Language processing is one of the higher functions of the human brain, involving thinking, cognition and social interaction. The electroencephalography (EEG) based brain-computer interfaces (BCIs) provide a way to decode the complex cognitive processes in real-time, including the language processing activities. In this paper, we propose a novel linguistic BCI system based on an improved EEGNet, aiming at real-time decoding of semantic processing by categorizing the EEG responses to semantic stimuli. We designed three experimental paradigms to explore the effects of semantic stimuli on EEG activity: (1) Both reaction (BR), keystrokes were performed on all semantic stimuli; (2) No reaction (NR), subjects performed no operations on semantic stimuli; (3) Single reaction (SR), keystrokes were performed on semantically incongruent stimuli. We recorded EEG data from eight healthy volunteers, and observed the differences in EEG activities under different paradigms by ERP and time-frequency analysis. The results showed that the N400 component was significantly enhanced by semantic incongruent stimuli, and the N400 in experimental paradigms of both reaction and single reaction had high reliability, and the time-frequency analysis showed that semantic incongruent stimuli evoked enhanced activity in the theta band and reduced activity in the alpha band. For the classification model, we applied the original EEGNet and the improved EEGNet model and used the Synthetic Minority Over-sampling Technique (SMOTE) to balance the dataset. The classification acurracies for BR, NR and SR paradigms using the improved EEGNet were 70%, 66.7% and 87.1%, respectively. This indicates that the improved EEGNet model outperforms the original model in semantic classification tasks. The research in this paper provides a new experimental paradigm and modeling approach for the development of linguistic BCI systems, which can help to further understand and apply the advanced cognitive functions of the human brain.