Abstract
Emotion detection in social media is complicated by multi-label expressions, topical context, and noisy annotations. We present a constrained optimization framework based on Non-negative Matrix Factorization (NMF) that integrates topical correlations, emotion co-occurrence "bindings," and word-usage biases. Our model significantly outperforms state-of-the-art methods on SemEval, ISEAR, and Twitter datasets, demonstrating exceptional robustness to label noise and the ability to automatically expand emotional lexicons.
Introduction
Detecting emotions in social media is notoriously difficult because human expression is messy. Traditional models often treat every post as an isolated island, but the authors identify four primary gaps in existing approaches:
- Multiple Emotions: Most models pick one emotion, but a single sentence often contains several (e.g., "Recent plane crashes make me sad and angry").
- Topical Correlation: Posts discussing the same event likely share an emotional baseline.
- Emotion Bindings: Emotions follow natural co-occurrence rules; joy and fear rarely coexist, while surprise and joy often do.
- Noisy Labels: Ground truth for emotion is subjective, leading to inherent noise in training data.
Methods
The authors propose a framework based on Non-negative Matrix Factorization (NMF), modified with a series of novel constraints. Instead of standard classification, they model the problem as the minimization of a complex objective function. Key components include:
- Topic Constraints ($\lambda_w$): A similarity matrix ensures topically related documents exhibit similar emotion distributions.
- Emotion Binding Matrix ($Q$): A constraint that penalizes the co-occurrence of unlikely emotion pairs based on prior linguistic knowledge.
- Bias Factors ($u, v$): Multivariate half-normal distributions account for "popular" vs. "rare" words, preventing common neutral words from skewing results.
- Lexicon Constraints ($\lambda_d$): Leverages existing word-emotion lexicons while allowing the model to learn the emotional weight of new words (e.g., associating "bestie" with joy).
Experimental Evaluation
The model was evaluated on three distinct datasets: SemEval (1,250 multi-label news headlines), ISEAR (7,666 sentences from a psychological study), and a Twitter Dataset (1,800 informal tweets labeled by Mechanical Turkers).
Key Results and Discussion
The proposed model consistently outperformed baselines, including SVMs and standard NMF models. On the SemEval and ISEAR datasets, the model achieved average F-scores of 0.63 and 0.74, respectively. On the messy Twitter dataset, the model improved recall by 42% over standard NMF while maintaining higher precision.
Component Importance
Ablation studies (Figure 6) reveal that the Topic Constraint provided the largest boost to performance. Knowing the subject matter helps the model "fill in the blanks" for sentences that might not contain explicit emotion words.
Structural Robustness
Because the model relies on structural constraints rather than just local labels, it is incredibly resilient to bad data. The authors found the model's performance stayed high even when 40% of the training labels were intentionally corrupted.
As shown in Figure 3, performance on the ISEAR dataset remains stable until noise levels exceed 40%, confirming the model's structural robustness compared to traditional methods.
Conclusion
This constrained optimization approach provides a robust, efficient, and context-aware solution for social media emotion analysis. While the model still struggles with high-level nuance like sarcasm, its ability to leverage topical context and handle noisy data represents a significant advancement. Future work will focus on refining bias factors to handle platform-specific quirks like emojis and hashtags.
Summary updated