TL;DR
This paper introduces a novel approach, EE-CVAE, to generate dialogue responses that elicit a wider range of positive emotions, moving beyond single-sentiment generation. By leveraging a conditional variational autoencoder with latent emotion variables and discriminators, their model achieves significantly higher accuracy in eliciting specific positive emotions compared to existing methods.
What Problem Does This Paper Address?
Generating emotionally intelligent dialogue is a key goal in natural language processing. While previous work has focused on generating responses that express a particular emotion or a general positive sentiment, this paper tackles the more nuanced problem of eliciting specific positive emotions in the user. The challenge lies in the fact that positive sentiment is not monolithic; it encompasses a spectrum of emotions like joy, trust, and anticipation. Existing models often rely on limited, human-annotated datasets for single-emotion elicitation, which restricts their ability to capture this richness and scale. This paper aims to bridge this gap by developing a model that can elicit a variety of fine-grained positive emotions using large, unannotated datasets.
Methods
The core of the proposed method is the Emotion-Eliciting Conditional Variational Auto-Encoder (EE-CVAE). This model builds upon the Conditional Variational Auto-Encoder (CVAE) framework, which is commonly used for dialogue generation. The key innovation is the introduction of a latent variable e specifically designed to control the generation of responses that elicit designated emotions.
Here's a breakdown of the novel aspects:
- Latent Emotion Variable (
e): Unlike traditional CVAEs that use a latent variablezfor semantic and discourse-level variations, EE-CVAE introduceseto capture emotional intention. This allows the model to learn and represent different emotional states. - Two Discriminators: To guide the generation process and ensure the elicited emotions are coherent and accurate, two discriminators are employed:
Demo(Emotion Discriminator): This acts as a sentence classifier trained to predict the intended elicited emotion. It's crucial for enforcing the generation of coherent emotions.Dsem(Semantic Discriminator): This discriminator helps ensure that the generated response is semantically relevant and coherent, similar to existing methods.- Unsupervised and Semi-Supervised Learning: The model leverages large unannotated datasets for pre-training and then fine-tunes on a smaller, reconstructed dataset with emotion labels. The latent variable
eis trained to entail designated emotions using a small set of labeled examples, whilezis learned in a fully unsupervised manner. - Wake-Sleep Training: The authors employ a wake-sleep training schedule, where the generator is trained before the discriminator, which is a common technique for VAE-based models.
Datasets
The paper utilizes two main datasets:
- Friends and Open Subtitles: These large, unannotated datasets are used for pre-training the generator module of the model. This allows the model to learn general language generation capabilities.
- MEmoR (reconstructed): This dataset, originally multimodal, is reconstructed to focus on textual data for this task. It contains dialogue clips from "The Big Bang Theory" with emotion labels. The authors specifically extract non-speaker emotions as the target for elicitation. The reconstructed dataset has 22,732 utterances, split into training (18,943), development (1,894), and testing (1,894) sets. They focus on 3 positive emotions from Plutchik's 9 primary emotions for model output.
Key Insights and Ideas
- Richness of Positive Emotions: The paper strongly emphasizes that "positive emotion" is not a single entity but a spectrum of distinct feelings (joy, trust, anticipation). Eliciting these nuances is crucial for more sophisticated and engaging dialogue.
- Latent Variables for Emotion Control: Introducing a dedicated latent variable for emotion is a powerful way to control the generation of emotionally targeted responses, especially when dealing with multiple fine-grained emotions.
- Benefit of Negative Emotions in Training: Counter-intuitively, including negative emotions in the training data (even if the goal is to elicit positive ones) improves the model's ability to elicit positive emotions. This is attributed to increased data volume and the model learning to differentiate emotions better.
- Pre-training is Key: Pre-training the generation module significantly improves performance, especially for the proposed CVAE structure, suggesting it's more effective at leveraging large amounts of data for better emotion signal separation.
Key Results
- Improved Emotion Accuracy: The EE-CVAE model achieves 44.8% accuracy in eliciting target positive emotions, significantly outperforming the baseline EmpDG model which achieves 29.0% accuracy in human evaluations.
- More Fluent and Longer Responses: The proposed model generates more fluent responses (lower perplexity) and produces longer utterances (average length of 13.17, 43% longer than the baseline), which tend to be more generic and shorter.
- Effective Emotion Elicitation: Human evaluations confirm that the model is better at eliciting specific positive emotions compared to the baseline.
- Pre-training Benefits: Pre-training leads to a substantial reduction in perplexity for both models, with the proposed model showing a markedly larger reduction, indicating better utilization of wake-sleep learning.
Limitations and Future Work
- Focus on Positive Emotions: While the dataset contains negative emotions, the model is currently designed to elicit only positive emotions. The authors acknowledge that modeling various emotions would allow a system to be more aware of the correlation between intention and response, potentially avoiding eliciting negative emotions.
- Dataset Reconstruction: The MEmoR dataset was reconstructed from multimodal data, and the authors acknowledge the need for further validation and potential for more sophisticated annotation strategies.
- Few-Shot Learning: The paper suggests that rich emotion elicitation is a challenging task for current neural models and highlights the need for more effective few-shot learning approaches.
- Future Work: The authors intend to model various emotions to understand their correlations and enable more nuanced dialogue strategies, such as avoiding eliciting sadness. They also plan to release their code and data.
Summary updated