Joint multi-level attentional model for emotion detection and emotion-cause pair extraction
Introduction
Emotion cause extraction (ECE) aims at identifying the corresponding clauses that induce a certain emotion expression in current context, and become a popular research topic in recent years [1], [2]. This task is first treated as a word-level sequence labeling problem [3]. Considering the limitation of word-level labeling, which is not appropriate for the document-level emotion analysis, Gui et al. (2016) re-designed the ECE task for cause clause detection, and modeled the task as a clause-level classification problem [4]. Furthermore, Xia et al. (2019) thought that there exists an implicit connection between the cause clauses and their corresponding emotion clauses in the document-level representation [5]. They expanded the original task into a new one called emotion-cause pair extraction (ECPE), which aims to predict emotion-cause pairs, and meanwhile introduced a new dataset for the task.
Emotion detection (ED) task [6] aims to predict the clause-level emotion category. As shown in Fig. 1, we can find that ED is closely related to ECPE in two aspects: (1) All emotion clauses in ECPE must carry certain kinds of emotions, which means that these clauses can not be non-emotion clauses in ED; (2) Cause clause is also related to the given emotion, and different semantic fragments can result in different emotions. For example, ‘became better’ usually indicates the positive emotions such as ‘happiness’. The information of the specific emotion, emotion clause and cause clause can affect each other, so we propose a joint task (called ED+ECPE) to improve the performance of both ECPE and ED tasks.
As shown in Fig. 1, we can see that emotion clause and cause clause are implicitly connected in a fine-grained manner, and they both contribute to the formation of emotions. For example, ‘cried excitedly’ reflects the emotion ‘happiness’, and ‘became better’ actually induces the action ‘cried excitedly’. To get a better emotion-cause pair representation and take advantage of the emotion formation process, a coherent method is required to model the inner relationship between emotion clause and cause clause with the guidance of the emotion information. To this end, we can use self attention to describe the connection between the specific emotion and emotion clause, and mutual inner attention to describe the connection between emotion clause and cause clause. Furthermore, these two attentional forms can be integrated coherently by a multi-level attentional module.
In this paper, we devise a hierarchical structure enhanced by a biaffine module to address the joint task of ED+ECPE. This structure consists of a contextual encoder and multiple clause encoders, which are implemented by long short-term memory networks (LSTMs). The outputs of the contextual encoder are further fed into a biaffine module to extract emotion-cause pairs. To capture the fine-grained emotion-sensitive relationship between emotion clause and cause clause, we introduce a multi-level attentional module, which requires the information of self attention and the mutual inner attention, for the hierarchical structure. Specifically, self attention is generated in clause encoders for ED and reflects the word-level importance to formulate the emotion in the emotion clause. Mutual inner attention is gathered from the alignment matrix between two clauses in an emotion-cause pair, and its value represents the connection strength. Then a multi-level attentional module utilizes these attention values to generate pair representations. Experimental results show our proposed architecture can bring consistent performance gains in comparison with several existing methods, achieving the state-of-the-art performance on two benchmark datasets.
Section snippets
Related work
In recent years, emotion cause extraction (ECE) has received extensive research attention in NLP community. Existing work can mainly be divided into three categories: word-level emotion cause extraction, clause-level emotion cause extraction and emotion-cause pair extraction.
Word-level emotion cause extraction: In early work, emotion cause extraction is defined as extracting the word-level causes that lead to the given emotions in the text [3]. Traditional methods such as machine learning and
Preliminaries
In this section, we first introduce the basic procedure and original hierarchical structure for handling ECE task, then we give a brief illumination of our ED+ECPE task.
Model
The overall structure of the proposed model is illustrated in Fig. 2. Specifically, we enhance the encoder with BERT and propose a joint model for ED+ECPE task. With the guidance of self attention mechanism, a multi-level attentional module is designed for modeling the word-level interaction between emotion clauses and cause clauses, to obtain a fused pair representation. Note that all possible emotion-cause pairs are evaluated by the score function in the biaffine module.
