Generation or Judgement? A Paradigm Perspective on LLM-Based Emotion-Cause Pair Extraction in Conversation

# Generation or Judgement? A Paradigm Perspective on LLM-Based Emotion–Cause Pair Extraction in Conversation
Source: [https://arxiv.org/html/2607.26967](https://arxiv.org/html/2607.26967)
###### Abstract

Emotion–cause pair extraction in conversation \(ECPEC\) identifies utterance pairs in which one utterance causes an emotion expressed in another\. Recent LLM\-based approaches formulate ECPEC at markedly different granularities, ranging from generating complete pair sets to judging individual candidate pairs\. In this paper, we make the surprising observation that task formulation substantially affects performance, where pair\-level judgement outperforms dialogue\-level generation in all 18 controlled comparisons\. We investigate the sources of this paradigm gap and find that many relations omitted by dialogue\-level generation remain recognizable under explicit pair queries, under which the model recognizes 92\.7%–98\.1% of emotion–cause relations\. This suggests that LLMs can recognize emotion–cause relations but struggle to discover and return complete pair sets\. Pair\-level judgement alleviates this burden, although its candidate rankings are more reliable than the binary decisions produced by a shared threshold\. Based on this diagnosis, we introduce an auxiliary retriever that selectively re\-examines ambiguous boundary cases, yielding consistent F1 improvements of 0\.50–1\.46 points across three datasets while maintaining an inference time of only 1\.49×\\timesthat of the baseline paradigm\. These findings show that task decomposition and candidate scope are critical to effectively utilizing LLMs for ECPEC\.

## 1Introduction

![Refer to caption](https://arxiv.org/html/2607.26967v1/x1.png)Figure 1:Paradigm gap and unrealized separation potential\.Judconsistently outperformsGen, while the shared decision boundary realizes only part of the separation permitted by candidate rankings\.Emotion–cause pair extraction in conversation \(ECPEC\) aims to identify emotion–cause utterance pairs, where a cause utterance triggers or explains the emotion conveyed by a target utterance\(Liet al\.[2023](https://arxiv.org/html/2607.26967#bib.bib29); Wanget al\.[2023](https://arxiv.org/html/2607.26967#bib.bib34)\)\. ECPEC may involve multiple speakers and multiple causes, requiring causal reasoning over conversational context\. Prior studies\(Jeong and Bak[2023](https://arxiv.org/html/2607.26967#bib.bib25); Liuet al\.[2023](https://arxiv.org/html/2607.26967#bib.bib31); Anet al\.[2023](https://arxiv.org/html/2607.26967#bib.bib22); Wanget al\.[2024a](https://arxiv.org/html/2607.26967#bib.bib35); Maet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib38)\)have primarily addressed this task using task\-specific architectures that explicitly model conversational context, speaker interactions, and emotion–cause dependencies\. Recently, LLMs have demonstrated strong language understanding and contextual reasoning capabilities\(Huang and Chang[2023](https://arxiv.org/html/2607.26967#bib.bib87); Wanget al\.[2024b](https://arxiv.org/html/2607.26967#bib.bib88); Louet al\.[2024](https://arxiv.org/html/2607.26967#bib.bib89)\), which provide a natural foundation for emotion–cause reasoning\. Thus, recent studies\(Luoet al\.[2024](https://arxiv.org/html/2607.26967#bib.bib79); Juet al\.[2025](https://arxiv.org/html/2607.26967#bib.bib37); Wanget al\.[2026](https://arxiv.org/html/2607.26967#bib.bib82); Wuet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib41)\)have increasingly explored LLM\-based approaches to ECPEC\.

Although LLM\-based ECPEC is still in its early stages, existing studies have explored markedly different ways of using LLMs\. Tu et al\.\(Tuet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib40)\), for example, evaluate GPT\-4 by asking it to predict the emotions and causes of all utterances from the complete conversation, whereas Luo et al\.\(Luoet al\.[2024](https://arxiv.org/html/2607.26967#bib.bib79)\)and Arefa et al\.\(Arefaet al\.[2024](https://arxiv.org/html/2607.26967#bib.bib77)\)decompose the task into emotion recognition followed by target\-specific cause prediction\. Ju et al\.\(Juet al\.[2025](https://arxiv.org/html/2607.26967#bib.bib37)\)formulate each target–candidate pair as a yes\-or\-no question and use GPT\-3\.5 to refine the candidate decision\. More elaborate systems embed LLMs into a speaker\-centric cognition–perception–reasoning agent\(Wanget al\.[2026](https://arxiv.org/html/2607.26967#bib.bib82)\)or use multi\-LLM debate to resolve emotion and pair predictions disputed by generative and discriminative models\(Wuet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib41)\)\. Despite this diversity, existing studies have not converged on a common way of posing ECPEC to LLMs, leaving open how to utilize their capabilities effectively for the task\.

![Refer to caption](https://arxiv.org/html/2607.26967v1/x2.png)Figure 2:Illustration of the three LLM\-based ECPEC paradigms\.To explore this open question, we rethink LLM\-based ECPEC from a paradigm perspective\. We distill existing approaches, according to their task\-decomposition granularity, into three representative paradigms: dialogue\-level generation \(Gen\), target\-level selection \(Sel\), and pair\-level judgement \(Jud\) \(§[2](https://arxiv.org/html/2607.26967#S2)\)\. Surprisingly, we find that paradigm choice leads to substantial differences in ECPEC performance, withJudconsistently outperformingGenacross LLM backbones and input modalities \(as shown in Figure[1](https://arxiv.org/html/2607.26967#S1.F1)\(a\)\), demonstrating the broad generality of this paradigm gap \(§[3](https://arxiv.org/html/2607.26967#S3)\)\.

To understand what drives this gap, we then delve into further analysis to examine how task decomposition affects the recognition of emotion–cause pairs and the recovery of complete cause sets \(§[4](https://arxiv.org/html/2607.26967#S4)–§[5](https://arxiv.org/html/2607.26967#S5)\)\. We find that many pairs omitted byGenremain recognizable under more explicit queries\. More generally, when directly asked whether a specified utterance causes the emotion in another, the LLM recognizes 92\.7%–98\.1% of gold emotion–cause pairs\. Moreover,Genmisses more pairs when a dialogue contains more gold pairs, a target utterance has multiple causes, or the cause is farther from the target\. These findings suggest thatGenstruggles mainly because it must discover and return all pairs at once, rather than because the LLM cannot recognize them\.Judreduces this burden by asking the LLM to examine one candidate at a time, but combining these decisions into a complete cause set remains difficult, where the model often scores gold causes above negative candidates, yet a single decision threshold cannot separate them reliably across targets \(Figure[1](https://arxiv.org/html/2607.26967#S1.F1)\(b\)\)\. Overall, the paradigm gap arises because finer\-grained task decomposition better exposes the pair\-recognition ability of LLMs, but this ability is more reliably reflected in candidate rankings than in binary judgements\.

Guided by these findings, we develop a lightweight remedy for the judgement paradigm that uses an auxiliary retriever to selectively re\-examine ambiguous cases around the decision boundary\. The remedy consistently improves F1 by 0\.50–1\.46 points across three datasets, showing that adapting decision boundaries is a viable improvement direction\. Finally, our efficiency analysis shows that restricting pair\-level queries to the local scope improves macro F1 overGen\-Full by 14\.47 points while limiting inference time to 1\.33×\\times, compared with 9\.08×\\timesforJud\-Full\. Our findings show that finer\-grained task decomposition, combined with a controlled candidate scope can better harness the reasoning capability of LLMs for ECPEC task while maintaining a practical balance between effectiveness and efficiency\.

## 2Background

This work focuses on the task ofemotion–cause pair extraction in conversation\(ECPEC\)\. Given a multi\-turn dialogue𝒟=\{\(ui,si,𝒙i\)\}i=1N\\mathcal\{D\}=\\\{\(u\_\{i\},s\_\{i\},\\bm\{x\}\_\{i\}\)\\\}\_\{i=1\}^\{N\}, whereuiu\_\{i\}denotes theii\-th utterance andsis\_\{i\}is speaker corresponding touiu\_\{i\},𝒙i\\bm\{x\}\_\{i\}contains textual features and, when available, acoustic and visual features\. The goal of ECPEC is to identify all emotion–cause pairs in𝒟\\mathcal\{D\}:

𝒫∗=\{\(ue,uc\)∣uc​causes the emotion in​ue\},\\mathcal\{P\}^\{\*\}=\\big\\\{\(u\_\{e\},u\_\{c\}\)\\mid u\_\{c\}\\text\{ causes the emotion in \}u\_\{e\}\\big\\\},\(1\)Notably, an emotion utterance may have multiple causes and may also serve as its own cause\. Encouraged by the strong reasoning capabilities of LLMs, recent ECPEC studies have increasingly explored LLM\-based solutions\. We distill LLM\-based ECPEC approaches into three paradigms according to their inference granularity \(Figure[2](https://arxiv.org/html/2607.26967#S1.F2)\), namely dialogue\-level generation, target\-level selection, and pair\-level judgement\.

#### Dialogue\-level generation\.

Dialogue\-level generation \(Gen\) takes the complete dialogue𝒟\\mathcal\{D\}as input and uses a single LLM call to jointly predict an emotion label and a set of cause utterances for every utteranceui∈𝒟u\_\{i\}\\in\\mathcal\{D\}:

\[\(i,yi,𝒞i\)\]i=1N=fGen​\(𝒟\),\\big\[\(i,y\_\{i\},\\mathcal\{C\}\_\{i\}\)\\big\]\_\{i=1\}^\{N\}=f^\{\\text\{\{Gen\}\}\}\(\\mathcal\{D\}\),\(2\)whereyiy\_\{i\}is the predicted emotion label ofuiu\_\{i\},𝒞i\\mathcal\{C\}\_\{i\}denotes the set of utterance indices predicted to cause the emotion expressed inuiu\_\{i\}, andfGenf^\{\\text\{\{Gen\}\}\}is the LLM inference function\.

