Wondering About Dimensionality in Sentiment and Emotion Classification

Hello, I’m looking for pre-trained models that can provide some dimensionality for sentiments and emotions (e.g., amount of fear between 0-1). I’m most interested in the j-hartmann/emotion-english-distilroberta-base model. I’m wondering if the output scores for each emotion could be interpreted as a dimensional score (e.g., joy = .98, sad = .41). Is that possible? If not, is there a way to fine-tune the model to accommodate this? Are there models I’m missing that already do what I’m asking?

There are 7 classes and labels. Coming from this model. The output looks like this:
[[{‘label’: ‘anger’, ‘score’: 0.004419783595949411},
{‘label’: ‘disgust’, ‘score’: 0.0016119900392368436},
{‘label’: ‘fear’, ‘score’: 0.0004138521908316761},
{‘label’: ‘joy’, ‘score’: 0.9771687984466553},
{‘label’: ‘neutral’, ‘score’: 0.005764586851000786},
{‘label’: ‘sadness’, ‘score’: 0.002092392183840275},
{‘label’: ‘surprise’, ‘score’: 0.008528684265911579}]]

My question is whether the values can be interpreted as the amount of each emotion. For instance, can you say the text input is 97.7% “joy”? If not, can models like this be fine-tuned to add such emotional dimension or is there one that already exists?

My research question involves understanding the degree of fear in text. So, rather than just knowing whether the model thinks there’s fear in the text (i.e., a binary “yes” or “no”), I’d like to know how much fear is being communicated in the text (i.e., 15%, 81%, etc.). I would also like a sentiment model that shows the same dimensionality for positive and negative sentiment. Is this possible?

Thank you very much! :slight_smile: