Introduction
Traditional sentiment analysis often simplifies human emotion into discrete categories like "happy" or "sad." However, psychological research suggests that emotions are better represented as a continuous spectrum across multiple dimensions. The most common framework uses Valence (the pleasantness of a stimulus), Arousal (the intensity of emotion), and Dominance (the degree of control).
Emotion regression aims to map text to continuous scores within this VAD space. This task is challenging because these dimensions are often correlated yet distinct. For instance, certain words might influence Arousal without affecting Valence. Most existing models treat these dimensions as independent, failing to capture the nuanced overlaps in emotional language.
The Adversarial Attention Network (AAN)
To address these challenges, researchers proposed the Adversarial Attention Network (AAN). The core idea is to move beyond isolated regression tasks and use a joint learning approach that explicitly separates shared emotional features from dimension-specific ones.
The AAN architecture consists of three primary components:
- Multi-layered Attention: Instead of a single attention mechanism, the model employs three: one for a specific dimension (e.g., Valence), one for another (e.g., Arousal), and a Shared Attention layer ($Att_S$) designed to capture words that contribute to both.
- Feature Extraction: A Bidirectional LSTM (BiLSTM) processes the input, and the attention weights are applied to the hidden states to produce fixed-length feature vectors.
- Adversarial Training: This is the model's "secret sauce." A Discriminator ($D$) is trained to identify which emotion dimension a feature vector belongs to. Simultaneously, the attention layers are trained to "fool" the discriminator. This adversarial "tug-of-war" forces the specific attention layers to focus on highly discriminative words unique to their respective dimensions, while the shared layer captures common emotional cues.
Experimental Setup
The model was evaluated using the EMOBANK corpus, a comprehensive dataset containing over 10,000 text entries. EMOBANK is unique because it provides annotations for two perspectives:
- Reader's Emotion: How the text makes the audience feel.
- Writer's Emotion: The intended sentiment of the author.
The dataset spans six diverse domains, including News, Fictions, Blogs, Essays, Letters, and Travel Guides, with each entry rated on a 1.0 to 5.0 scale for Valence, Arousal, and Dominance.
Results and Analysis
The AAN was benchmarked against several state-of-the-art models, including Deep CNNs and Regional CNN-LSTMs. Performance was measured using Pearson’s correlation coefficient ($r$).
As shown in the results table, the AAN consistently outperformed all baselines across every domain and for both Reader and Writer perspectives. For example, in the News domain for Reader's Valence, AAN achieved an $r$-value of 0.424, a significant improvement over the best baseline of 0.392.
Ablation studies further confirmed that the adversarial component is crucial; a version of the model using joint learning without the discriminator performed significantly worse. The results also highlighted domain-specific nuances. In the News domain, writers often maintain low Arousal for objectivity, even when the content evokes high Arousal in readers—a distinction the AAN captured effectively.
Conclusion
The Adversarial Attention Network demonstrates that attention mechanisms can serve as more than just a focus tool; they can be a medium for adversarial learning to disentangle complex, overlapping features. By training emotion dimensions jointly and using a discriminator to enforce specificity, the AAN sets a new standard for multi-dimensional emotion regression. Future work aims to extend this framework to handle more than two dimensions simultaneously and apply it to other heterogeneous NLP tasks like multi-domain sentiment analysis.
Summary updated