AI Research Analysis

Experimental results demonstrate that DeepSeek-R1 excels in emotional adaptability, balancing contextual and emotional understanding, achieving maximum empathy (5.0) with robust coherence (1.8). ChatGPT 4 also reaches high empathy (5.0) but often at the cost of coherence. Lexicon integration consistently boosts empathy scores across models (e.g., Llama 2 empathy +81%), but can introduce trade-offs with coherence. Our framework surpasses baseline models in empathy, coherence, informativeness, and fluency, particularly in emotionally sensitive psychotherapy contexts.

A detailed comparison of LLMs with various lexicons reveals specific strengths and weaknesses. VADER significantly enhances empathy (e.g., Flan-T5 +392%, ChatGPT 4 +462%) but often reduces coherence. WordNet generally improves coherence (+18%) more than VADER or SentiNet, though with moderate empathy gains. SentiNet maximizes informativeness (+22%) but can negatively impact fluency. DeepSeek-R1 consistently shows superior performance in balancing these metrics, demonstrating better contextual reasoning and emotional intelligence compared to other models across different lexicon combinations, making it highly effective for structured yet empathetic responses.

The Emotion-Aware Embedding Fusion framework has direct applications in AI-driven therapy chatbots, crisis intervention systems, and hybrid AI-therapist collaboration. It enables chatbots (e.g., on platforms like Woebot, Wysa) to generate contextually relevant and empathetic responses by retrieving past session data. In crisis scenarios, it can detect high-risk emotional states for real-time triage. For human therapists, it provides emotion tracking and trend analysis over multiple sessions, enhancing long-term patient assessments and supporting cognitive behavioral therapy. This model contributes to accessible and effective mental health support globally, especially in underserved regions.