Introduction
In this project, we address this challenge through the design and development of a novel AI framework for the adaptable, robust, and explainable elicitation of multi-granular assembles of emotions.
Key points of AWARE compared to state-of-the-art emotion detection models.
- Ability to generate multifaceted emotion profiles with corresponding intensity scores.
- Example : “New treatment plan helps me to feel better and relaxed than ever”.
Emotion Profile : [ Joy : 0.530, Trust : 0.455, Anticipation : 0.048 …]
- Example : “New treatment plan helps me to feel better and relaxed than ever”.
- Robustness to intensifiers, inhibitors and negations.
- Negation :
- “Clinical staff was supportive.”. --- Joy
“Clinical staff was not supportive.”. --- Sadness
- Modifiers :
- “I’m feeling relaxed after talking to them." ---- Joy : 0.421
"I’m feeling extremely relaxed after talking to them." ---- Joy : 0.677
"I’m feeling little relaxed talking to them." ---- Joy : 0.332
- Negation :
- Ability to adapt to a given domain.
- Explainability in emotion extraction.
- “I’m feeling relaxed after the radiation therapy.” . --> relaxed
- "The leg is painful when walking." --> painful
Emotion AWARE is based on a hybrid/neuro-symbolic architecture which bridges language models and rule-based approaches. Here the basic idea is to extract significant implicit and explicit features from a given text and compare it with emotional concepts space and use most matching concepts to decide the emotional profile.
Framework
Following is the framework of Emotion AWARE.
How Emotion AWARE works.
Following is an example of Emotion AWARE works. Here it build emotion profile for sentence, “The movie had a great start, but the ending was awful” which has mixed emotions.
Technologies and areas
Python, Deep learning - Language models(BERT), Text similarity matching, word/sentence embedding, lexicon matching
Team
Gihan Gamage(me), A. Prof. Daswin De Silva, Prof. Damminda Alahakoon
Publications
Publication in review
