Abstract
In this paper, we present a new challenging task for emotion analysis: emotion cause extraction (ECE). Rather than simple sentiment classification, we focus on detecting the stimulant or reason behind an emotion. We propose an emotion cause detection method using an event extraction framework, where events are represented via a tree structure. To address the inherent class imbalance in training data, we employ an under-sampling-based bagging algorithm with multi-kernel SVMs. Our approach achieves an F-measure 7.04% higher than state-of-the-art methods.
Introduction: The "Why" Behind Emotions
While traditional sentiment analysis identifies what an emotion is (e.g., happiness, anger), identifying the why—the emotion cause—is crucial for deeper understanding in fields like customer service and social science. Emotion Cause Extraction (ECE) is difficult due to two primary factors: data scarcity and extreme class imbalance, as most clauses in a text do not contain an emotion cause.
Methodology
The proposed framework treats ECE as an event extraction task, moving beyond keyword matching to structural analysis.
Event Tree Representation
Sentences are first parsed into 7-tuple events (Subject, Object, Adverb, Verb, etc.). These are then converted into Event Trees (ET) that capture the hierarchical grammatical structure of the clause. This structural representation allows the model to differentiate between similar word sets used in different syntactic contexts.
Multi-Kernel SVM and Ensemble Learning
To process these trees, we utilize a Support Vector Machine (SVM) with a multiplicative multi-kernel. This kernel combines a Tree Kernel (for syntactic structure) and a Polynomial Kernel (for lexical features). The multiplicative approach captures the interaction between vocabulary and grammar more effectively than simple addition.
To combat the "needle in a haystack" problem where neutral clauses far outnumber cause clauses, we implement an under-sampling-based bagging algorithm.
As shown in the framework above, the system repeatedly pairs the limited positive samples with different random subsets of negative samples. Multiple classifiers are trained on these balanced subsets, and their outputs are combined in an ensemble to produce the final result.
Dataset and Observations
Due to the lack of existing resources, we developed a new dataset from 20,000 Chinese news articles, resulting in 2,105 emotional instances. A key observation from this data was that 78% of causes are adjacent to the emotion keyword, and over 93% of causes are centered around verbs, justifying the event-based approach.
Results
The multiplicative multi-kernel bagging approach ($B_{*multi}$) achieved a benchmark F-measure of 0.7456, representing a 7.04% improvement over previous methods. The ensemble strategy proved particularly effective at maintaining high recall without sacrificing precision, demonstrating robustness even with limited training data.
Conclusion
By representing emotion causes as structured events and utilizing an ensemble of multi-kernel SVMs, we significantly improved the state-of-the-art in ECE. Future work will focus on identifying implicit (noun-based) causes and extending the framework to handle long-distance relationships between emotions and their triggers.
Summary updated