---
title: 'ECPE: Emotion-Cause Pair Extraction'
url: https://www.emergentmind.com/topics/emotion-cause-pair-extraction-ecpe
type: topic
---

# ECPE: Emotion-Cause Pair Extraction

Emotion-Cause Pair Extraction (ECPE) is the task of extracting all potential pairs of emotions and their corresponding causes in a document without pre-annotated emotion clauses, extending Emotion Cause Extraction (ECE) from cause retrieval conditioned on a given emotion to joint identification of emotion clauses, cause clauses, and their pairings [1906.01267]. In the literature summarized here, ECPE is usually formulated at the clause level, although later work generalizes it to span-level extraction, dialog-level quadruples, and conversational alignment; methodologically, the field has moved from two-step multi-task pipelines to end-to-end link prediction, graph reasoning, machine reading comprehension, question answering, and instruction-tuned large language models with explicit knowledge injection [2003.03507][2306.03969][2604.19547][2507.14887].

## 1. Origins, formal definition, and task scope

The task was introduced to address two shortcomings of earlier ECE formulations: the emotion must be annotated before cause extraction, and the procedure of first annotating emotion and then extracting the cause ignores the fact that emotion and cause are mutually indicative [1906.01267]. In its canonical clause-level form, a document is written as \( d = [c_1, c_2, ..., c_{|d|}] \), and the target output is a set of emotion-cause pairs,
\[
P = \{ \cdots, (c^e, c^c ), \cdots \},
\]
where \( c^e \) is an emotion clause and \( c^c \) is a corresponding cause clause [1906.01267]. The original benchmark construction reported 1,945 documents, and 89.77% have a single emotion-cause pair, a distributional property that later became central to discussions of sparsity and positional bias [1906.01267].

This clause-level definition remained the default reference point for several years, but subsequent work made its granularity and scope explicit. ECSP states that emotion expressions and causes are often spans rather than whole clauses, and that ECPE and ECE can be regarded as two special cases at the clause-level [2003.03507]. Later dialog and conversation settings preserved the core objective of linking emotion-bearing text to its cause, while changing the structural unit from clauses to utterances and, in some cases, adding emotion type and cause type labels [2306.03969][2604.19547]. This suggests that ECPE is best understood as a family of structurally related extraction problems rather than a single immutable benchmark definition.

## 2. From two-step pipelines to end-to-end extraction

The earliest ECPE systems used a two-step architecture. The original method first performed individual emotion extraction and cause extraction via multi-task learning and then conducted emotion-cause pairing and filtering with a logistic regression classifier over candidate pairs \(P_{\text{all}} = E \times C\) [1906.01267]. On that benchmark, the Inter-EC variant reported an ECPE F1 of 0.6128, improving over Indep at 0.5818 and showing that interaction between emotion and cause extraction improves performance [1906.01267].

A major shift was the move to end-to-end modeling. E2EECPE formulated pair extraction as a directional link prediction task with biaffine attention and auxiliary emotion and cause extraction tasks; on the Chinese news benchmark it reported pair extraction \(P=0.6478\), \(R=0.6105\), \(F1=0.6280\), and was 6–7 times faster per epoch than two-stage baselines [2002.10710]. An End-to-End Network for Emotion-Cause Pair Extraction later adapted the NTCIR-13 ECE corpus to ECPE in English and reported 50.17 F1 for pair extraction, improving over ECPE 2-stage at 43.67 by approximately 6.5 points while using about 790K trainable parameters compared with about 6.37M for ECPE-MLL [2103.01544].

Two-stage architectures nevertheless remained competitive when they explicitly addressed error propagation and contextual matching. The Dual-Questioning Attention Network questioned candidate emotions and causes to the context independently through attention networks and reported 0.6362 F1 on the ECPE benchmark, against 0.6205 for Inter-EC-W [2104.07221]. Emotion Prediction Oriented ECPE made emotion prediction the organizing signal, introduced synchronization between emotion prediction and pair extraction, and separated genuine pair supervision from fake pair supervision; with BERT it reported \(P=0.7621\), \(R=0.7519\), \(F1=0.7564\), compared with 0.7452 for ECPE-MLL and 0.7202 for PairGCN [2302.12417]. A plausible implication is that the field did not abandon pipelines so much as demand tighter coupling between clause identification and pair reasoning.

