Source-linked AI summary
Emotion-Cause Pair Extraction: A New Task to Emotion Analysis in Texts
Rui Xia, Zixiang Ding
TL;DR
ECE requires emotion annotations before cause extraction and overlooks the mutual indication between emotions and causes. The paper introduces ECPE and a two-step multi-task framework for extracting and filtering emotion-cause pairs, achieving 61.28% F1 on a benchmark corpus and demonstrating feasibility and effectiveness.
Problem
Traditional ECE requires emotion annotations before cause extraction and ignores that emotions and causes are mutually indicative.
Method
ECPE first extracts emotion and cause clauses through multi-task learning, then forms candidate pairs and filters out pairs without causal relationships.
Results
61.28% F1 was achieved for emotion-cause pair extraction on a benchmark emotion cause dataset without using test-set emotion annotations; cause extraction was comparable to traditional ECE methods.
Takeaways & Limitations
The experiments support the feasibility of ECPE and the effectiveness of the proposed approach.
Takeaways & Limitations
The two-step strategy is indirect, and errors in its first step affect the second; future work targets end-to-end extraction.
Abstract
from arXiv · showhide
Emotion cause extraction (ECE), the task aimed at extracting the potential causes behind certain emotions in text, has gained much attention in recent years due to its wide applications. However, it suffers from two shortcomings: 1) the emotion must be annotated before cause extraction in ECE, which greatly limits its applications in real-world scenarios; 2) the way to first annotate emotion and then extract the cause ignores the fact that they are mutually indicative. In this work, we propose a new task: emotion-cause pair extraction (ECPE), which aims to extract the potential pairs of emotions and corresponding causes in a document. We propose a 2-step approach to address this new ECPE task, which first performs individual emotion extraction and cause extraction via multi-task learning, and then conduct emotion-cause pairing and filtering. The experimental results on a benchmark emotion cause corpus prove the feasibility of the ECPE task as well as the effectiveness of our approach.
1 Introduction
Traditional ECE requires emotion annotations before extracting causes and does not exploit the mutual indication between emotions and causes. ECPE instead extracts emotion-cause pairs without advance emotion annotations, using a two-step framework evaluated on a benchmark corpus.
- Traditional ECE detects cause clauses given annotated emotions, limiting applications when emotion annotations are unavailable at test time.
- ECPE extracts all potential pairs of emotion clauses and corresponding cause clauses directly from a document without providing emotion annotations in advance.
- The proposed framework first extracts emotion and cause clauses through multi-task learning, then pairs candidates and filters out non-causal pairs.
- 61.28% F1 was achieved for emotion-cause pair extraction on a benchmark emotion cause dataset without using test-set emotion annotations.
- Without test-set emotion annotations, cause extraction was comparable to traditional ECE methods, while showing advantages over methods designed to remove annotation dependence.
2 Related Work
Earlier ECE research extracted word- or clause-level causes given annotated emotions. This work instead extracts emotions and corresponding causes together, reflecting their mutual dependence.
- ECE research progressed from word-level sequence labeling to clause-level cause detection, with clause-level modeling addressing multi-clause and long-distance causes.
- Benchmark Chinese emotion-cause corpora supported rule-based, traditional machine-learning, and later studies of ECE.
- Prior work generally extracted causes conditional on emotion annotations, whereas this work extracts emotions and corresponding causes simultaneously.
- The proposed direction investigates whether indicating causes can improve emotion extraction and vice versa because causes and emotions are not mutually independent.
3 Task
ECPE extracts emotion-cause clause pairs from a document, unlike traditional ECE, which extracts a cause clause after the emotion clause is annotated. The task is defined at clause level.
- Given a document of clauses, ECPE extracts a set of pairs consisting of an emotion clause and its corresponding cause clause.
- Traditional ECE extracts the cause clause given the annotation of the emotion clause.
- ECPE is more difficult because the emotion clause annotation is not provided before extraction.
- Both tasks are defined at the clause level because emotion causes are difficult to describe at the word or phrase level.
4 Approach
The approach extracts emotion and cause clauses through multi-task learning, then forms candidate pairs and filters them for causal relationships. It includes independent and interactive variants to model the correlation between the two extraction tasks.
- Step 2: Emotion-Cause Pairing and Filtering: Step 2 applies a Cartesian product to the extracted emotion and cause sets, creating candidate pairs for filtering.Each pair uses emotion-clause and cause-clause representations plus their clause-distance features; logistic regression removes pairs without causal relationships.
- Step 1: Individual Emotion and Cause Extraction: Step 1 uses two multi-task learning networks to extract sets of emotion clauses and cause clauses for each document.The networks are Independent Multi-task Learning and Interactive Multi-task Learning.
- Independent Multi-task Learning: The document is represented hierarchically with word-level Bi-LSTMs for clause context and attention-based clause representations.A clause-level Bi-LSTM layer then processes the clause representations for the extraction tasks.
- Independent Multi-task Learning: Independent multi-task learning predicts emotion and cause labels with separate clause-level Bi-LSTM components and a weighted sum of their prediction losses.The loss combines emotion and cause cross-entropy errors using a tradeoff parameter λ.
- Interactive Multi-task Learning: Interactive multi-task learning captures emotion-cause correlation by allowing one extraction component to enhance the other.Inter-EC uses emotion extraction to improve cause extraction, while Inter-CE uses cause extraction to enhance emotion extraction.
5 Experiments
Experiments evaluate the constructed ECPE corpus, the multi-task models, interaction between emotion and cause extraction, pair filtering, and comparison with traditional ECE methods.
- Dataset and Metrics: The ECPE corpus merges documents with identical text content and labels each emotion-cause pair; experiments use 90% training data, 10% testing data, and averages over 20 repetitions.Precision, recall, and F1 are used for ECPE and the two individual extraction tasks.
- Overall Performance: Inter-EC improves ECPE and both sub-tasks over Indep, mainly by increasing recall for cause extraction and consequently ECPE recall.The results also report improved emotion extraction, indicating benefit from supervision between the two tasks.
- Overall Performance: Inter-CE improves ECPE relative to Indep mainly through higher emotion-extraction precision, which also improves ECPE precision.This supports using cause-extraction predictions to enhance emotion extraction.
- Upper-Bound of Emotion and Cause Interaction: The interaction results support mutual indication between emotion and cause, while upperbound experiments show substantially greater gains when ground-truth labels replace predicted inputs.The comparison attributes the gap to errors in predicted emotion extraction and notes that cause extraction has more room for improvement.
- Evaluation on the ECE Task: Without emotion annotations on test data, the method achieves 65.07% F1 and outperforms CANN-E by 27.1%, while remaining comparable to most traditional ECE methods.Removing emotion annotations causes CANN's F1 score to drop by about 34.69%.
6 Conclusions and Future Work
The paper introduces ECPE and a two-step method that extracts emotions and causes before pairing and filtering them. It reports effectiveness on a benchmark ECE corpus but identifies error propagation and indirect optimization as limitations.
- ECPE extracts potential emotion-cause pairs without requiring emotion annotations in advance.
- The method first extracts emotions and causes with multi-task learning, then combines them through Cartesian product and filters false pairs.
- A benchmark ECE corpus was adapted to construct a corpus suitable for ECPE.
- The two-step strategy is indirect, and first-step mistakes affect second-step results.
- Future work will pursue a one-step end-to-end model that directly extracts emotion-cause pairs.