<p>Intention recognition and slot filling are two major tasks in natural language understanding. Aiming at the existing joint model of intent recognition and slot filling, which fails to fully consider the influence of local features, relative position information and graph structure information on text features hidden in the dialog data, as well as the unsatisfactory intersection effect of the information of intent recognition and slot filling, a co-interaction for intent recognition and slot filling with global–local-spatial–temporal feature fusion model (GLST-Co-In-IRSF) is proposed. The multi-window recursive module (Mul-win) is used to construct a sliding window to generate a subsequence and extract detail information so as to obtain local features. GC-LSTM is utilized to obtain the relationship between nodes and nodes, nodes and edges, so as to obtain the spatio-temporal semantic features. Bi-directional connections between the two tasks of intent recognition and slot filling are modeled, and bi-directional connections are explicitly established. Finally, decoding is performed using intent classifier and conditional random field (CRF). The experimental results show that on the ATIS and Snips datasets, the GLST-Co-In-IRSF model achieves the lowest improvement in the three metrics of intention recognition accuracy, slot filling F1 value, and sentence accuracy over the JointBERT+CRF, JointIDSF, GC-LSTM, Slot-gate, and ESIE-BERT respectively, with the minimum improvement of 0.43%, 0.57% and 1.15%.</p>