#### Target\-level selection\.

Target\-level selection \(Sel\) decomposes the prediction for a dialogue𝒟\\mathcal\{D\}containingNNutterances intoNNutterance\-conditioned LLM calls\. Each call fixes one utteranceuiu\_\{i\}as the target, treats each eligible utteranceuj∈𝒟u\_\{j\}\\in\\mathcal\{D\}, includinguiu\_\{i\}itself, as a candidate cause, and predicts the emotion labelyiy\_\{i\}and cause\-index set𝒞i\\mathcal\{C\}\_\{i\}:

\(yi,𝒞i\)=fSel​\(𝒟,ui\)\.\(y\_\{i\},\\mathcal\{C\}\_\{i\}\)=f^\{\\text\{\{Sel\}\}\}\(\\mathcal\{D\},u\_\{i\}\)\.\(3\)

#### Pair\-level judgement\.

Pair\-level judgement \(Jud\) further decomposes the target\-conditioned prediction inSelinto independent binary decisions over individual target–candidate pairs\. For each target utteranceuiu\_\{i\}and each eligible candidate utteranceuju\_\{j\}, it uses a separate LLM call to determine whetheruju\_\{j\}causes the emotion expressed inuiu\_\{i\}:

fJud​\(𝒟,ui,uj\)∈\{yes,no\}\.f^\{\\text\{\{Jud\}\}\}\(\\mathcal\{D\},u\_\{i\},u\_\{j\}\)\\in\\\{\\texttt\{yes\},\\texttt\{no\}\\\}\.\(4\)AlthoughJudtheoretically incurs𝒪​\(N2\)\\mathcal\{O\}\(N^\{2\}\)LLM calls by exhaustively enumerating all target–candidate pairs in𝒟\\mathcal\{D\}, our later analysis shows that restricting the candidate causes for each targetuiu\_\{i\}to\{uj∣max⁡\(1,i−4\)≤j≤i\}\\\{u\_\{j\}\\\!\\mid\\\!\\max\(1,i\\\!\-\\\!4\)\\\!\\leq\\\!j\\\!\\leq\\\!i\\\}preserves performance while reducing the number of calls to𝒪​\(N\)\\mathcal\{O\}\(N\)\.

## 3How Does Task Paradigm Shape ECPEC?

This section first compares the three LLM\-based ECPEC paradigms against representative ECPEC methods and examines the performance differences among the paradigms \(§[3\.2](https://arxiv.org/html/2607.26967#S3.SS2)\)\. We then investigate whether these paradigm\-level differences persist across different experimental settings \(§[3\.3](https://arxiv.org/html/2607.26967#S3.SS3)\)\.

### 3\.1Experimental Setup

We evaluate the three paradigms on ECF\(Wanget al\.[2023](https://arxiv.org/html/2607.26967#bib.bib34)\), ConvECPE\(Liet al\.[2023](https://arxiv.org/html/2607.26967#bib.bib29)\), and RECCON\(Poriaet al\.[2021](https://arxiv.org/html/2607.26967#bib.bib32)\)\. LLaMA\-3\.1\-8B\(Llama Team[2024](https://arxiv.org/html/2607.26967#bib.bib84)\)serves as the primary backbone for all three paradigms, while Qwen\-2\.5\-7B\(Qwen Team[2025](https://arxiv.org/html/2607.26967#bib.bib85)\)and GLM\-4\-9B\(Team GLM[2024](https://arxiv.org/html/2607.26967#bib.bib86)\)are included to assess backbone generality\. Unless otherwise stated, all experiments use text\-only inputs\. A key experimental setting is the candidate scope, which determines the range of utterances considered as candidate causes for each target emotion utterance\. Across the three datasets, 87\.0%–93\.9% of gold cause utterances are either the target utterance itself or one of its four preceding utterances, as detailed in Appendix Table[10](https://arxiv.org/html/2607.26967#A2.T10)\. We therefore conduct all experiments under two candidate scopes: the local scope considers utterancesuju\_\{j\}satisfying0≤i−j≤40\\\!\\leq\\\!i\-j\\\!\\leq\\\!4for each targetuiu\_\{i\}, and the full scope considers every utterance in the dialogue, including both preceding and following utterances\. Further details on experiment protocols are provided in Appendix[B](https://arxiv.org/html/2607.26967#A2)\.

ECFConvECPERECCONMethodPRF1PRF1PRF1Conventional MethodMRC44\.4657\.6550\.2025\.9247\.0533\.4252\.1952\.8652\.47CENTER47\.9043\.6544\.7537\.9845\.6041\.4447\.3946\.8847\.13SCALE55\.0160\.6757\.7040\.3151\.7445\.3256\.3161\.6058\.83Prior LLM\-based MethodsDEC\-Debate♯69\.5155\.5261\.7350\.1850\.5450\.36\-\-\-GMEC\-GPT♯55\.9250\.9353\.31\-\-\-\-\-\-MSPF\+♯\-\-\-50\.3157\.9153\.8456\.2954\.9155\.59Controlled LLM ParadigmsGen\-Full60\.0453\.8256\.7628\.4523\.5525\.7750\.3655\.0152\.58Gen\-Local64\.9851\.3157\.3450\.3731\.7538\.9556\.5556\.4256\.49Sel\-Full64\.6957\.6160\.9452\.6255\.0953\.8352\.3357\.2254\.66Sel\-Local66\.5255\.1560\.3056\.1353\.3254\.6957\.6753\.2055\.34Jud\-Full51\.4674\.3260\.8133\.2779\.4846\.9052\.0165\.9358\.15Jud\-Local67\.7457\.3962\.1458\.1656\.9257\.5363\.6054\.7858\.86Table 1:Overall ECPEC performance comparison\.Shadingandunderliningindicate the best and second\-best results\.♯denotes the results taken from the original papers\.
### 3\.2LLMs Are Competitive for ECPEC

Table[1](https://arxiv.org/html/2607.26967#S3.T1)compares controlled implementations of the three LLM\-based paradigms with conventional ECPEC methods and prior LLM\-based methods\. The results first demonstrate that LLMs provide a competitive foundation for ECPEC\. Under the local scope,Judachieves F1 scores of 62\.14, 57\.53, and 58\.86 on ECF, ConvECPE, and RECCON, respectively, outperforming the strongest baselines in Table[1](https://arxiv.org/html/2607.26967#S3.T1), which obtain 61\.73, 53\.84, and 58\.83\. Thus, appropriately designed LLM paradigms can achieve state\-of\-the\-art ECPEC performance\. Performance, however, varies substantially across the three paradigms\. Under local scope, their F1 scores range from 57\.34 to 62\.14 on ECF, from 38\.95 to 57\.53 on ConvECPE, and from 55\.34 to 58\.86 on RECCON\. The corresponding ranges under full scope are 56\.76–60\.94, 25\.77–53\.83, and 52\.58–58\.15\.Judachieves the highest F1 in four of the six settings\. The exceptions are ECF and ConvECPE under full scope, whereSelachieves 60\.94 and 53\.83 F1 respectively, compared with 60\.81 and 46\.90 forJud\. The paradigm\-level difference is particularly pronounced on ConvECPE, where the F1 spread reaches 18\.58 points under local scope and 28\.06 points under full scope\. Overall, adopting an LLM alone does not guarantee ECPEC performance, while the paradigm through which the task is learned and performed critically shapes its effectiveness\.

LLaMA\-3\.1\-8BQwen\-2\.5\-7BGLM\-4\-9BInputLocalFullLocalFullLocalFullText\+4\.80\+4\.05\+5\.13\+8\.32\+1\.98\+5\.65Multimodal \+ MLP\+3\.41\+5\.75\+3\.77\+6\.60\+1\.90\+5\.25Multimodal \+ GraphSmile\+4\.94\+7\.81\+4\.46\+7\.11\+5\.49\+6\.48Table 2:F1 improvement ofJudoverGenacross settings on ECF\.
### 3\.3The Paradigm Gap Persists Across Settings

UsingGenandJudas the two endpoints of the task\-decomposition spectrum, we next examine whether the advantage of pair\-level judgement generalizes across backbone architectures and input representations\. Using ECF, which also provides multimodal dialogue data, we evaluate three LLM backbones under three input configurations: text\-only input, MLP\-fused multimodal features, and GraphSmile\-enhanced\(Liet al\.[2025b](https://arxiv.org/html/2607.26967#bib.bib66)\)multimodal representations\. For each configuration, we reportΔ​F​1=F​1Jud−F​1Gen\\Delta F1=F1\_\{\\textsc\{Jud\}\}\-F1\_\{\\textsc\{Gen\}\}\. As shown in Table[2](https://arxiv.org/html/2607.26967#S3.T2),Δ​F​1\\Delta F1is positive in all 18 backbone–input–scope combinations, ranging from 1\.90 to 8\.32 points\. The advantage persists with both multimodal input configurations and reaches 4\.46–7\.81 points with GraphSmile\-enhanced representations\. Thus, providing additional multimodal evidence does not eliminate the performance difference between the two paradigms\. Moreover, the advantage is larger under full scope than under local scope in eight of the nine backbone–input configurations, indicating that pair\-level decomposition generally becomes more beneficial as the candidate scope expands\.

#### Takeaway\. LLMs provide a competitive foundation for ECPEC, but their effectiveness varies substantially with task decomposition\. Pair\-level judgement \(Jud\) consistently outperforms dialogue\-level generation \(Gen\) across different settings\.

## 4Why DoesGenMiss Emotion–Cause Pairs?

This section diagnoses whyGenmisses valid emotion–cause pairs\. We first determine whether these missed pairs are beyond the LLM’s recognition ability or merely overlooked byGen\(§[4\.1](https://arxiv.org/html/2607.26967#S4.SS1)\)\. We then locate where the omissions arise in the generation process \(§[4\.2](https://arxiv.org/html/2607.26967#S4.SS2)\), and examine the conditions under whichGenis prone to omission \(§[4\.3](https://arxiv.org/html/2607.26967#S4.SS3)\)\.