## 3. Structural reasoning, graph modeling, and positional bias

A recurring empirical problem in ECPE is that the relative distance distribution of emotions and causes is extremely imbalanced in the typical ECPE dataset, which encourages models to overfit local proximity rather than document semantics [2205.02132]. MGSAG addressed this with a Multi-Granularity Semantic Aware Graph that jointly incorporated fine-grained keyword–clause interactions and coarse-grained clause–clause relations without position features. On the standard Chinese ECPE dataset it reported EC Pair F1 0.6846, above ECPE-MLL at 0.6740, and on the position-insensitive split it reported 0.4301 F1 on \(Test_{NoBias}\), compared with 0.3988 for ECPE-MLL [2205.02132].

Other work attacked the same issue through stronger cross-task alignment. A²Net introduced Feature-Task Alignment with a Partition Filter Network and Inter-Task Alignment via bidirectional KL-divergence between ECPE predictions and pseudo ECPE labels derived from EE and CE outputs. On the Chinese benchmark it reported 76.34% F1 for ECPE, exceeding MGSAG’s 75.21% and ECPE-MLL’s 74.52%, while also reporting 90.80% for EE and 78.35% for CE [2209.04112]. PBJE argued that sequential feature encoding creates an imbalance in inter-task feature interaction and therefore generated clause and pair features simultaneously within a heterogeneous undirected graph processed by an RGCN; it reported Pair F1 76.37 and CE F1 78.78 on the Chinese benchmark [2212.01844].

Several models reframed the task to avoid the \(n^2\) pairing matrix or to inject explicit clause semantics. MM-R transformed ECPE into a document-level machine reading comprehension problem with a three-turn extraction-and-rethink process and reported Pair F1 80.62, Emotion F1 93.70, and Cause F1 81.35 on the benchmark corpus [2209.07972]. EA-GAT defined outer-clause, inter-clause, and intra-clause relationships and proposed a clause-level encoder that could be integrated into previous ECPE architectures, reporting average gains of 2.1% on the Chinese benchmark and 1.03% on the English benchmark corpus [2208.13549]. CGR-Net added co-evolving reasoning with bidirectional feedback between ECPE and its subtasks on a Multi-Task Relational Graph and reported ECPE F1 76.48, EE F1 87.75, and CE F1 78.75 [2306.04340]. MM-ECPE(BERT) combined BERT, sentiment lexicon signals, a position-aware interaction module layer, and knowledge graph filtering, reporting pair extraction \(P=77.48\), \(R=76.35\), \(F1=76.91\), and on the imbalanced test set F1 45.46 [2404.06812].

Taken together, these results show that “structure” in ECPE has been operationalized in several distinct ways: graph topology, clause-to-clause relation typing, multi-task label-space alignment, MRC query flow, and explicit position-aware attention. The common theme is not merely richer encoders, but deliberate resistance to shortcut learning from local distance alone.

## 4. Question answering and large language models

Question answering formulations treat ECPE as span or clause retrieval conditioned on a textual query. Guided-QA cast ECPE as extractive QA: first predict the best emotion clause using a fixed question, then use the predicted emotion as a question to predict the most potential cause. On the 10-split ECPE dataset, Guided-QA with RoBERTa reported Emotion F1 0.884, Cause F1 0.767, and EC Pair F1 0.744, while Table 4 reported a runtime of 2h30 for Guided-QA versus 8.5h for ECPE-MLL on Tesla P100 [2301.01982]. The formulation is notably simple: interaction between emotion and cause is modeled implicitly by question conditioning rather than by task-specific graph or tensor modules.

Prompted large language models introduced a different set of trade-offs. DECC used chain-of-thought prompting to decompose ECPE into recognizing, locating, analyzing, and summarizing, and combined inducing inference with logical pruning and in-context learning. On the rebalanced Chinese ECPE dataset, GPT3.5-DECC (4-shot) reported 61.38 F1 with an F1 drop of -12.25%, compared with -33.9% for UECA-Prompt and -39.8% for MTST-ECPE, supporting the claim that decomposed reasoning is more robust under reduced position bias [2401.17716].

MEKiT focused on a limitation specific to LLM-based ECPE: although large language models excel in text comprehension and generation, their performance on the ECPE task is often underperform smaller language model, with the main reason identified as the lack of auxiliary knowledge [2507.14887]. MEKiT injected internal emotional knowledge generated by COMET and refined with SBERT, as well as external causal knowledge filtered from large open-domain instruction datasets such as FLAN, into an instruction-tuning dataset called EmoCausBlend and fine-tuned the backbone with LoRA using standard next-token prediction loss [2507.14887]. On NTCIR-13 ECPE, the best prior tuned LLM baseline in the summary table, Gemma-2-9B-it*, reported \(P=61.05\), \(R=54.55\), \(F1=57.62\), whereas MEKiT reported \(P=65.04\), \(R=58.31\), \(F1=61.49\), an absolute F1 gain of 3.87% [2507.14887]. The reported applicability across Vicuna, LLaMA2/3, Qwen, and Gemma suggests that, for ECPE, knowledge injection and instruction design may be at least as important as raw model scale.