Datasets
We conduct our experiments on two benchmark datasets: the ECE corpus and ECPE corpus. The ECE corpus is the mostly used corpus for emotion cause extraction [4] and the ECPE corpus is a newly released dataset for emotion-cause pair extraction [5]. The two datasets are split by the same proportion of 8:1:1. For the ECE, there are 21,050 annotated documents, including 16,840 documents for training, 2110 for development, and 2100 for testing. Correspondingly, there are 24,354 clauses for training,
Conclusion
We investigated a joint structure for the emotion detection and emotion-cause pair extraction tasks by constructing a multi-task learning problem. We introduce a multi-level attentional module, which requires the information of self attention and the mutual inner attention, to capture the relationship between the emotion clause and cause clause. To utilize the hierarchical structure information, latent variable was introduced to model the emotion distribution over the clauses in a document.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
We wish to draw the attention of the Editor to the following facts which may be considered as potential conflicts of interest and to significant financial contributions to this work. We confirm that the manuscript has been read and approved by all named authors and that there are no other persons who satisfied the criteria for
CRediT authorship contribution statement
Hao Tang: . Donghong Ji: Writing - review & editing, Investigation. Qiji Zhou: .
Acknowledgments
We thank the reviewers for their valuable comments. This work is supported by the National Natural Science Foundation of China (Grant No.61702121, 61772378).
Hao Tang is currently pursuing Ph.D. at Wuhan University, China. His current research interests are machine learning and natural language process.
References (37)
- et al.
Comparison of the support vector machine and relevant vector machine in regression and classification problems
Proceedings of the 8th International Conference on Control, Automation, Robotics and Vision, ICARCV Kunming, China
(2004) - et al.
Emocause: An easy-adaptable approach to extract emotion cause contexts
- et al.
A co-attention neural network model for emotion cause analysis with emotional context awareness
- et al.
From independent prediction to reordered prediction: Integrating relative position and global label information to emotion cause identification
Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019.
(2019) - et al.
Emotion cause detection with linguistic constructions
- et al.
Event-driven emotion cause extraction with corpus construction
- et al.
Emotion-cause pair extraction: A new task to emotion analysis in texts
- et al.
Neural feature extraction for contextual emotion detection
- et al.
Detecting emotion causes with a linguistic rule-based approach
Comput. Intell.
(2013) - et al.
Text-based emotion classification using emotion cause extraction
Expert Syst. Appl.
(2014)
Emotion cause detection for chinese micro-blogs based on ECOCC model
A rule-based approach to emotion cause detection for chinese micro-blogs
Expert Syst. Appl.
Emotion cause detection with linguistic construction in chinese weibo text
Conditional random fields: Probabilistic models for segmenting and labeling sequence data
Extracting causes of emotions from text
Proceedings of the Sixth International Joint Conference on Natural Language Processing, IJCNLP 2013, Nagoya, Japan
Detecting emotion stimuli in emotion-bearing sentences
Detecting concept-level emotion cause in microblogging
A bootstrap method for automatic rule acquisition on emotion cause extraction
Cited by (49)
Experiencer-Driven and Knowledge-Aware Graph Model for emotion–cause pair extraction
2023, Knowledge-Based SystemsEmotion-cause pair extraction with bidirectional multi-label sequence tagging
2023, Applied Intelligence
Hao Tang is currently pursuing Ph.D. at Wuhan University, China. His current research interests are machine learning and natural language process.
DongHong Ji is Professor of School of Cyber Science and Engineering at Wuhan University. His research interests are machine learning, logics and reasoning, natural language processing and their applications in text analysis. He is coinvestigator of several other externally funded projects.
QiJi Zhou is currently pursuing Ph.D. at Wuhan University, China. His current research interests are semantic parsing and natural language process.