![Refer to caption](https://arxiv.org/html/2607.26967v1/x3.png)Figure 3:Emotion–cause pairs omitted byGen\.### 4\.1GenOmits Pairs It Can Recognize

We ask whether the emotion–cause pairs omitted byGenare beyond LLM recognition or remain recoverable under finer\-grained task paradigms\. Section[3](https://arxiv.org/html/2607.26967#S3)shows thatSelandJudrecognize substantially more gold pairs thanGen, but aggregate recall does not reveal whether their additional true positives correspond to the emotion–cause pairs omitted byGen\. We therefore use both paradigms as references to re\-examine the omissions ofGen\. To evaluate this, we define the recoverable omission rate \(ROR, see Appendix[C](https://arxiv.org/html/2607.26967#A3)\) ofGenas the proportion of gold pairs missed byGenbut correctly identified by another paradigm\. Figure[3](https://arxiv.org/html/2607.26967#S4.F3)partitions the gold emotion–cause pairs according to whether they are recognized byGenand the corresponding reference paradigm, with the value to the right of each bar reporting the ROR\. Across the six dataset–scope conditions,Judrecovers 25\.8–74\.9% of the gold pairs omitted byGen, whereasSelrecovers 24\.3–51\.8%\. The ROR is higher withJudas the reference in all six conditions\. The highest RORs occur on ConvECPE, whereJudrecovers 54\.6% of the gold pairs omitted byGenunder local scope and 74\.9% under full scope\. In the latter setting,Judrecovers 1,097 of the 1,464 gold pairs missed byGen, whereasGenitself correctly identifies only 451 gold pairs\. These results show that many emotion–cause pairs omitted byGencan still be identified by LLMs in practice, but are not reliably exposed through holistic generation\.

![Refer to caption](https://arxiv.org/html/2607.26967v1/x4.png)Figure 4:Results under progressive task decomposition\.
### 4\.2Omissions Arise During Set\-Level Prediction

We next locate where these omissions arise by progressively exposing the four key decisions hidden withinGen\. The first asks the model to generate all emotion–cause pairs directly from the dialogue\. The second provides a gold emotion target and its emotion label, removing target discovery and emotion\-label prediction\. The third additionally enumerates all candidate causes, replacing implicit cause search with candidate\-set prediction\. The final condition presents one target–candidate pair at a time, removing the need to jointly select and return a complete cause set\. We probe the same LLaMA\-3\.1\-8B checkpoint, which is jointly fine\-tuned on the three gold\-target formats, with dialogue generation evaluated as an additional probe\. As shown in Figure[4](https://arxiv.org/html/2607.26967#S4.F4), providing the gold target raises pair recall by 1\.51–15\.67 points across the six dataset–scope conditions, while listing the candidate utterances adds another 3\.20–24\.67 points\. These gains suggest that target and candidate discovery account for part of the omissions\. The largest increase, however, consistently occurs when candidate\-set prediction is decomposed into isolated pair judgements, improving recall by 30\.83–55\.04 points and yielding final recall of 92\.72%–98\.07%\. This pattern locates the dominant omission bottleneck in jointly selecting and returning a complete cause set\.

ECFConvECPERECCONMetricLowMidHighLowMidHighLowMidHigh\#Dialogues887779101110629271Avg\. \#Gold Pairs2\.56\.614\.537\.458\.989\.34\.26\.912\.3GenRecall55\.361\.150\.343\.018\.419\.162\.859\.449\.5SelRecall59\.062\.555\.249\.253\.558\.765\.561\.651\.5JudRecall69\.177\.673\.871\.979\.582\.673\.270\.760\.3GenP/G1\.020\.960\.840\.950\.780\.811\.801\.110\.87Table 3:Performance versus the number of gold pairs\.
### 4\.3When IsGenMore Likely to Miss Pairs?

We next examine conditions under whichGenis more likely to miss gold emotion–cause pairs\.

#### Number of gold pairs\.

We first examine whetherGenomits more gold pairs as the number of pairs in a dialogue increases\. Within each dataset, we partition dialogues into approximately equal\-sized low\-, medium\-, and high\-count groups according to their numbers of gold pairs\. We report recall to measure the proportion of gold pairs recovered and the predicted\-to\-gold ratio \(P/G\) to assess whetherGen’s output\-set size grows proportionally with the gold set\. As shown in Table[3](https://arxiv.org/html/2607.26967#S4.T3), recall under full\-scope in the high\-count group is lower than in the low\-count group by 5\.0–23\.9 points\. P/G likewise decreases from 1\.02 to 0\.84, from 0\.95 to 0\.81, and from 1\.80 to 0\.87\. These decreases indicate that, in high\-count dialogues,Genrecovers a smaller proportion of the required pairs and produces fewer predictions relative to the gold\-set size\. The cross\-paradigm comparison is most revealing on ConvECPE, whereGenrecall decreases by 23\.9 points, whereas the recalls ofSelandJudincrease by 9\.5 and 10\.7 points, respectively\. By contrast, all three paradigms decline on RECCON, suggesting that high\-count dialogues in this dataset are also intrinsically more difficult\.

Any CauseCause RecallComplete\#Causes12≥\\geq312≥\\geq312≥\\geq3Gen52\.261\.361\.352\.243\.429\.852\.225\.57\.8Sel67\.177\.478\.267\.155\.939\.367\.134\.37\.7Jud82\.289\.789\.182\.274\.354\.982\.258\.823\.2Table 4:Cause recovery \(%\) versus the number of causes\.
#### Number of gold causes\.

We next examine whetherGenbecomes less complete as the number of gold causes per target increases\. We group targets by cause count and report whether at least one cause is recovered \(Any Cause\), the proportion recovered \(Cause Recall\), and whether all causes are recovered \(Complete\)\. Table[4](https://arxiv.org/html/2607.26967#S4.T4)reports the average result for three datasets under full\-scope,Gen’s any\-cause recovery remains stable, while its cause recall falls from 52\.2% to 29\.8% and complete recovery from 52\.2% to 7\.8%\. Among targets with at least three causes, it recovers some but not all causes in 53\.5% of cases, indicating difficulty in constructing a complete cause set\. This problem affects all three paradigms, althoughJudmitigates it with 54\.9% cause recall and 23\.2% complete recovery in the same group\.

![Refer to caption](https://arxiv.org/html/2607.26967v1/x5.png)Figure 5:Pair recall \(%\) versus emotion–cause distances\.
#### Emotion–cause distance\.

Finally, we examine whetherGenis more likely to miss a gold pair when its emotion target and cause are farther apart\. We report pair recall for local \(d≤4d\\leq 4\) and long\-range \(d\>4d\>4\) gold pairs under full\-scope, wheredddenotes the utterance\-index distance\. As shown in Figure[5](https://arxiv.org/html/2607.26967#S4.F5),Gen’s recall decreases from local to long\-range pairs by 33\.4, 12\.6, and 46\.3 points on the three datasets\. Long\-range pairs are therefore more difficult forGen, as they are for the other two paradigms\. However, even among local pairs,JudoutperformsGenby 20\.8, 61\.2, and 11\.0 points across the three datasets\. Thus, long distance aggravatesGen’s omissions but cannot explain them alone, as substantial omissions persist even when the cause is nearby\.

Takeaway\.LLMs can recognize emotion–cause pairs, butGenomits many during set\-level prediction, especially with more gold pairs per dialogue, more gold causes per target, or longer emotion–cause distances\.

## 5Why DoesJudWork? When Does It Fail?

This section first briefly summarizes why pair\-levelJud, at the opposite end of the task\-decomposition spectrum from dialogue\-levelGen, is more effective \(§[5\.1](https://arxiv.org/html/2607.26967#S5.SS1)\)\. We then investigate whereJudreaches its limits from two perspectives\. We first examine whetherJudcorrectly ranks and separates gold causes from negative candidates within a given candidate scope, and then analyze how its pair\-level errors accumulate as the scope expands \(§[5\.2](https://arxiv.org/html/2607.26967#S5.SS2)\)\.

### 5\.1Explicit Pair Query Reduce Extraction Burden

Genmust discover emotion targets and their causes before jointly selecting and returning a complete pair set\. By explicitly specifying both the target and candidate cause,Judinstead reduces each query to judging one explicit emotion–cause pair\. As shown in Table[1](https://arxiv.org/html/2607.26967#S3.T1)and Figure[4](https://arxiv.org/html/2607.26967#S4.F4), many gold pairs omitted under coarser paradigms are recovered when presented individually to the same checkpoint\. Together with the analysis in Section[4](https://arxiv.org/html/2607.26967#S4), these results show thatJudis effective because it removes the set\-level prediction bottleneck and more directly elicits the LLM’s pair\-recognition capability\.

### 5\.2The Capability Boundary ofJud

Expanding the candidate scope from local to full covers more gold pairs and improvesJud’s recall, but substantially lowers its precision and ultimately reduces F1\. This recall–precision trade\-off suggests that the advantage ofJudreaches a limit\. We locate this boundary from two perspectives: whether gold causes are correctly ranked above and separated from negatives, and how pair\-level errors accumulate as the candidate space expands\.

ECFConvECPERECCONOutcomeLocalFullLocalFullLocalFullCase\-160\.3374\.3159\.7079\.5555\.9068\.98Case\-22\.842\.126\.291\.874\.102\.18Case\-329\.3220\.2920\.3316\.2127\.8623\.61Case\-47\.513\.2713\.692\.3712\.145\.24ISR80\.4874\.5265\.8050\.1271\.3170\.08RSR47\.4143\.8234\.0310\.1737\.9442\.50Table 5:Pairwise ranking and threshold outcomes \(%\)\.#### Ranking versus thresholding\.

