As the number of sensors becomes massive in Internet of Things (IoT) networks, the amount of data is humongous. To process data in real-time while protecting user privacy, federated learning (FL) has been regarded as an enabling technique to push edge intelligence into IoT networks with massive devices. However, FL latency increases dramatically due to the increase of the number of parameters in deep neural network and the limited computation and communication capabilities of IoT devices. To address this issue, we propose a semi-federated learning (SemiFL) paradigm in which network pruning and over-the-air computation are efficiently applied. To be specific, each small base station collects the raw data from its served sensors and trains its local pruned model. After that, the global aggregation of local gradients is achieved through over-the-air computation. We first analyze the performance of the proposed SemiFL by deriving its convergence upper bound. To reduce latency, a convergence-constrained SemiFL latency minimization problem is formulated. By decoupling the original problem into several sub-problems, iterative algorithms are designed to solve them efficiently. Finally, numerical simulations are conducted to verify the effectiveness of our proposed scheme in reducing latency and guaranteeing the identification accuracy.