TL;DR
This paper introduces a novel framework for learning distributed representations of emotion categories in an "emotion space," moving beyond traditional one-hot encodings. By leveraging soft labels from pre-trained models, they demonstrate that these representations capture nuanced emotion relations more effectively than semantic word embeddings, leading to better performance in tasks like emotion mapping.
What Problem Does This Paper Address?
Current approaches to text emotion analysis often treat emotion categories as discrete and independent, typically representing them with one-hot vectors. This simplification overlooks the inherent relationships and complexities between different emotions. For instance, while "joy" and "sadness" are distinct, they exist on a spectrum and can influence each other. Existing methods fail to capture these inter-emotion dynamics, limiting their ability to understand the full spectrum of human emotional expression in text. This paper aims to bridge this gap by developing a method to represent emotion categories in a continuous "emotion space" where their relationships are explicitly modeled.
Methods
The core of the paper is a general framework for learning distributed emotion representations. The key idea is to treat each emotion category as a point in an emotion space, and the representation of an emotion category ($V_K$) is derived by averaging the representations of all instances labeled with that category.
- Soft Labels: Instead of using hard, one-hot labels, the authors utilize "soft labels" generated by a pre-trained neural network model. These soft labels, which represent probability distributions over emotion categories, are considered more informative and less prone to overfitting than manual annotations.
- Instance Representation: The distributed representation of an individual text instance ($V(x)$) is directly taken as the soft label output by the neural network.
- Category Representation: The distributed representation for an emotion category $K$ ($V_K$) is computed as the average of the soft label representations of all instances belonging to category $K$. For multi-label data, a weighted average is used, where weights are the reciprocal of the number of labels for each instance.
- Distance Metric: Squared Euclidean distance is used as the metric to measure the difference between instance and category representations.
Datasets
The paper uses several datasets for training and evaluation:
- GoEmotions: A large dataset of 58k Reddit comments with 27 emotion categories, multi-labeled. This dataset is used for the arrangement and mapping experiments.
- AffectiveText: A dataset of 1250 news headlines, multi-labeled with six emotion categories and valence.
- ISEAR: A dataset with single labels for seven emotion categories.
- Affect in Tweets: A dataset of tweets with ten emotion categories.
The paper notes that most existing datasets are annotated with discrete basic emotion categories, which inherently ignore inter-emotion relations.
Key Insights and Ideas
- Emotion Space vs. Semantic Space: The paper clearly distinguishes between "semantic space" (where word embeddings reside) and "emotion space" (where emotion categories are represented). Semantic space captures word meaning, while emotion space captures the relationships and nuances between emotional states.
- Soft Labels as Richer Representations: The use of soft labels from neural networks is crucial. These probabilities provide a more nuanced view of an instance's emotional content than a single hard label, allowing for the derivation of continuous emotion representations.
- Linear Separability in Emotion Space: A key finding is that in the learned emotion space, different categories (positive, negative, ambiguous) become linearly separable, unlike in semantic space where they are mixed. This suggests that the learned representations capture a more structured understanding of emotions.
- Ambiguous Emotions as Intermediaries: Ambiguous emotions are found to lie between positive and negative emotions in the emotion space, reflecting their transitional nature.
Key Results
- Improved Emotion Relation Expression: Experiments show that the proposed distributed representations in emotion space significantly outperform word vectors in semantic space for expressing emotion relations.
- Mapping Accuracy: In a mapping experiment, the proposed representations correctly mapped 18 out of 21 emotion categories from the GoEmotions taxonomy to Ekman's basic emotions, a substantial improvement over semantic space embeddings (which mapped 3-10 correctly).
- Cross-Corpora Consistency: The emotion relations derived from different datasets (AffectiveText, GoEmotions, ISEAR, Affect in Tweets) showed high Pearson correlation coefficients (0.867-0.949), indicating that the learned emotion representations are consistent across diverse domains and annotation standards.
- Model Agnosticism: The effectiveness of the emotion space representations was demonstrated across different neural network architectures (TextCNN, BiLSTM, BERT), suggesting the approach is robust to the underlying model.
Limitations and Future Work
- Corpus-Specific Representations: The current distributed representations are derived from specific emotion classification datasets. The authors suggest exploring methods to build universal emotion representations that are independent of any single corpus.
- Reliance on Soft Labels: The method relies on soft labels predicted by a neural network. Future work could investigate more general methods for obtaining these representations.
- NLP Applications: The authors plan to explore further NLP applications of their emotion representations, such as improving emotion classification performance and studying emotion spaces across different languages.
- Complexity of Emotion Relations: The paper acknowledges that some emotion relations, like embarrassment, are complex and can map to multiple emotions (e.g., sadness and disgust), highlighting the ongoing challenge in fully capturing human emotional nuance.
Summary updated