Publications.
Papers, preprints, and technical reports. Where work has been published or submitted, the venue and a link are below.
-
2026
Confidence- Aware Empathic Conversation frameworks by Semantic- Heuristic Gating and Anchoring Dynamics
M.moses,T.meshack, et al. · journal
In this paper, we design and propose a Confidence-Aware Empathic Conversation Framework to address the limitations of smaller-parameter Large Language Models (LLMs) such as FLAN-T5-Large in affective and mental health applications. We propose a novel architecture that gates LLM outputs …
In this paper, we design and propose a Confidence-Aware Empathic Conversation Framework to address the limitations of smaller-parameter Large Language Models (LLMs) such as FLAN-T5-Large in affective and mental health applications. We propose a novel architecture that gates LLM outputs through a dual-metric scoring system combining keyword-based heuristic empathy checking and vector-space semantic relevance evaluation, unified as Stotal = H + (3 × R). We implement a Dynamic Sentiment-Aware Anchoring strategy that pre-fills assistant responses with emotionally appropriate anchors depending on detected user sentiment, restricting the decoder search space to a corresponding emotional subspace. We further employ a Confidence Gate with a strict threshold (τ = 2.5) that triggers a safe fallback response when no candidate meets the required quality bar. We evaluate our system through a multi-turn interaction session on academic and workrelated stress scenarios using FLAN-T5-Large for generation and all-MiniLM-L6-v2 for semantic embedding evaluation. Our experimental results show that the proposed Hard-Anchoring strategy effectively prevents incoherent and emotionally misaligned outputs, with all generated responses remaining well above the confidence threshold. We achieve a mean system confidence of 4.74 across five turns, with the highest confidence score of 6.79 recorded when the system detected a user shift toward a positive emotional state and reinforced it appropriately. These results demonstrate that our framework substantially improves empathic coherence, response reliability, and emotional alignment in human-AI dialogue, offering a principled path toward safer AI deployment in mental health support contexts.