Multimodal Understanding and Personalized Dispatching for Property Service Orders via Pretraining Based Fusion

In the domain of property services, traditional rule-based work order dispatching systems are challenged by issues of accuracy and efficiency. To address these concerns, this study introduces a personalized dispatching framework that integrates multimodal understanding with pre-training-based fusion techniques. Utilizing Meta’s Llama 2 as the pre-training model, we have developed a fusion algorithm that integrates data from diverse modalities—i.e. text, image, video, and audio—with multi-dimensional worker profiles. This pre-training fusion framework enables our system to interpret and process the multi-level demands of work orders, facilitating rapid and precise dispatching tailored to the skills, locations, and availability of workers, as well as the urgency and type of work orders. Experimental evaluation on a large-scale property service dataset indicates our method’s superiority over traditional rule-based systems, achieving an approximate 19% increase in dispatch accuracy and 20% increase in dispatch speed. Moreover, our model optimizes workers’ routes, supporting the processing of multiple work orders and routes, thus enhancing dispatching efficiency. This study not only offers an innovative approach to personalized dispatching of property service orders but also illustrates the practical potential of pre-training-based fusion techniques. Future research will divide into optimizing fusion strategies and expanding this approach to a wider range of service dispatching scenarios.