To determine whether this degradation arises mainly from ranking or thresholding, we examine the judgement score andYes\-or\-Nodecision assigned byJudto each candidate utterance\. Reliable cause\-set prediction requires every gold cause to outrank every negative candidate and the decision threshold to separate them\. For each target utterance, we therefore compare every gold cause with every negative candidate and divide the comparisons into four cases: gold cause is accepted and the negative candidate rejected \(Case\-1\); gold cause scores higher, but both are accepted \(Case\-2\); gold cause scores higher, but both are rejected \(Case\-3\); or gold cause is ranked below or tied with the negative candidate \(Case\-4\)\. As shown in Table[5](https://arxiv.org/html/2607.26967#S5.T5), Case\-4 accounts for only 2\.37%–13\.69% of comparisons, whereas Case\-2 and Case\-3 together account for 18\.08%–32\.16% and are more frequent in every setting\. Case\-3 also consistently dominates Case\-2, showing that thresholding errors more often reject correctly ranked gold causes than accept negatives\. Thus, failures arise more often from failing to separate correctly ordered gold–negative comparisons than from ranking them incorrectly\. Pairwise comparisons, however, do not show whether all candidates for a target can be separated simultaneously\. We therefore define the Ideal Separability Ratio \(ISR\) as the proportion of targets whose lowest\-scoring gold cause outranks their highest\-scoring negative\. An oracle target\-specific threshold could fully separate these targets without changing their rankings\. ISR ranges from 50\.12% to 80\.48%, whereas the Realized Separability Ratio \(RSR\), which measures complete separation under the threshold used byJud, reaches only 10\.17%–47\.41%\. The resulting 27\.58–39\.95 point gap identifies target utterances whose rankings are already separable but whose cause sets are not realized by the threshold\.

ECFConvECPERECCONMetricLocalAddedFullLocalAddedFullLocalAddedFull\#Gold1,7581151,8731,6662491,9151,6551121,767\#Neg\.8,54323,02631,5696,13986,29392,4328,12019,34627,466TP1,361311,3921,446761,5221,140251,165TPR \(%\)77\.4226\.9674\.3286\.7930\.5279\.4868\.8822\.3265\.93FP1,223901,3132,6204333,0531,011641,075FPR \(%\)14\.320\.394\.1642\.680\.503\.3012\.450\.333\.91Precision \(%\)52\.6725\.6251\.4635\.5614\.9333\.2753\.0028\.0952\.01Recall \(%\)72\.661\.6674\.3275\.513\.9779\.4864\.521\.4165\.93Table 6:Error accumulation under candidate\-space expansion\.
#### Error accumulation\.

We next ask whether pair\-level errors accumulate when many independently judged candidates are combined into cause sets\. We partition the candidates from the same full\-scope run into Local \(0≤i−j≤40\\\!\\leq\\\!i\\\!\-\\\!j\\\!\\leq\\\!4\) and Added \(i−j≥5i\\\!\-\\\!j\\\!\\geq\\\!5ori−j<0i\\\!\-\\\!j\\\!<\\\!0\) regions\. As shown in Table[6](https://arxiv.org/html/2607.26967#S5.T6), the Added region ofJudhas an FPR of only 0\.33%–0\.50%, substantially below the Local FPR of 12\.45%–42\.68%\. However, the Added region contains 19,346–86,293 negatives, 2\.38–14\.06 times as many as the Local region, but only 112–249 gold pairs\. Consequently, this low per\-pair FPR still produces 64–433 additional false positives for only 25–76 true positives\. These predictions contribute just 1\.41–3\.97 points to full\-scope recall while lowering precision by 0\.99–2\.29 points\. Although 85\.8%–94\.0% of full\-scope false positives originate from the harder Local region, candidate expansion introduces a secondary accumulation effect: even rare errors become consequential when independent judgements are made over a large negative space\.

#### Takeaway\. Judis effective because it reduces ECPEC to recognizing one explicit pair at a time\. However, it remains constrained by a shared decision boundary and error accumulation as the candidate space expands\.

## 6A Diagnosis\-Guided Remedy

Based on the above analyses, we develop a lightweight remedy around three design choices\.1We adopt pair\-levelJudrather than dialogue\-levelGen, as explicit pair queries better expose the relation recognition capability of LLMs\.2We use local as the default candidate scope because it retains most gold pairs while limiting the error accumulation and inference cost caused by full\-scope expansion\.3We introduce a localized approximation to target\-specific boundary adaptation\. Directly estimating a reliable target\-specific boundary is difficult, while Table[5](https://arxiv.org/html/2607.26967#S5.T5)shows that 16\.21%–29\.32% of instances fall into Case\-3, where gold causes are ranked above negative candidates yet both are rejected\. This prevalent under\-selection motivates a localized correction that reconsiders only the highest\-ranked rejected candidate\. We implement this correction using an auxiliary retriever initialized from the trainedJudcheckpoint and further fine\-tuned on a mixture of original pair\-judgement instances and boundary\-candidate instances\. At inference, the retriever re\-examines the highest\-ranked rejected candidate\. If accepted, the candidate is added to the original prediction; otherwise, the prediction remains unchanged\.

MetricECFConvECPERECCONJudF162\.1457\.5358\.86\+ Retriever F163\.60\(\+1\.46\)58\.03\(\+0\.50\)59\.39\(\+0\.53\)Δ\\DeltaPrecision\-2\.50\-0\.93\-3\.13Δ\\DeltaRecall\+4\.65\+1\.93\+3\.57Correct additions rate44\.938\.934\.4Table 7:Effectiveness of the proposed remedy under local scope\. Correct additions rate denotes the percentage of newly accepted pairs that are gold pairs\.#### Effectiveness\.

We evaluate the final correction against its pipeline\-alignedJud\-Local baseline for each dataset\. As shown in Table[7](https://arxiv.org/html/2607.26967#S6.T7), the correction yields modest but consistent Pair\-F1 improvements of 0\.50–1\.46 points across the three datasets, increasing macro F1 by 0\.83 points\. The gains are recall\-driven: recall increases by 1\.93–4\.65 points, while precision decreases by 0\.93–3\.13 points\. Among the candidates newly accepted by the retriever, 34\.4%–44\.9% are gold causes\. Although the accepted additions contain false positives, the recovered gold causes are sufficient to produce a net F1 improvement on every dataset\. The overall gains remain incremental because the correction reconsiders only one rejected boundary candidate per target and recovers only 11\.71%–19\.12% of the gold causes presented for reconsideration\. Nevertheless, the consistent improvements obtained without changing the original candidate rankings show that the shared decision boundary is an actionable source of error\. More comprehensive estimation of target\-specific decision boundaries therefore represents a promising direction for future ECPEC research\.

ConfigurationQueriesInput tokensOutput tokensTimeGen\-Full1\.01\.00×\\times1\.00×\\times1\.00×\\timesGen\-Local1\.01\.03×\\times0\.99×\\times0\.98×\\timesSel\-Full12\.822\.00×\\times1\.27×\\times1\.39×\\timesSel\-Local12\.822\.32×\\times1\.25×\\times1\.36×\\timesJud\-Full303\.7792\.27×\\times2\.16×\\times9\.08×\\timesJud\-Local53\.988\.90×\\times0\.38×\\times1\.33×\\times\+ Retriever66\.2111\.19×\\times0\.38×\\times1\.49×\\timesTable 8:Inference\-cost comparison of the selected configurations\. Input tokens, output tokens, and inference time are normalized toGen\-Full\.
#### Inference efficiency\.

Beyond effectiveness, another practical concern is the inference cost of the three paradigms and the proposed remedy\. Table[8](https://arxiv.org/html/2607.26967#S6.T8)compares their inference costs\. AlthoughJud\-Full incurs substantially higher costs thanGen\-Full, restricting candidate causes to the local scope reduces its inference time from 9\.08×\\timesto 1\.33×\\times\. Notably, despite processing 88\.90×\\timesmore input tokens,Jud\-Local consumes only 0\.38×\\timesas many output tokens because each query requires only a short binary response\. The retriever uses next\-token classification without text generation, leaving the output\-token cost unchanged\. It increases the number of queries and input tokens by 22\.8% and 25\.1% overJud\-Local, respectively, while adding only 11\.8% inference time\. Its total inference time remains 1\.49×\\timesthat ofGen\-Full, far below the 9\.08×\\timescost ofJud\-Full\. Thus, the remedy trades a measurable but controlled increase in inference cost for consistent F1 gains\.

These gains nevertheless remain modest because the remedy performs only a localized correction rather than directly estimating a reliable target\-specific decision boundary\. Its main value is therefore to demonstrate that the shared decision boundary is an actionable source of error, while more comprehensive boundary adaptation remains an open challenge\. The effectiveness and efficiency results jointly show that utilizing the existing pair\-recognition ability of LLMs requires appropriate task decomposition and controlled candidate scopes, as well as decision mechanisms that can reliably translate this ability into final predictions\.

## 7Related Work

#### Conventional ECPEC methods\.

Emotion–cause analysis originated from emotion cause extraction, which identifies the cause of a given emotion\(Leeet al\.[2010](https://arxiv.org/html/2607.26967#bib.bib10); Guiet al\.[2016](https://arxiv.org/html/2607.26967#bib.bib9)\)\. Xia and Ding\(Xia and Ding[2019](https://arxiv.org/html/2607.26967#bib.bib18)\)later introduced emotion–cause pair extraction \(ECPE\) to jointly identify emotions and their causes\. Subsequent studies developed pipeline, end\-to\-end, graph\-based, sequence\-labeling, and machine\-reading\-comprehension methods for document\-level ECPE\(Dinget al\.[2020](https://arxiv.org/html/2607.26967#bib.bib6); Chenet al\.[2020](https://arxiv.org/html/2607.26967#bib.bib3); Fanet al\.[2021](https://arxiv.org/html/2607.26967#bib.bib8); Chenet al\.[2022](https://arxiv.org/html/2607.26967#bib.bib5); Zhuet al\.[2024](https://arxiv.org/html/2607.26967#bib.bib20)\)\. Poria et al\.\(Poriaet al\.[2021](https://arxiv.org/html/2607.26967#bib.bib32)\)introduced emotion\-cause recognition in conversations through RECCON, and Li et al\.\(Liet al\.[2023](https://arxiv.org/html/2607.26967#bib.bib29)\)subsequently formalized ECPEC as the joint extraction of emotion–cause utterance pairs and released the ConvECPE dataset\. Later ECPEC methods capture pair relations, speaker interactions, dialogue context, and event structure using mixture\-of\-experts, machine\-reading\-comprehension, and graph\-based architectures\(Jeong and Bak[2023](https://arxiv.org/html/2607.26967#bib.bib25); Liuet al\.[2023](https://arxiv.org/html/2607.26967#bib.bib31); Anet al\.[2023](https://arxiv.org/html/2607.26967#bib.bib22); Wanget al\.[2024a](https://arxiv.org/html/2607.26967#bib.bib35)\)\. Multimodal extensions further incorporate acoustic and visual information and strengthen cross\-modal interaction, task coupling, and global pair modeling\(Wanget al\.[2023](https://arxiv.org/html/2607.26967#bib.bib34); Liet al\.[2024](https://arxiv.org/html/2607.26967#bib.bib30); Huet al\.[2024](https://arxiv.org/html/2607.26967#bib.bib80); Liet al\.[2025a](https://arxiv.org/html/2607.26967#bib.bib81); Tuet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib40); Xieet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib39); Maet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib38)\)\. This line of work primarily advances task\-specific architectures, whereas we study how different inference paradigms shape the use of LLMs for ECPEC\.

