt limit 448. We use final-epoch checkpoints; teacher/verifier operations are offline.

Q2. Target preservation, audit coverage, and policy difficulty.

During replacement screening, the source emotion quotation and label identify the target. Final cause-selection training supplies its source ID and quotation:

ae=(esrc,qe),x+=T(D+,ae).a_e=(e_{\mathrm{src}},q_e),\qquad x^+=T(D^+,a_e).

The augmented training interface returns the source-target ID in emotion_idx, while cause_idxs use augmented-document positions. The relocated emotion-bearing clause position is tracked separately. Appendix B.6's statement that the emotion index remains unchanged omitted this distinction. In unaugmented test documents, source IDs and current emotion-clause positions coincide.

For illustration, if a same-clause source has esrc=ksrc=2e_{\mathrm{src}}=k_{\mathrm{src}}=2 and its gold/emotion-bearing clause moves to position 5, chosen is {"emotion_idx":2,"cause_idxs":[5]} and rejected is {"emotion_idx":2,"cause_idxs":[2]}. Both concern the same specified target; position 2 now contains the trap.

The original 180-record audit examined replacement documents: 94 same-clause and 7 multi-cause-derived records. The 6,178 retained records include 3,241 same-clause (1,620 swapped) and 242 multi-cause-derived records; 184 retain multiple distinct gold causes. Chosen outputs preserve the complete remapped gold set.

We additionally audited 300 final augmented training inputs:

Final-input groupNValidAmbiguousClearly invalid
Same-clause, gold original1009640
Same-clause, gold relocated1009280
Different-clause, both orientations1009820
Total300286140

The original 47-case final-input re-audit is included in the 100 swapped same-clause cases. Target/referent consistency, gold/trap output indices, and visibility after tokenization/truncation passed in 47/47; preference validity was 43/4/0. Two cases are multi-cause-derived, insufficient to establish a multi-cause-wide validity rate. All 6,178 records pass remapping and necessary-span visibility checks. The table measures sampled preference validity, not target-index errors or a population error rate.

The source quote may contain causal facts. Target binding and these audits do not rule out policy shortcuts from quote–gold text overlap; emotion-only-anchor diagnostics test verifier sensitivity, not policy reliance on that overlap.

Document-level support is a screening proxy: other causes can maintain support, and a support decrease does not guarantee a non-cause. Direct human-labeled testing gives 88.53% balanced accuracy on 380 unambiguous replacement views. Among 139 gate-passing cases, semantic-intervention assessment finds 125 satisfying / 9 failing / 5 ambiguous cases. On 200 separable sources, changing causes while preserving emotion expression yields mean Drop 0.43, versus 0.19 for emotion-expression deletion and 0.06 for placebo rewriting.

For frozen pre-DPO ECE-SFT, we score complete outputs by

m=log⁡P(yw∣x+)−log⁡P(yl∣x+).m=\log P(y_w\mid x^+)-\log P(y_l\mid x^+).

Hardest-valid versus random-valid lowers mean teacher-forced margin 0.43→0.31 and raises P(m<0)P(m<0) from 18.79% to 22.21% across 6,178 sources. Chosen/rejected share an input within each comparison; across selections, source/layout are matched while trap text can differ. This directly supports greater average completion-level confusion for the target policy.

AuthorsOpenReview ↗

Response to PeU3: Clarifications and New Experimental Evidence

Continuation (2/2): responses to Q3–Q4.

Q3. Negative construction, matched controls, and final distances.

Let HruleH_{\mathrm{rule}} contain structurally valid, nonduplicate candidates, with Aorig=S(D,e)A_{\mathrm{orig}}=S(D,e) and Acf(j)=S(Dcf(j),e)A_{\mathrm{cf}}^{(j)}=S(D_{\mathrm{cf}}^{(j)},e). Eligible candidates are

V={j∈Hrule:Aorig≥0.8, Aorig−Acf(j)≥0.3}.V=\{j\in H_{\mathrm{rule}}:A_{\mathrm{orig}}\ge0.8,\ A_{\mathrm{orig}}-A_{\mathrm{cf}}^{(j)}\ge0.3\}.

  • Vanilla-DPO: original document; uniformly sample a singleton rejected cause outside the complete gold set and emotion clause; chosen contains all gold causes.
  • CPA: maximize AcfA_{\mathrm{cf}} over VV; ties follow original candidate order.
  • Random-valid: uniformly sample once from VV with Random(42+f) and fix the resulting file.
  • No candidate Drop gate (“No eligibility”): retain the CPA source cohort and source threshold 0.8; remove only the candidate Drop threshold and maximize AcfA_{\mathrm{cf}} over HruleH_{\mathrm{rule}}.

