A lightweight multi-task transformer for joint classification and priority ranking of community service requests on edge devices

Urban community service platforms receive thousands of textual requests every day, yet most deployed systems still run two separate pipelines: one for category classification and one for urgency ranking. The duplication wastes computation and sits poorly with the memory and power envelope of a community-level edge gateway. We address this gap with EdgeCSR-Net, a multi-task transformer that couples a depth-wise separable encoder, a grouped-query attention block and a convolutional positional encoding, so that thirty-six-way classification and four-level urgency ranking emerge from a single forward pass. Ranking is trained as a pairwise comparison and the resulting score is mapped onto the four operational levels by thresholds fixed on the validation split. The uncompressed encoder carries about 2.02 million parameters; a task-aware pipeline of structured pruning, distillation from a BERT-base teacher and INT8 quantization-aware training trims this to a 1.43-million-parameter, 1.5 MB deployable model. Rather than hand-setting the balance between the two objectives, we weight each loss by a learned task-noise term, an arrangement usually called homoscedastic uncertainty weighting, and pair it with a gradient-projection rule that suppresses conflicting updates on the shared encoder. Evaluated on a corpus of 47,832 requests from three metropolitan districts, EdgeCSR-Net reaches 88.1% macro-F1 and 0.897 NDCG@10. Against BERT-base it gives up 1.4 macro-F1 points and 0.015 NDCG@10 while carrying 77 times fewer parameters; against RoBERTa-base the corresponding figures are 1.8 points, 0.019 and 87 times. Under a matched INT8 pipeline it is significantly more accurate than TinyBERT and DistilBERT, statistically tied with MobileBERT, and six to ten times faster than all three. On an NVIDIA Jetson Xavier NX the quantized model finishes inference in 7.6 ms per request while occupying 34 MB of peak memory; on a Raspberry Pi 4B it takes 52.3 ms. Temporal and leave-one-district-out protocols cost only 1.7 and 0.8 macro-F1 points respectively, which indicates that neither the joint architecture nor its compression schedule buys real-time latency at the price of cross-region robustness.