#### LLM\-based ECPEC methods\.

LLMs have recently been incorporated into ECPEC as primary predictors, auxiliary reasoners, or components of hybrid systems\. Several SemEval\-2024 systems use GPT\-based emotion recognition followed by a separate cause extractor\(Levchenkoet al\.[2024](https://arxiv.org/html/2607.26967#bib.bib75); Kazakovet al\.[2024](https://arxiv.org/html/2607.26967#bib.bib76)\), while others use LLMs for multi\-stage cause prediction or instruction\-tuned joint prediction\(Arefaet al\.[2024](https://arxiv.org/html/2607.26967#bib.bib77); Chenget al\.[2024](https://arxiv.org/html/2607.26967#bib.bib78); Luoet al\.[2024](https://arxiv.org/html/2607.26967#bib.bib79)\)\. Ju et al\.\(Juet al\.[2025](https://arxiv.org/html/2607.26967#bib.bib37)\)develop a single\-stage generative framework augmented with pair\-level yes\-or\-no decisions obtained from GPT\-3\.5\. At the dialogue level, Tu et al\.\(Tuet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib40)\)evaluate GPT\-4 by asking it to predict emotions and causes for the complete conversation\. Wu et al\.\(Wuet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib41)\)instead generate causes for individual targets and apply pair\-level debate to disputed candidates\. More elaborate systems also embed LLMs into multi\-stage agent frameworks\(Wanget al\.[2026](https://arxiv.org/html/2607.26967#bib.bib82)\)\. These approaches often combine multiple inference granularities and differ substantially in backbone, modality, supervision, and candidate scope\. We therefore distill their inference operations into dialogue\-level generation, target\-level selection, and pair\-level judgement, and compare these paradigms under controlled settings\.

## 8Conclusion

This work rethinks LLM\-based ECPEC from a paradigm perspective by systematically comparing dialogue\-level generation, target\-level selection, and pair\-level judgement\. We identify a consistent paradigm gap: finer\-grained task decomposition better exposes the ability of LLMs to recognize emotion–cause pairs, allowing pair\-level judgement to consistently outperform dialogue\-level generation\. Our analyses show that dialogue\-level generation is primarily limited by the burden of discovering and returning complete pair sets, whereas pair\-level judgement is limited by the difficulty of converting reliable candidate rankings into binary decisions under a shared decision boundary\. Guided by this diagnosis, we develop an auxiliary retriever that selectively re\-examines ambiguous boundary cases, yielding consistent F1 improvements across all three datasets\. Overall, our analysis suggests that, among the evaluated configurations,Jud\-Local offers the most practical foundation for LLM\-based ECPEC, balancing the performance benefits of pair\-level decomposition against its inference cost\. Developing reliable target\-specific decision boundaries remains an important direction for translating strong candidate rankings into more accurate emotion–cause pair predictions\. We leave this direction to future work\.

