Introduction
Standard neural dialogue systems, typically based on Sequence-to-Sequence (Seq2Seq) architectures, often produce generic, non-committal responses like "I don't know" or "I'm sorry." While grammatically sound, these systems lack the emotional intelligence required for natural human-computer interaction. Researchers from the University of Alberta address this by proposing a framework for controlled emotion expression, allowing chatbots to generate contextually relevant replies aligned with specific, pre-defined emotions.
Methodology
The authors extend the standard Seq2Seq model with an attention mechanism and explore three methods for injecting emotional context:
- Enc-bef (Emotion Before): Prepends a special emotion token to the input sequence (e.g.,
[SAD] I missed my deadline). - Enc-aft (Emotion After): Appends the emotion token to the end of the input.
- Dec (Emotion in Decoder): Injects a continuous emotion vector ($v_e$) directly into the LSTM decoder at every time step.
The Dec approach (illustrated above) is the most sophisticated, as it ensures the target emotion influences the entire generation process rather than just the initial context. To facilitate training, the researchers developed a Bi-LSTM emotion tagger to label the OpenSubtitles dataset, using the CBET dataset (81,000 instances across 9 emotions) for supervision.
Experimental Results
The models were evaluated on their ability to accurately express the requested emotion. The Dec model outperformed the token-based injection methods, achieving an average accuracy of 78.46%.
As shown in the table, performance varied significantly by emotion. "Sadness" (94.22%), "Guilt" (91.22%), and "Thankfulness" (91.06%) were the most successfully replicated. In contrast, "Joy" proved the most challenging (48.53%), often being confused with "thankfulness" or "love." Qualitative analysis revealed that the models learned specific linguistic patterns, such as starting guilty responses with "I am sorry."
Discussion and Conclusion
While the Dec model successfully generates emotionally charged responses, limitations remain. The system still exhibits the Seq2Seq tendency toward repetitive phrasing within emotional categories. Furthermore, the current implementation requires manual emotion selection. Future work aims to automate the selection of the most appropriate emotional response based on user input, moving closer to truly empathetic conversational agents.
Summary updated