TL;DR
MoodSwipe is a smart soft keyboard that detects the emotional tone of your messages in real-time and suggests alternative phrasings based on a target emotion you select. By turning emotion labeling into a useful feature for the user, the system creates a virtuous cycle for collecting high-quality, self-reported emotional data for dialogue research.
What Problem Does This Paper Address?
While emotional intelligence is vital for effective leadership and healthy relationships, text-based communication often strips away the nuance of tone and intent. Most existing affective computing tools focus on helping a receiver understand a message, but few help the sender express themselves accurately.
Furthermore, the researchers highlight a major bottleneck in NLP: collecting labeled emotion data for dialogues is notoriously difficult. Users find it intrusive to stop and label their feelings mid-conversation. MoodSwipe addresses both by providing a tool that helps users rephrase messages (e.g., turning a heated argument into a neutral description) while silently collecting user-validated emotion labels in the background.
Methods
The MoodSwipe system consists of two main components:
- Emotion Classification: The keyboard uses real-time classifiers to detect the emotion of the current input. The authors compared a Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) architecture. These models categorize text into seven emotions: Anger, Joy, Sadness, Fear, Anticipation, Tired, and Neutral.
- Text Suggestion: The system uses a retrieval-based model powered by Lucene. When a user wants to change the "mood" of their text, the system searches a massive database of social dialogues for similar messages that carry the target emotion. It ranks these candidates using the BM25 algorithm.
The user interface is the key innovation here. As shown in the workflow below, a color bar above the keyboard indicates the detected emotion. Users can swipe this bar to cycle through different emotional "filters" and see suggested text replacements.
Datasets
- EmotionPush Dataset: The primary source, containing 162,031 message logs collected from a previous emotion-based chat app.
- Manual Labels: 8,818 messages were manually labeled by native speakers to train and validate the classifiers.
- Comparison Corpora: The researchers also utilized LJ40K (blog posts) and a custom Tweet dataset (filtered by hashtags) to test how well models trained on other domains generalize to chat data.
Key Insights and Ideas
- Domain Specificity is Crucial: Models trained on blogs or tweets performed significantly worse on dialogue data than models trained on actual chat logs. Chat language is unique—it's informal, fragmented, and context-dependent.
- Incentivized Labeling: The "Select" and "Swipe" actions provide a clever way to gather ground-truth data. If a user swipes to "Joy" and selects a suggestion, the system can infer that the user's original intent was Joy, or that the selected text is a valid representation of that emotion.
- Color-Emotion Mapping: The system utilizes a specific color palette derived from psychological studies to represent emotions, making the interface intuitive (e.g., Red for Anger, Navy Blue for Sadness).
Key Results
- High Accuracy: The CNN and LSTM models trained specifically on dialogue data achieved accuracies over 0.90 for the three major emotions (Joy, Anger, and Sadness).
- Human Evaluation: In Amazon Mechanical Turk trials, workers ranked MoodSwipe's suggestions against the original user input. While original inputs were still preferred, MoodSwipe's suggestions were considered "good" (better than the original) 26% to 28% of the time.
- Anger Management: The system was particularly effective at suggesting "Anger" messages, with a "Good Suggestion Rate" of approximately 40%, suggesting the tool is highly useful for users looking to express (or perhaps moderate) frustration.
Limitations and Future Work
The authors acknowledge that the current retrieval-based model is a "filter" that depends entirely on the size and variety of the underlying database. If the database doesn't contain a similar message, the suggestions may feel irrelevant.
Future work aims to move beyond simple retrieval toward generative models that can synthesize new text tailored to a specific emotion. Additionally, the researchers plan to leverage the automatically collected user labels to further refine the classifiers, creating a self-improving system.
Summary updated