All seven conditions share identical (fold, parent document, emotion, targeted gold cause) sets. Per-fold records are 478/619/639/602/640/629/612/627/673/659 (6,178 total); additional updates are 20/26/27/26/27/27/26/27/29/28 (263 summed across folds). This matches sources/updates, not tokens/FLOPs.

C1 uses CPA's starting policy, augmented inputs, chosen outputs, and updates with chosen-only cross-entropy. C2 keeps those inputs/outputs but remaps Vanilla's generic negative into the augmented document. C3 fixes candidates, uses gold-original placement, and retains dummy/order controls.

ConditionOverall F1LONG F1
Vanilla-DPO83.6270.42
No candidate Drop gate84.0972.02
Random-valid84.4873.41
C1: chosen-only augmented SFT84.1272.25
C2: same-input generic-negative DPO84.4272.96
C3: fixed placement85.0674.50
CPA-DPO85.4375.46

CPA exceeds C1/C2 by 1.31/1.01 Overall F1. C3 measures a joint placement/distance effect. No-gate→Random-valid changes both gating and selection, so it does not isolate gating.

Final distances track the emotion-bearing clause, d=∣pe+−pc+∣d=|p_e^+-p_c^+|:

DistanceTargeted goldSelected trap
03,2410
11,4461,796
23561,061
≥3\ge31,1353,321

LONG (d≥2d\ge2) is 24.13% for targeted gold and 70.93% for traps. All 6,380 chosen gold occurrences have counts 3,273/1,574/376/1,157 (LONG 24.03%). Source targeted-gold LONG is 5.10%: augmentation increases long-distance exposure, which C1/C2 retain. Role balancing does not equalize gold/trap distances.

Q4. Independent HN, Rebalanced-CN, and relocation.

Independent HN contains 2,000 instances (200/fold), 1,941 targets/1,814 documents, and 10,000 comparisons. Each instance has one gold and five negatives; multiple-gold targets use a fixed-index gold. The 59 repeated-target instances use different negative sets. Negatives are held-out original clauses excluding gold/emotion clauses and gold-text duplicates. Two annotators assess non-causality independently; a third adjudicates. Potential causes/ambiguous candidates are excluded. Five eligible negatives are sampled without replacement per instance. There is no within-fold training-parent overlap. Teacher/verifier outputs, the CPA pool, and evaluated-policy scores are excluded from negative selection.

The evaluated policy uses

sθ(c)=log⁡Pθ(是∣x,c)−log⁡Pθ(否∣x,c).s_\theta(c)=\log P_\theta(\text{是}\mid x,c)-\log P_\theta(\text{否}\mid x,c).

Accuracy is the fraction of 10,000 comparisons with sθ(cg)>sθ(cn)s_\theta(c_g)>s_\theta(c_n): 87.92%→91.35%. Comparisons sharing targets/documents are not independent samples.

Rebalanced-CN retains every held-out document containing a LONG pair and samples purely local documents with probability 0.30, using Random(42+f) in document-ID order. Full documents/all labels are retained; normalized-text matches to that fold's training documents are removed without replenishment. The shared view has 817 documents/940 gold pairs, including 314 LONG pairs (33.40%), with zero within-fold training parent/text overlap. Vanilla/CPA F1 is 74.65/79.95.

Relocation moves one existing distractor, excluding every gold/emotion clause, immediately before/after the target emotion or to the document start/end. Other clauses retain relative order; labels are remapped, with no rewriting/duplication. The 1,800 base documents yield 7,200 operation views, including 523 no-ops. A 200-view audit gives 194 valid / 6 ambiguous / 0 clearly invalid. Using the same clean/perturbed set per fold,

W=110∑f=110max⁡r∈R(Ff,clean−Ff,r).W=\frac1{10}\sum_{f=1}^{10}\max_{r\in\mathcal R}(F_{f,\mathrm{clean}}-F_{f,r}).

We take the worst drop within each fold before averaging: 4.27→2.72 F1 points for Vanilla-DPO→CPA-DPO, supporting reduced sensitivity under these operations.

AuthorsOpenReview ↗

Looking Forward to Further Discussion

Dear Reviewer PeU3,

Thank you again for your careful review and constructive suggestions. We sincerely appreciate the opportunity to clarify our methodology, strengthen the experimental validation, and improve the reproducibility of our work. We have carefully considered your concerns and provided detailed responses in our rebuttal.

For Q1 regarding reproducibility, we have clarified the exact verifier identity, its frozen configuration without task-specific training, and the complete ECE-SFT and DPO training settings.

For Q2 regarding verifier reliability and preference validity, we have provided additional validation of the verifier, a 300-instance human audit of the final augmented inputs, and direct evidence of increased difficulty for the target SFT policy under hardest-valid selection.