## References

- J\. An, Z\. Ding, K\. Li, and R\. Xia \(2023\)Global\-view and speaker\-aware emotion cause extraction in conversations\.IEEE/ACM Transactions on Audio, Speech, and Language Processing31,pp\. 3814–3823\.External Links:ISSN 2329\-9304,[Document](https://dx.doi.org/10.1109/TASLP.2023.3319990)Cited by:[§B\.3](https://arxiv.org/html/2607.26967#A2.SS3.p2.1),[§1](https://arxiv.org/html/2607.26967#S1.p1.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- Arefa, M\. A\. Ansari, C\. Saxena, and T\. Ahmad \(2024\)JMI at SemEval 2024 task 3: two\-step approach for multimodal ECAC using in\-context learning with GPT and instruction\-tuned llama models\.InProceedings of the 18th International Workshop on Semantic Evaluation \(SemEval\-2024\),A\. Kr\. Ojha, A\. S\. Doğruöz, H\. Tayyar Madabushi, G\. Da San Martino, S\. Rosenthal, and A\. Rosá \(Eds\.\),Mexico City, Mexico,pp\. 1561–1576\.External Links:[Link](https://aclanthology.org/2024.semeval-1.223/),[Document](https://dx.doi.org/10.18653/v1/2024.semeval-1.223)Cited by:[§1](https://arxiv.org/html/2607.26967#S1.p2.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px2.p1.1)\.
- C\. Busso, M\. Bulut, C\. Lee, A\. Kazemzadeh, E\. Mower, S\. Kim, J\. N\. Chang, S\. Lee, and S\. S\. Narayanan \(2008\)IEMOCAP: interactive emotional dyadic motion capture database\.Language Resources and Evaluation42\(4\),pp\. 335–359\.External Links:ISSN 1574\-020X, 1574\-0218,[Document](https://dx.doi.org/10.1007/s10579-008-9076-6)Cited by:[§B\.1](https://arxiv.org/html/2607.26967#A2.SS1.SSS0.Px2.p1.1)\.
- S\. Chen, X\. Shi, J\. Li, S\. Wu, H\. Fei, F\. Li, and D\. Ji \(2022\)Joint alignment of multi\-task feature and label spaces for emotion cause pair extraction\.InProceedings of the 29th International Conference on Computational Linguistics,N\. Calzolari, C\. Huang, H\. Kim, J\. Pustejovsky, L\. Wanner, K\. Choi, P\. Ryu, H\. Chen, L\. Donatelli, H\. Ji, S\. Kurohashi, P\. Paggio, N\. Xue, S\. Kim, Y\. Hahm, Z\. He, T\. K\. Lee, E\. Santus, F\. Bond, and S\. Na \(Eds\.\),Gyeongju, Republic of Korea,pp\. 6955–6965\.External Links:[Link](https://aclanthology.org/2022.coling-1.606/)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- Y\. Chen, W\. Hou, S\. Li, C\. Wu, and X\. Zhang \(2020\)End\-to\-end emotion\-cause pair extraction with graph convolutional network\.InProceedings of the 28th International Conference on Computational Linguistics,D\. Scott, N\. Bel, and C\. Zong \(Eds\.\),Barcelona, Spain \(Online\),pp\. 198–207\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.coling-main.17),[Link](https://aclanthology.org/2020.coling-main.17/)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- Z\. Cheng, F\. Niu, Y\. Lin, Z\. Cheng, X\. Peng, and B\. Zhang \(2024\)MIPS at SemEval\-2024 task 3: multimodal emotion\-cause pair extraction in conversations with multimodal language models\.InProceedings of the 18th International Workshop on Semantic Evaluation \(SemEval\-2024\),A\. Kr\. Ojha, A\. S\. Doğruöz, H\. Tayyar Madabushi, G\. Da San Martino, S\. Rosenthal, and A\. Rosá \(Eds\.\),Mexico City, Mexico,pp\. 667–674\.External Links:[Link](https://aclanthology.org/2024.semeval-1.97/),[Document](https://dx.doi.org/10.18653/v1/2024.semeval-1.97)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px2.p1.1)\.
- Z\. Ding, R\. Xia, and J\. Yu \(2020\)End\-to\-end emotion\-cause pair extraction based on sliding window multi\-label learning\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),B\. Webber, T\. Cohn, Y\. He, and Y\. Liu \(Eds\.\),Online,pp\. 3574–3583\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.290),[Link](https://aclanthology.org/2020.emnlp-main.290/)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- C\. Fan, C\. Yuan, L\. Gui, Y\. Zhang, and R\. Xu \(2021\)Multi\-task sequence tagging for emotion\-cause pair extraction via tag distribution refinement\.IEEE/ACM Transactions on Audio, Speech, and Language Processing29,pp\. 2339–2350\.External Links:ISSN 2329\-9304,[Document](https://dx.doi.org/10.1109/TASLP.2021.3089837),[Link](https://ieeexplore.ieee.org/document/9457144/)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- L\. Gui, D\. Wu, R\. Xu, Q\. Lu, and Y\. Zhou \(2016\)Event\-driven emotion cause extraction with corpus construction\.InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing,J\. Su, K\. Duh, and X\. Carreras \(Eds\.\),Austin, Texas,pp\. 1639–1649\.External Links:[Document](https://dx.doi.org/10.18653/v1/D16-1170),[Link](https://aclanthology.org/D16-1170/)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen \(2022\)LoRA: low\-rank adaptation of large language models\.InInternational Conference on Learning Representations,Cited by:[§B\.4](https://arxiv.org/html/2607.26967#A2.SS4.SSS0.Px1.p1.1)\.
- G\. Hu, Z\. Zhu, D\. Hershcovich, L\. Hu, H\. Seifi, and J\. Xie \(2024\)UniMEEC: towards unified multimodal emotion recognition and emotion cause\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 5248–5261\.External Links:[Link](https://aclanthology.org/2024.findings-emnlp.302/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-emnlp.302)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- J\. Huang and K\. C\. Chang \(2023\)Towards reasoning in large language models: a survey\.InFindings of the Association for Computational Linguistics: ACL 2023,A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 1049–1065\.External Links:[Link](https://aclanthology.org/2023.findings-acl.67/),[Document](https://dx.doi.org/10.18653/v1/2023.findings-acl.67)Cited by:[§1](https://arxiv.org/html/2607.26967#S1.p1.1)\.
- D\. Jeong and J\. Bak \(2023\)Conversational emotion\-cause pair extraction with guided mixture of experts\.InProceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics,A\. Vlachos and I\. Augenstein \(Eds\.\),Dubrovnik, Croatia,pp\. 3288–3298\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.eacl-main.240)Cited by:[§B\.3](https://arxiv.org/html/2607.26967#A2.SS3.p2.1),[§1](https://arxiv.org/html/2607.26967#S1.p1.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- X\. Ju, D\. Zhang, J\. Li, S\. Li, and G\. Zhou \(2025\)Enhanced generative framework with llms for multimodal emotion\-cause pair extraction in conversations\.IEEE Transactions on Multimedia27\(\),pp\. 4924–4935\.External Links:[Document](https://dx.doi.org/10.1109/TMM.2025.3543080)Cited by:[§B\.3](https://arxiv.org/html/2607.26967#A2.SS3.p1.1),[§1](https://arxiv.org/html/2607.26967#S1.p1.1),[§1](https://arxiv.org/html/2607.26967#S1.p2.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px2.p1.1)\.
- R\. Kazakov, K\. Petukhova, and E\. Kochmar \(2024\)PetKaz at SemEval\-2024 task 3: advancing emotion classification with an LLM for emotion\-cause pair extraction in conversations\.InProceedings of the 18th International Workshop on Semantic Evaluation \(SemEval\-2024\),A\. Kr\. Ojha, A\. S\. Doğruöz, H\. Tayyar Madabushi, G\. Da San Martino, S\. Rosenthal, and A\. Rosá \(Eds\.\),Mexico City, Mexico,pp\. 1127–1134\.External Links:[Link](https://aclanthology.org/2024.semeval-1.164/),[Document](https://dx.doi.org/10.18653/v1/2024.semeval-1.164)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px2.p1.1)\.
- D\. Kong, N\. Yu, Y\. Yuan, G\. Fu, and C\. Gong \(2022\)Discourse\-aware emotion cause extraction in conversations\.arXiv\.External Links:2210\.14419,[Document](https://dx.doi.org/10.48550/arXiv.2210.14419)Cited by:[§B\.3](https://arxiv.org/html/2607.26967#A2.SS3.p2.1)\.
- S\. Y\. M\. Lee, Y\. Chen, and C\. Huang \(2010\)A text\-driven rule\-based system for emotion cause detection\.InProceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text,D\. Inkpen and C\. Strapparava \(Eds\.\),Los Angeles, CA,pp\. 45–53\.External Links:[Link](https://aclanthology.org/W10-0206/)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- S\. Levchenko, R\. Wolert, and P\. Andruszkiewicz \(2024\)PWEITINLP at SemEval\-2024 task 3: two step emotion cause analysis\.InProceedings of the 18th International Workshop on Semantic Evaluation \(SemEval\-2024\),A\. Kr\. Ojha, A\. S\. Doğruöz, H\. Tayyar Madabushi, G\. Da San Martino, S\. Rosenthal, and A\. Rosá \(Eds\.\),Mexico City, Mexico,pp\. 1097–1105\.External Links:[Link](https://aclanthology.org/2024.semeval-1.159/),[Document](https://dx.doi.org/10.18653/v1/2024.semeval-1.159)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px2.p1.1)\.
- B\. Li, H\. Fei, F\. Li, T\. Chua, and D\. Ji \(2024\)Multimodal emotion\-cause pair extraction with holistic interaction and label constraint\.ACM Transactions on Multimedia Computing, Communications, and Applications21\(11\),pp\. 1–19\.External Links:ISSN 1551\-6857,[Document](https://dx.doi.org/10.1145/3689646)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- G\. Li, X\. Duan, C\. Wang, and S\. Peng \(2025a\)An adaptive hybrid machine reading comprehension framework for multimodal emotion\-cause pair extraction in conversations\.Engineering Applications of Artificial Intelligence162,pp\. 112333\.External Links:ISSN 0952\-1976,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.engappai.2025.112333),[Link](https://www.sciencedirect.com/science/article/pii/S0952197625023413)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- J\. Li, X\. Wang, and Z\. Zeng \(2025b\)Tracing intricate cues in dialogue: joint graph structure and sentiment dynamics for multimodal emotion recognition\.IEEE Transactions on Pattern Analysis and Machine Intelligence47\(10\),pp\. 8786–8803\.External Links:ISSN 1939\-3539,[Document](https://dx.doi.org/10.1109/TPAMI.2025.3581236)Cited by:[§3\.3](https://arxiv.org/html/2607.26967#S3.SS3.p1.2)\.
- W\. Li, Y\. Li, V\. Pandelea, M\. Ge, L\. Zhu, and E\. Cambria \(2023\)ECPEC: emotion\-cause pair extraction in conversations\.IEEE Transactions on Affective Computing14\(3\),pp\. 1754–1765\.External Links:ISSN 1949\-3045,[Document](https://dx.doi.org/10.1109/TAFFC.2022.3216551)Cited by:[§B\.1](https://arxiv.org/html/2607.26967#A2.SS1.p1.1),[§1](https://arxiv.org/html/2607.26967#S1.p1.1),[§3\.1](https://arxiv.org/html/2607.26967#S3.SS1.p1.3),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- Y\. Li, H\. Su, X\. Shen, W\. Li, Z\. Cao, and S\. Niu \(2017\)DailyDialog: a manually labelled multi\-turn dialogue dataset\.InProceedings of the Eighth International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\),G\. Kondrak and T\. Watanabe \(Eds\.\),Taipei, Taiwan,pp\. 986–995\.External Links:[Link](https://aclanthology.org/I17-1099/)Cited by:[§B\.1](https://arxiv.org/html/2607.26967#A2.SS1.SSS0.Px3.p1.1)\.
- C\. Liu, C\. Niu, J\. Xie, Y\. Jia, and H\. Zan \(2023\)Emotion\-cause pair extraction in conversations based on multi\-turn mrc with position\-aware gcn\.In2023 International Conference on Asian Language Processing \(IALP\),pp\. 25–30\.External Links:ISSN 2159\-1970,[Document](https://dx.doi.org/10.1109/IALP61005.2023.10337182)Cited by:[§B\.3](https://arxiv.org/html/2607.26967#A2.SS3.p1.1),[§1](https://arxiv.org/html/2607.26967#S1.p1.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- Llama Team \(2024\)The llama 3 herd of models\.External Links:2407\.21783,[Link](https://arxiv.org/abs/2407.21783)Cited by:[§B\.4](https://arxiv.org/html/2607.26967#A2.SS4.SSS0.Px1.p1.1),[§3\.1](https://arxiv.org/html/2607.26967#S3.SS1.p1.3)\.
- R\. Lou, K\. Zhang, and W\. Yin \(2024\)Large language model instruction following: a survey of progresses and challenges\.Computational Linguistics50\(3\),pp\. 1053–1095\.Cited by:[§1](https://arxiv.org/html/2607.26967#S1.p1.1)\.
- M\. Luo, H\. Zhang, S\. Wu, B\. Li, H\. Han, and H\. Fei \(2024\)NUS\-emo at SemEval\-2024 task 3: instruction\-tuning LLM for multimodal emotion\-cause analysis in conversations\.InProceedings of the 18th International Workshop on Semantic Evaluation \(SemEval\-2024\),A\. Kr\. Ojha, A\. S\. Doğruöz, H\. Tayyar Madabushi, G\. Da San Martino, S\. Rosenthal, and A\. Rosá \(Eds\.\),Mexico City, Mexico,pp\. 1589–1596\.External Links:[Link](https://aclanthology.org/2024.semeval-1.226/),[Document](https://dx.doi.org/10.18653/v1/2024.semeval-1.226)Cited by:[§1](https://arxiv.org/html/2607.26967#S1.p1.1),[§1](https://arxiv.org/html/2607.26967#S1.p2.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px2.p1.1)\.
- T\. Ma, W\. Feng, X\. Wang, and Z\. Cheng \(2026\)Emotion–cause pair extraction in conversations via semantic decoupling and alignment\.InFindings of the Association for Computational Linguistics: ACL 2026,M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),pp\. 12697–12711\.External Links:[Link](https://aclanthology.org/2026.findings-acl.617/),ISBN 979\-8\-89176\-395\-1Cited by:[§B\.3](https://arxiv.org/html/2607.26967#A2.SS3.p1.1),[§1](https://arxiv.org/html/2607.26967#S1.p1.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- S\. Poria, N\. Majumder, D\. Hazarika, D\. Ghosal, R\. Bhardwaj, S\. Y\. B\. Jian, P\. Hong, R\. Ghosh, A\. Roy, N\. Chhaya, A\. Gelbukh, and R\. Mihalcea \(2021\)Recognizing emotion cause in conversations\.Cognitive Computation13\(5\),pp\. 1317–1332\.External Links:ISSN 1866\-9956, 1866\-9964,[Document](https://dx.doi.org/10.1007/s12559-021-09925-7)Cited by:[§B\.1](https://arxiv.org/html/2607.26967#A2.SS1.p1.1),[§3\.1](https://arxiv.org/html/2607.26967#S3.SS1.p1.3),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- Qwen Team \(2025\)Qwen2\.5 technical report\.External Links:2412\.15115,[Link](https://arxiv.org/abs/2412.15115)Cited by:[§B\.4](https://arxiv.org/html/2607.26967#A2.SS4.SSS0.Px1.p1.1),[§3\.1](https://arxiv.org/html/2607.26967#S3.SS1.p1.3)\.
- Team GLM \(2024\)ChatGLM: a family of large language models from glm\-130b to glm\-4 all tools\.External Links:2406\.12793,[Link](https://arxiv.org/abs/2406.12793)Cited by:[§B\.4](https://arxiv.org/html/2607.26967#A2.SS4.SSS0.Px1.p1.1),[§3\.1](https://arxiv.org/html/2607.26967#S3.SS1.p1.3)\.
- G\. Tu, J\. Wang, L\. Yang, B\. Liang, E\. Cambria, W\. Li, and R\. Xu \(2026\)Multi\-task mutual learning for multimodal emotion\-cause pair extraction in conversations\.Information Fusion127,pp\. 103877\.External Links:ISSN 15662535,[Document](https://dx.doi.org/10.1016/j.inffus.2025.103877)Cited by:[§1](https://arxiv.org/html/2607.26967#S1.p2.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px2.p1.1)\.
- B\. Wang, K\. Tang, and P\. Zhu \(2024a\)Enhancing emotion\-cause pair extraction in conversations via center event detection and reasoning\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 10773–10783\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.findings-emnlp.632)Cited by:[§B\.3](https://arxiv.org/html/2607.26967#A2.SS3.p1.1),[§1](https://arxiv.org/html/2607.26967#S1.p1.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- B\. Wang, L\. Wu, S\. Guo, K\. Wang, Q\. Wang, T\. Zhang, J\. Liu, and H\. U\. Khan \(2026\)HCPRA: a hierarchical cognition–perception–reasoning agent framework for emotion\-cause pair extraction in conversations\.InProceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval,pp\. 1892–1902\.External Links:[Document](https://dx.doi.org/10.1145/3805712.3809665),ISBN 979\-8\-4007\-2599\-9Cited by:[§B\.3](https://arxiv.org/html/2607.26967#A2.SS3.p2.1),[§1](https://arxiv.org/html/2607.26967#S1.p1.1),[§1](https://arxiv.org/html/2607.26967#S1.p2.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px2.p1.1)\.
- F\. Wang, Z\. Ding, R\. Xia, Z\. Li, and J\. Yu \(2023\)Multimodal emotion\-cause pair extraction in conversations\.IEEE Transactions on Affective Computing14\(3\),pp\. 1832–1844\.External Links:ISSN 1949\-3045,[Document](https://dx.doi.org/10.1109/TAFFC.2022.3226559)Cited by:[§B\.1](https://arxiv.org/html/2607.26967#A2.SS1.p1.1),[§1](https://arxiv.org/html/2607.26967#S1.p1.1),[§3\.1](https://arxiv.org/html/2607.26967#S3.SS1.p1.3),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- X\. Wang, M\. Salmani, P\. Omidi, X\. Ren, M\. Rezagholizadeh, and A\. Eshaghi \(2024b\)Beyond the limits: a survey of techniques to extend the context length in large language models\.InProceedings of the Thirty\-Third International Joint Conference on Artificial Intelligence, IJCAI\-24,K\. Larson \(Ed\.\),pp\. 8299–8307\.Note:Survey TrackExternal Links:[Document](https://dx.doi.org/10.24963/ijcai.2024/917),[Link](https://doi.org/10.24963/ijcai.2024/917)Cited by:[§1](https://arxiv.org/html/2607.26967#S1.p1.1)\.
- S\. Wu, J\. Lin, X\. Wang, D\. Jiang, and E\. Cambria \(2026\)Bridging discrepancies in debate: joint debate for enhancing multimodal emotion\-cause pair extraction in conversation\.IEEE Transactions on Affective Computing,pp\. 1–15\.External Links:ISSN 1949\-3045, 2371\-9850,[Document](https://dx.doi.org/10.1109/TAFFC.2026.3681902)Cited by:[§B\.3](https://arxiv.org/html/2607.26967#A2.SS3.p1.1),[§1](https://arxiv.org/html/2607.26967#S1.p1.1),[§1](https://arxiv.org/html/2607.26967#S1.p2.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px2.p1.1)\.
- R\. Xia and Z\. Ding \(2019\)Emotion\-cause pair extraction: a new task to emotion analysis in texts\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics,A\. Korhonen, D\. Traum, and L\. Màrquez \(Eds\.\),Florence, Italy,pp\. 1003–1012\.External Links:[Document](https://dx.doi.org/10.18653/v1/P19-1096),[Link](https://aclanthology.org/P19-1096/)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- B\. Xie, J\. Wang, H\. Guo, P\. Chen, H\. Zhang, B\. Jiang, Y\. Wang, and L\. Chen \(2026\)MSPF: a multi\-semantic prompting fusion framework for emotion\-cause pair extraction in conversations\.Information Processing & Management63\(1\),pp\. 104356\.External Links:ISSN 03064573,[Document](https://dx.doi.org/10.1016/j.ipm.2025.104356)Cited by:[§B\.3](https://arxiv.org/html/2607.26967#A2.SS3.p1.1),[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.
- P\. Zhu, B\. Wang, K\. Tang, H\. Zhang, X\. Cui, and Z\. Wang \(2024\)A knowledge\-guided graph attention network for emotion\-cause pair extraction\.Knowledge\-Based Systems286,pp\. 111342\.External Links:ISSN 09507051,[Document](https://dx.doi.org/10.1016/j.knosys.2023.111342),[Link](https://linkinghub.elsevier.com/retrieve/pii/S0950705123010900)Cited by:[§7](https://arxiv.org/html/2607.26967#S7.SS0.SSS0.Px1.p1.1)\.

