Introduction
Traditional Emotion Knowledge Graphs (EKGs) typically focus on static representations, linking emotion keywords to related terms or assigning fixed sentiment scores like valence and arousal. However, human emotions are inherently dynamic and fluid, evolving throughout a conversation based on context and external stimuli. Existing models often fail to capture these "fuzzy boundaries" and the specific triggers that cause an emotional shift.
To address this gap, this research introduces EmoTransKG, a novel framework that formalizes emotion transitions as a graph problem. By shifting the focus from static labels to the process of transformation, the authors provide a structured way to represent how dialogue acts as a catalyst for emotional change.
The EmoTransKG Framework
EmoTransKG redefines the standard Knowledge Graph triplet $(h, r, t)$ to suit conversational dynamics. Unlike traditional KGs with a closed set of relations (e.g., "is-a"), EmoTransKG utilizes an open class of relations derived directly from dialogue.
- Head Entity ($h$): The speaker's initial emotion state.
- Tail Entity ($t$): The speaker's subsequent emotion state.
- Relation ($r$): The sequence of open-text utterances or events that occur between these two states.
This structure allows for "multi-edge" logic, where various conversational paths can lead from the same starting emotion to the same ending emotion, capturing the complexity of human interaction.
EmoTransNet and Emotion Retrofitting
To navigate this graph, the authors developed EmoTransNet, a Transformer-based translational model. EmoTransNet uses a RoBERTa backbone to extract features from textual utterances and a Transformer encoder to capture the temporal sequence of the "relation" events. The model projects the head emotion into a high-dimensional space, concatenating it with the relation representation to predict the resulting tail emotion.
A key innovation of this work is Emotion Retrofitting. EmoTransNet is designed as a model-agnostic "plug-in" for existing Conversational Emotion Recognition (CER) systems. Instead of replacing baseline models, it refines their predictions by incorporating knowledge-based transformation logic.
Experimental Evaluation
The system was evaluated on two major conversational benchmarks: IEMOCAP (dyadic scripted/improvised performances) and EmoryNLP (transcripts from the TV show Friends). These datasets offer contrasting dynamics; IEMOCAP exhibits high "emotional inertia" (70% of transitions result in the same emotion), while EmoryNLP features rapid transformations (only 36% inertia).
Key Results
EmoTransNet was tested against 10 different baseline CER models, including DialogueRNN, COSMIC, and DAG-ERC. The results demonstrated:
- Consistent Performance Gains: Incorporating EmoTransKG led to statistically significant improvements (p < 0.05) across all 10 models on both datasets.
- Quantitative Success: For instance, the DialogueRNN model's weighted F1 score on IEMOCAP improved from 66.03 to 67.19.
- Qualitative Insights: t-SNE visualizations confirmed that EmoTransNet successfully clusters similar emotional transformations, proving it captures the underlying semantic logic of how dialogue influences emotional states.
Conclusion and Future Work
EmoTransKG represents a significant step toward modeling the fluidity of human experience in AI. While the current model faces challenges such as data imbalance (a bias toward "no change" transitions) and the subjectivity of basic emotion definitions, it provides a robust foundation for knowledge-based emotion recognition.
Future research will focus on expanding linguistic coverage using larger datasets like DailyDialog and MELD. Additionally, the authors propose a "bi-directional" retrofitting scheme to refine past emotion predictions based on subsequent transformations, further enhancing the temporal consistency of conversational AI.
Summary updated