Rehabilitation chatbots must extract structured events and estimate users’ self-efficacy in real time, yet Japanese dialogue—rich in honorifics, ellipsis, and loanwords—plus limited annotations and edge-device constraints make this difficult. We propose Synergy-BERT, a lightweight multi-task framework that jointly performs event extraction (event, rationale, facilitator, barrier) and self-efficacy assessment. The model augments a Japanese BERT encoder with a BiLSTM, five learnable task tokens, and a synergistic attention module enabling bidirectional interaction between tasks and the sequence; BIO tags are decoded with a CRF, and scores are produced by a compact MLP. A progressive fine-tuning schedule freezes most pre-trained layers before partially unfreezing for domain transfer, mitigating gradient interference and improving convergence stability. Experiments on real Japanese rehabilitation dialogues show consistent improvements over a strong BERT baseline while preserving low-latency inference suitable for edge deployment. The framework provides a practical, pluggable core for rehabilitation assistants and opens avenues for multimodal inputs, energy-aware optimization, and cross-lingual transfer.