## Appendix APrompt Templates

This section presents the prompt templates used for the three controlled paradigms\.

Prompt for dialogue\-level generation \(Gen\)You are an expert in emotion\-\-cause pair extraction in conversations\.Given the following conversation, identify the emotion expressed in each utterance and the utterance IDs that cause that emotion\.Emotion labels must be selected from:\[EMOTION LABEL SET\]Rules:1\. If an utterance is neutral, set "cause\_ids" to \[\]\.2\. If an utterance expresses a non\-neutral emotion, "cause\_ids" should contain the IDs of all utterances that trigger or explain that emotion\.3\. A cause may be the emotion utterance itself\.4\.\[SCOPE INSTRUCTION\]5\. Return only a valid JSON list without any explanation\.Conversation:\[DIALOGUE\]Output format:\[ \{"utterance\_id": \[ID\], "emotion": "\[EMOTION\]", "cause\_ids": \[\[CAUSE IDS\]\]\} \]

Prompt for target\-level selection \(Sel\)You are an expert in emotion\-\-cause pair extraction in conversations\.Given the following conversation and target utterance, identify the emotion expressed in the target and select all utterances that cause that emotion\.Emotion labels must be selected from:\[EMOTION LABEL SET\]Rules:1\. If the target utterance is neutral, \[IDS\] must be \[\]\.2\. If the target utterance expresses a non\-neutral emotion, \[IDS\] should contain the IDs of all utterances that trigger or explain that emotion\.3\. A cause may be the target utterance itself\.4\.\[SCOPE INSTRUCTION\]5\. Return only the filled template without any explanation\.Conversation:\[DIALOGUE\]Target utterance:\[TARGET ID\]\.\[TARGET SPEAKER\]:\[TARGET TEXT\]Output template:\[TARGET TEXT\] contains \[CLASS\] emotion, corresponding to \[CAUSE IDS\]\.

Prompt for pair\-level judgement \(Jud\)You are an expert in emotion\-\-cause pair extraction in conversations\.Given the following conversation, determine whether the candidate cause utterance triggers or explains a non\-neutral emotion expressed in the target utterance\.A valid cause may be an event, statement, behavior, or contextual clue that triggers or explains the target emotion\.Conversation:\[DIALOGUE\]Target utterance:\[TARGET ID\]\.\[TARGET SPEAKER\]:\[TARGET TEXT\]Candidate cause utterance:\[CANDIDATE ID\]\.\[CANDIDATE SPEAKER\]:\[CANDIDATE TEXT\]Question:Does the candidate cause utterance cause a non\-neutral emotion expressed in the target utterance?Answer only "yes" or "no"\.Answer:

For local scope,\[SCOPE INSTRUCTION\]is replaced with:

> Only consider the target utterance itself and up to four preceding utterances as possible causes\.

For full scope, it is replaced with:

> Consider every utterance in the conversation as a possible cause, regardless of its position relative to the target utterance\.

## Appendix BDetailed Experimental Protocol

### B\.1Datasets

We conduct experiments on ECF\(Wanget al\.[2023](https://arxiv.org/html/2607.26967#bib.bib34)\), ConvECPE\(Liet al\.[2023](https://arxiv.org/html/2607.26967#bib.bib29)\), and RECCON\(Poriaet al\.[2021](https://arxiv.org/html/2607.26967#bib.bib32)\)\. These benchmarks differ substantially in dialogue length, pair density, and the number of annotated emotion–cause pairs, providing complementary evaluation conditions\. Table[9](https://arxiv.org/html/2607.26967#A2.T9)reports the statistics of three benchmarks\.

StatisticRECCONConvECPEECF\#Dialogue1,1061511,374\#Utterance11,1047,43313,619\#Emotion–Cause Pairs9,1249,4739,794Train/Val/Test75/5/2080/–/2070/10/20Table 9:Statistics of three datasets\.#### ECF\.

ECF is a multimodal ECPEC benchmark derived from the television seriesFriends\. It provides textual, acoustic, and visual information together with utterance\-level emotion labels and emotion–cause pair annotations\. Our primary experiments use only the dialogue transcripts so that the three paradigms receive the same textual evidence\. The additional multimodal experiment uses the acoustic and visual inputs provided by the benchmark\.

#### ConvECPE\.

ConvECPE is constructed from IEMOCAP\(Bussoet al\.[2008](https://arxiv.org/html/2607.26967#bib.bib60)\)and contains relatively long dyadic conversations\. Compared with ECF and RECCON, its test dialogues are substantially longer and therefore induce a much larger candidate space under the full scope\. This property makes ConvECPE particularly useful for examining omission under set generation and false\-positive accumulation under pair\-level judgement\.

