Mobile Edge NLU with On-device Inference for Humanitarian Assistance during Disasters
Karunchai Mabuntham, Worawan Marurngsith · 2022
Natural language understanding (NLU) models allow computers to help disaster response teams in understanding requests spoken in different languages. Multilingual assistants can automatically translate messages spoken in a detected language into a language understood by the NLU model, prior to translating the extracted information back to the response team. The success of humanitarian assistance during disasters depends largely on the performance and dependability of the assistants, especially in a resource-limited situation. Recent works have successfully pushed natural language computation to special-purpose AI edge nodes, the computing devices located at the edge of a network, to achieving availability and resilience to network loss. However, limited works have offloaded the natural language workload to process on mobile devices and quantify the resource requirement.Thus, this work proposed a use case of mobile devices connected as an NLU edge computing node for understanding disaster response messages. Its resource requirement has been measured based on a transformer-based Thai-English translation model. The results show that the model achieve an acceptable performance in language translation and classification of needed humanitarian supplies. Moreover, using the cloud AI-as-a-Service, we could extract key essences of the disaster response messages in acceptable response time.