For Q3 regarding the contribution of preference construction, we have added controlled experiments, including chosen-only augmented SFT (C1), same-input generic-negative DPO (C2), and fixed-placement DPO (C3). These comparisons help distinguish the benefits of our preference construction from positional augmentation and simpler negative-selection strategies. We have also clarified the matched source coverage, training budgets, and final emotion–cause distance distributions.

For Q4 regarding evaluation protocols, we have provided detailed explanations of Independent HN construction and annotation, Rebalanced-CN fold isolation, and distractor-relocation operations and aggregation.

We hope these clarifications and additional experimental results address your main concerns. We would greatly appreciate your feedback on whether the new evidence has adequately addressed these issues, and please let us know if any concerns remain. We would be happy to provide further explanations and continue the discussion.

Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Addressing Key Concerns with New Evidence

Dear Area Chair and Reviewers,

Thank you for the careful and constructive reviews. The central concerns are whether CPA's preferences preserve the intended labels and whether their benefit exceeds augmentation, data selection, and additional supervision. Below, we map these concerns to the corresponding responses and the evidence bearing on each judgment.

1. Incremental value beyond augmentation and additional supervision.

PeU3 Q3; S3bs mechanism-validation question; 1diZ Q2; KE55 additional-supervision concern — new controlled evidence.

CPA-DPO improves Overall Pair-F1 from 83.62 to 85.43 (+1.81) over Vanilla-DPO. Our controls use the same 6,178 source-pair identities and fold-wise additional update counts (263 summed across ten folds). Against chosen-only training on identical augmented inputs (C1), CPA gains +1.31 Overall / +3.21 LONG F1; against same-input generic-negative DPO (C2), +1.01 / +2.50. These gains persist beyond the tested simpler uses of the augmented data. With candidates fixed, randomized placement improves over fixed placement (C3) by 0.37 / 0.96 F1, measuring the placement policy's overall contribution.

2. Target preservation and final-label validity in same-clause cases.

S3bs Q1; PeU3 Q2; 1diZ Q3 — protocol clarification plus final-input audit.

Our responses distinguish the source emotion ID/quote identifying the target from its current clause position; cause indices are remapped to the augmented document. The 100 swapped same-clause inputs have 92 valid / 8 ambiguous / 0 clearly invalid preferences; the full stratified audit of 300 final DPO inputs finds 286 / 14 / 0, respectively. This directly examines the special case underlying the label-preservation concern, beyond index checks or replacement-stage auditing, while retaining the observed ambiguities.

3. Direct evidence for verifier competence and target-policy difficulty.

KE55 verifier/hardness concerns; PeU3 Q1–Q2 — two separate direct tests.

Human-labeled emotion-support evaluation gives 88.53% balanced accuracy on 380 unambiguous replacement views. For the frozen pre-DPO policy, selecting maximum residual support among eligible candidates reduces the mean teacher-forced chosen–rejected log-probability margin from 0.430.43 to 0.310.31 and increases rejected-completion outranking from 18.79%18.79\% to 22.21%22.21\%, relative to random-valid. The first test evaluates verifier competence at emotion-support assessment; the second supports residual support as an informative, imperfect proxy for target-policy difficulty.

4. Evaluation independence and the meaning of cross-dataset results.

PeU3 Q4; S3bs Q3; 1diZ evaluation concerns — clarified protocols and additional checks.

Our responses specify teacher/verifier-independent HN construction (2,000 instances; 10,000 comparisons; accuracy 87.92%→91.35%87.92\% \rightarrow 91.35\%) and relocation aggregation (1,800 documents; mean fold-wise worst F1 drop 4.27→2.724.27 \rightarrow 2.72). ECPE-EN is an English-trained benchmark evaluation; its CPA–Vanilla gain remains +0.93 F1 after excluding exact training-text matches, versus +0.97 originally. RECCON-DD is in-domain evaluation; only DD-to-IE is zero-shot transfer. These distinctions delimit the generalization claims.

5. Contribution, backbone dependence, and reproducibility.

XbAv Q1–Q3 and presentation concerns; shared provenance concerns — broader evidence and explicit specifications.

Five additional backbones show +0.81–1.16 Overall / +2.66–3.56 LONG F1 over their respective Vanilla-DPO baselines, extending the evidence beyond the original model. Our responses discuss prior position-bias work and identify ECPE-specific preference construction under standard DPO as the contribution; matched controls provide the primary attribution evidence. For reproducibility and presentation, the responses specify the frozen Qwen2.5-32B-Instruct verifier and its absence of ECPE-specific fine-tuning, ECE-SFT settings, negative-construction rules, fold isolation, and the target-index and evaluation definitions questioned in the reviews.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.