Abstract: Assembly semantics, as a vital content of an assembly model, is mainly set interactively by designers, which is usually time-consuming and inefficient. To account for this, an intelligent approach to identifying and setting dual-level assembly semantics is proposed. First, the existing graph attention network is improved where it is extended to a dual-level identification network, identifying all typical kinematic pair interfaces on each part model that have various geometric shapes but consistent kinematic semantics. After that, the existing back-propagation artificial neural network structure is modified for improving performance, recognizing all assembly constraint types (as well as their associated geometric entities) that are embodied in each kinematic pair interface. Based on the above-identified information, the mating kinematic pair interfaces and mating assembly-constraint geometric entities between arbitrary two part models can be searched automatically, and the full assembly semantics between them can be rapidly and semi-automatically set. To train aforementioned network model, a CAD part dataset with manual labeling is constructed. Experiments show that the accuracy of identification on the kinematic pair interface or on the assembly constraint type (as well as its associated geometric entity) is more than 93%. Besides, compared with several state-of-the-art works, the proposed approach also has some advantages and potentials to set the assembly semantics for various assembly models rapidly.