Intent　recognition　and　slot　filling　play　crucial　roles　in　natural　language　processing.　Therefore,　this　study　aims　to　enhance　the　performance　of　these　two　tasks.　Given　the　close　interrelation　between　intent　recognition　and　slot　filling,　this　paper　takes　a　feature　encoding　model　as　the　baseline　and　augments　it　with　a　GRU　network　decoding　layer　and　a　TextCNN-based　local　semantic　information　feature　representation　layer.　This　joint　neural　network　model　facilitates　feature　extraction　in　both　temporal　and　spatial　dimensions　and　incorporates　a　keyword　attention　mechanism,　enabling　it　to　capture　the　expression　of　contextual　semantic　information　more　precisely.　Model　training　is　conducted　using　the　PGB　adversarial　training　method　to　enhance　the　model＇s　resilience　against　attacks.　Additionally,　an　asynchronous　training　strategy　is　employed　to　enable　multiple　models　to　learn　and　adapt　independently,　accelerating　the　model　training　process　and　enhancing　its　ability　to　learn　and　capture　contextual　semantic　information　more　effectively.　©　2024　SPIE.