## 5. Granularity shifts, domain transfer, and dialog or conversation variants

One major line of extension concerns representational granularity. ECSP introduced Emotion-Cause Span-Pair extraction and classification, motivated by the observation that emotion expression and cause are not the whole clause in most cases and that clause-pair extraction therefore limits real-world applicability [2003.03507]. Its span-based ETC model reported F1 = 52.11 for emotion-cause span-pair extraction and F1 = 48.97 for ECSP, while the same framework, when restricted to clause-level, reported F1 = 86.05 on ECPE and F1 = 89.57 on ECE [2003.03507]. This suggests that clause-level ECPE and span-level ECSP are best viewed as related operating points on a granularity continuum.

A second line concerns domain transfer. “Emotion-Cause Pair Extraction in Customer Reviews” created a manually annotated English dataset of 1,000 reviews across 50 diverse products because there are no existing English ECPE datasets in the review domain [2112.03984]. The system combined Word2Vec with the NRC Emotion Lexicon to construct emotion-aware word embeddings, used a Bi-LSTM for emotion classification, another Bi-LSTM for cause clause identification conditioned on predicted emotion probabilities, and applied Agglomerative Clustering to summarize representative cause clauses [2112.03984]. No explicit F-score, precision, or recall numbers were provided, but the work established feasibility for review-specific English ECPE [2112.03984].

A third line shifts from monologic documents to dialogs and conversations. ECQED argued that conventional ECPE ignores emotion type and cause type and is limited to a single text piece, then extended the task to Emotion-Cause Quadruple Extraction in Dialogs with quadruples of the form \((\text{emotion\_utterance}, \text{cause\_utterance}, \text{emotion\_type}, \text{cause\_type})\) [2306.03969]. Its structural and semantic heterogeneous graph plus parallel grid tagging scheme reported F1 = 63.68% for quadruple extraction and F1 = 69.37% for pair extraction on RECCON, with a 9.21% F1 drop when the SSHG component was removed [2306.03969]. SCALE addressed ECPEC by semantic decoupling of emotion-oriented and cause-oriented representations followed by optimal transport–based global alignment, reporting F1 = 58.83 on RECCON-DD, 34.69 on RECCON-IE, and 57.70 on ECF [2604.19547]. Relative to document ECPE, these dialog formulations emphasize many-to-many conversational causality, speaker structure, and globally consistent matching.

## 6. Evaluation limits, explanatory validity, and current debates

A recent critique argues that the dominant binary pair/non-pair formulation captures only part of what “explanation” should mean. “They Are Not the Same: Direct Causes Are Not Grounded Emotion Explanations” examined binary ECPE under a three-role diagnostic audit with emo-cause, emo-context, and non-pair labels [2605.25208]. In IEMO-MECP, 90.9% of original positives remain emo-cause and 95.0% of original negatives remain non-pair, confirming that the binary ECPE task is largely preserved; however, emo-context appears on both sides of the original boundary, with 8.8% among original positives and 4.6% among original negatives, and its overall frequency rises from 5% to almost 10% among the most uncertain pairs [2605.25208].

The central claim is not that binary ECPE is useless, but that it is easy to over-read as evidence grounded emotion explanation. Across evaluated ECPE models, direct triggers are recovered more reliably than contextual support, and under shortcut pressure binary-trained models assign higher pair scores to nearby lexically similar non-pair candidates than to evidence supported but structurally harder emo-cause and emo-context pairs [2605.25208]. The paper therefore concludes that high binary ECPE performance indicates that a model can identify direct triggers; it does not indicate that the model has explained the emotion [2605.25208].

This debate reframes several earlier modeling trends. Position-bias mitigation, graph reasoning, multi-turn verification, and knowledge injection all improve pair extraction, but they do not by themselves resolve the distinction between direct trigger extraction and grounded explanation. A plausible implication is that future ECPE work will need to separate claims about direct-cause extraction from claims about explanatory adequacy, potentially combining role-wise evaluation with richer knowledge sources and dialog-level context modeling.

Source: https://www.emergentmind.com/topics/emotion-cause-pair-extraction-ecpe