#### RECCON\.

RECCON is a widely used benchmark for emotion\-cause pair extraction and consists of two subsets\. RECCON\-DD, annotated from DailyDialog\(Liet al\.[2017](https://arxiv.org/html/2607.26967#bib.bib51)\), serves as the primary corpus for model training and evaluation, while RECCON\-IE, annotated from IEMOCAP, is a smaller subset used exclusively to assess model generalization\. In this work, we choice the RECCON\-DD subset only and construct each gold pair from an annotated emotion utterance and one of its annotated cause utterances\.

### B\.2Candidate Scope

For each target emotion utteranceuiu\_\{i\}, the candidate scope determines which utterances are considered as potential causes\. Under local scope, the candidate set is restricted to\{uj∣0≤i−j≤4\}\\\{u\_\{j\}\\\!\\mid\\\!0\\\!\\leq\\\!i\\\!\-\\\!j\\\!\\leq\\\!4\\\}, which includes the target itself and its four preceding utterances\. Under full scope, every utterance in the dialogue is treated as a candidate cause, including utterances occurring after the target\.

StatisticECFConvECPERECCON\-DD\#Gold Pairs1,8731,9151,767Mean\|d\|\|d\|0\.831\.851\.380≤d≤40\\leq d\\leq 493\.9%87\.0%93\.7%Table 10:Utterance\-distance statistics of gold emotion–cause pairs, whered=i−jd=i\-jfor target utteranceuiu\_\{i\}and cause utteranceuju\_\{j\}\.Table[10](https://arxiv.org/html/2607.26967#A2.T10)reports the utterance\-distance distribution of gold emotion–cause pairs\. Most gold causes occur close to their target emotion utterances\. Specifically, 93\.9%, 87\.0%, and 93\.7% of the gold pairs in ECF, ConvECPE, and RECCON\-DD, respectively, satisfy0≤i−j≤40\\leq i\-j\\leq 4\. Based on this consistent locality pattern, we use local scope as a empirical setting shared across datasets\. We also report results under full scope to evaluate the three paradigms\. Under local scope, gold pairs outside the candidate scope remain part of the evaluation and are counted as false negatives if not predicted\.

### B\.3Competitive Baselines

We compare the three controlled LLM paradigms with six representative ECPEC baselines\.MRC\(Liuet al\.[2023](https://arxiv.org/html/2607.26967#bib.bib31)\)reformulates ECPEC as a three\-stage machine reading comprehension process that sequentially identifies emotions, extracts their causes, and verifies candidate pairs\. A position\-aware GCN models dialogue structure and speaker interactions\.CENTER\(Wanget al\.[2024a](https://arxiv.org/html/2607.26967#bib.bib35)\)detects center events and constructs event\-aware graphs to model contextual regions and interactions among candidate emotion–cause pairs\. It jointly optimizes center event detection, emotion and cause extraction, and pair prediction\.SCALE\(Maet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib38)\)models global contextual, local temporal, and intra\-speaker dependencies using a shared conversation graph\. It derives emotion\- and cause\-oriented representations and aligns them through optimal transport\.GMEC\-GPT\(Juet al\.[2025](https://arxiv.org/html/2607.26967#bib.bib37)\)enhances the generative GMEC framework by using GPT\-3\.5 to refine its predictions\. The refinement incorporates prediction confidence, retrieved examples, and visual descriptions generated by MiniGPT\-4\.DEC\-Debate\(Wuet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib41)\)combines an LLM\-based generative model with a multimodal discriminative model\. Their conflicting emotion and pair predictions are resolved through multi\-round LLM debates and majority voting\.MSPF\+\(Xieet al\.[2026](https://arxiv.org/html/2607.26967#bib.bib39)\)performs pair verification by fusing tagging\-based, synonym\-based, and causal\-claim prompt representations\. It further filters candidate pairs using a validation\-selected utterance\-distance threshold\.

Since this work focuses on comparing LLM task paradigms rather than providing an exhaustive ECPEC leaderboard, we include only a compact set of representative baselines\. We therefore omit additional agent\-based LLM systems, such asHCPRA\(Wanget al\.[2026](https://arxiv.org/html/2607.26967#bib.bib82)\), and task\-specific architectures, such asDAM\(Konget al\.[2022](https://arxiv.org/html/2607.26967#bib.bib26)\),GSECE\-UC\(Anet al\.[2023](https://arxiv.org/html/2607.26967#bib.bib22)\), andPRG\-MoE\(Jeong and Bak[2023](https://arxiv.org/html/2607.26967#bib.bib25)\)\.

### B\.4Reproducibility

#### Training details\.

We use LLaMA\-3\.1\-8B\(Llama Team[2024](https://arxiv.org/html/2607.26967#bib.bib84)\)as the primary backbone and further evaluate Qwen\-2\.5\-7B\(Qwen Team[2025](https://arxiv.org/html/2607.26967#bib.bib85)\)and GLM\-4\-9B\(Team GLM[2024](https://arxiv.org/html/2607.26967#bib.bib86)\)\. For the paradigm\-specific comparison, a separate LoRA adapter\(Huet al\.[2022](https://arxiv.org/html/2607.26967#bib.bib83)\)is fine\-tuned for each backbone and task paradigm using the same training split\. All experiments use deterministic single\-pass inference with temperature0\. ForJud, a candidate pair is predicted as positive when the normalized probability of theYestoken is greater than 0\.5; ties are assigned to No\. The threshold is fixed across datasets and is not tuned on validation or test data\. The training hyperparameters are summarized in Table[11](https://arxiv.org/html/2607.26967#A2.T11)\.

LLaMA\-3\.1\-8BQwen\-2\.5\-7BGLM\-4\-9BHyperparameterGenSelJudGenSelJudGenSelJudLoRArr88168816888LoRAα\\alpha161632161632161616Dropout0\.000\.000\.050\.000\.000\.050\.000\.000\.05Learning rate5e\-55e\-55e\-55e\-55e\-55e\-55e\-55e\-55e\-5Batch size1 \(8\)1 \(8\)1 \(8\)1 \(8\)1 \(8\)1 \(8\)1 \(8\)1 \(8\)1 \(8\)Epochs662662662Table 11:Hyperparameters for paradigm\-specific fine\-tuning\.
#### Implementation\.

Our implementation uses Python 3\.10\.20, PyTorch 2\.11\.0, Transformers 5\.9\.0, PEFT 0\.19\.1, and CUDA 12\.8\. All experiments are conducted on a server equipped with eight NVIDIA GeForce RTX 5090 GPUs \(32 GB each\) with two Intel Xeon Platinum 8457C CPUs and 1 TB of system memory\. We use a random seed of 13 and report results from a single run\.

#### Evaluation metrics\.

Predictions from all three paradigms are converted into sets of ordered emotion–cause utterance\-index pairs\. Duplicate pairs are removed, and indices outside the dialogue or the evaluated candidate scope are discarded\. A prediction is correct only when both utterance indices exactly match a gold pair\. Let𝒫\\mathcal\{P\}and𝒢\\mathcal\{G\}denote the predicted and gold pair sets over the complete test set\. We compute

P=\|𝒫∩𝒢\|\|𝒫\|,R=\|𝒫∩𝒢\|\|𝒢\|,F1=2​P​RP\+R\.P=\\frac\{\|\\mathcal\{P\}\\cap\\mathcal\{G\}\|\}\{\|\\mathcal\{P\}\|\},\\quad R=\\frac\{\|\\mathcal\{P\}\\cap\\mathcal\{G\}\|\}\{\|\\mathcal\{G\}\|\},\\quad F\_\{1\}=\\frac\{2PR\}\{P\+R\}\.\(5\)

## Appendix CDefinition of Recoverable Omission Rates

For each experimental condition, let𝒴\\mathcal\{Y\}denote the gold relation set\. We use the subscriptsGG,SS, andJJto refer toGen,Sel, andJud, respectively, and let𝒴^G\\widehat\{\\mathcal\{Y\}\}\_\{G\},𝒴^S\\widehat\{\\mathcal\{Y\}\}\_\{S\}, and𝒴^J\\widehat\{\\mathcal\{Y\}\}\_\{J\}denote their predicted relation sets\. The gold relations omitted byGenare defined as

F​NG=𝒴∖𝒴^G\.FN\_\{G\}=\\mathcal\{Y\}\\setminus\\widehat\{\\mathcal\{Y\}\}\_\{G\}\.\(6\)The gold relations correctly identified bySelandJudare

T​PS=𝒴∩𝒴^S,T​PJ=𝒴∩𝒴^J\.TP\_\{S\}=\\mathcal\{Y\}\\cap\\widehat\{\\mathcal\{Y\}\}\_\{S\},\\qquad TP\_\{J\}=\\mathcal\{Y\}\\cap\\widehat\{\\mathcal\{Y\}\}\_\{J\}\.\(7\)
#### Gen→\\rightarrowSel\.

The recoverable omission rate fromGentoSelmeasures the proportion of gold relations omitted byGenbut correctly identified bySel:

ROG→S=\|F​NG∩T​PS\|\|F​NG\|\.\\mathrm\{RO\}\_\{G\\rightarrow S\}=\\frac\{\|FN\_\{G\}\\cap TP\_\{S\}\|\}\{\|FN\_\{G\}\|\}\.\(8\)

#### Gen→\\rightarrowJud\.

The recoverable omission rate fromGentoJudmeasures the proportion of gold relations omitted byGenbut correctly identified byJud:

ROG→J=\|F​NG∩T​PJ\|\|F​NG\|\.\\mathrm\{RO\}\_\{G\\rightarrow J\}=\\frac\{\|FN\_\{G\}\\cap TP\_\{J\}\|\}\{\|FN\_\{G\}\|\}\.\(9\)
Both rates use the same set ofGenfalse negatives as the denominator and are therefore directly comparable\. A higher rate indicates that a larger proportion of the relations omitted byGenremain identifiable under the corresponding finer\-grained paradigm\. All sets are aggregated over the test set before computing the rates\.

## Appendix DUse of AI Assistants

AI assistants were used only for language polishing, formatting assistance, and consistency checking\. All research ideas, experimental designs, implementations, analyses, and conclusions were developed and verified by the authors\.