Abstract
Generating appropriate emotions is essential for dialog systems to provide human-like interaction. While previous systems focused on learning empathetic manners from anonymous data, they often suffer from affective instability—inconsistent emotional responses that decrease user engagement. This research proposes a new task, Personality-affected Emotion Generation, which generates emotions based on a system's assigned personality. We investigate this through a personality-affected mood transition model and introduce the Personality EmotionLines Dataset (PELD), which contains both emotion and personality annotations. Our results show that modeling personality as a transition weight improves emotion generation performance by 13% in macro-F1 and 5% in weighted-F1 over BERT-base models.
The Problem: Affective Instability
While current AI chatbots have become proficient at recognizing user emotions, they often lack a stable internal identity. Most models mimic conversational patterns from aggregated data, leading to a phenomenon known as affective instability. A chatbot might oscillate between cheerful and pessimistic responses in a way that feels jarring or "robotic," even if the user's attitude remains constant. Psychological research suggests that human emotional expression is rooted in personality traits; therefore, a truly human-like system must bridge the gap between reacting to the user (empathy) and acting according to an internal trait (personality).
Methodology: Personality-affected Mood Transition
The core innovation of this work is treating personality not as a static feature, but as a transition weight that dictates how a speaker’s mood evolves throughout a conversation.
1. The VAD Bridge
To integrate personality and emotion into a unified framework, the model maps all factors into the Valence-Arousal-Dominance (VAD) space. This 3D coordinate system allows the model to mathematically relate stable personality traits (the "Big Five" OCEAN model) to transient emotional states.
2. Mood Transition Regression
Instead of jumping directly to an emotion label, the model predicts a "mood state" (a region in VAD space). It calculates how the current dialog context shifts the speaker's mood, using the speaker's personality to weight that shift. This ensures that the AI maintains a "memory" of its internal state, preventing sudden emotional flips.
3. Multi-Granularity Extraction
The architecture utilizes a fine-tuned BERT encoder combined with a specialized attention layer. This layer aligns three distinct levels of information: the semantic meaning of the text, word-level emotional embeddings, and utterance-level emotion labels.
4. Joint Training
The system is trained to solve two tasks simultaneously: predicting the next mood state (regression) and selecting the correct discrete emotion label (classification). To handle the prevalence of "Neutral" responses in datasets, the authors employ Focal Loss to prioritize the learning of rarer emotions like "Disgust" or "Fear."
The PELD Dataset
To evaluate the framework, the authors introduced the Personality EmotionLines Dataset (PELD).
- Source: Built from the scripts of the TV show Friends, merging data from the MELD and EmoryNLP datasets.
- Content: It contains 6,510 dialogue triples (Utterance 1, Utterance 2, and the Response) for the six main characters.
- Annotations: Each triple includes emotion labels for every utterance and "Big Five" personality scores, providing a unique resource for studying how specific personas (e.g., Chandler vs. Rachel) react differently to identical situations.
Key Insights and Results
Extensive experiments on the PELD dataset yielded several significant insights:
- Performance Boost: The proposed model improved emotion generation performance by 13% in macro-F1 and 5% in weighted-F1 over the standard BERT-base model.
- Minority Emotion Accuracy: The model was significantly better at generating rare emotions. For instance, it improved the F1 score for "Disgust" from 0.012 to 0.167, a category usually ignored by models over-optimized for neutral data.
- Personality as a Filter: In dialogue, many emotions can be contextually appropriate. Personality acts as a search constraint that narrows down these choices, making the AI's responses feel more deliberate and "in character."
- Negative Emotion Sensitivity: A notable finding was that personality traits have a much stronger influence on how people transition out of negative emotions (like sadness or anger) than how they handle positive ones.
Limitations and Future Work
While the model represents a significant step forward, it currently relies entirely on text. In real human interaction, personality is also expressed through voice, gestures, and facial expressions. Furthermore, the model occasionally struggles when a response requires deep "lore" or relationship history beyond the immediate 3-turn dialog window. Future work will focus on integrating these findings into a full Natural Language Generation (NLG) pipeline to ensure that the actual words spoken perfectly match the predicted emotion and personality.
Summary updated