Post-implant telephone follow-up requires an agent to respond naturally, complete a structured standard operating procedure (SOP), and avoid unsupported medical advice within the same call. We present SAFER, a live speech-first system for Chinese neuromodulation follow-up that exposes three capabilities in one demonstration: real-time voice interaction with barge-in, an inspectable in-call SOP trace, and guarded handling of unscripted medical or device questions. In the dual-lane architecture, user-facing speech does not wait for asynchronous state commitment, while questions assigned to the guarded branch receive fixed non-generative deflection phrases; callback-eligible clinical or safety subclasses also create a human-follow-up record. Controlled studies report a median end-to-end voice latency of 1.07 seconds, a 0.983 transcript-level completeness-accuracy field score, and 23/30 joint routing, correctness, coverage, and grounding successes on independently expert-reviewed synthetic gold-ANSWER queries. The hosted demo lets reviewers conduct a call and inspect the evolving in-call record, available evidence text, latency, and patient-facing routing behavior. These results characterize a simulated prototype; they do not establish clinical safety or real-world robustness.

We propose to TransForm Scene Graphs into more descriptive Captions (TFSGC). In TFSGC, we apply multi-head attention (MHA) to design the Graph Neural Network (GNN) for embedding scene graphs. After embedding, different graph embeddings contain diverse specific knowledge for generating the words with different part-of-speech, e.g., object/attribute embedding is good for generating nouns/adjectives. Motivated by this, we design a Mixture-of-Expert (MOE)-based decoder, where each expert is built on MHA, for discriminating the graph embeddings to generate different kinds of words. Since both the encoder and decoder are built based on the MHA, as a result, we construct a simple and homogeneous encoder-decoder unlike the previous heterogeneous ones which usually apply Fully-Connected-based GNN and LSTM-based decoder. The homogeneous architecture enables us to unify the training configuration of the whole model instead of specifying different training strategies for diverse sub-networks as in the heterogeneous pipeline, which releases the training difficulty. Extensive experiments on the MS-COCO captioning benchmark validate the effectiveness of our TFSGC. The code is in: https://anonymous.4open.science/r/ACL23_TFSGC.