This　paper　presents　a　self-training　word　embedding　text　classification　model　based　on　knowledge　graph　expansion　for　text　classification.　Current　mixed　word　embedding　methods　are　overly　dependent　on　the　Fasttext　pre-training　model　and　here　is　still　a　problem　of　missing　words　with　rich　semantic　information　are　not　mapped.　First,　we　propose　a　method　for　extracting　missing　nouns　based　on　shape　near　word　filtering.　Second,　we　design　a　self-training　word　embedding　method　based　on　knowledge　graph　that　mixes　with　pre-training　word　embedding　to　obtain　a　high-quality　mixed　word　vector　with　rich　semantics　and　rich　semantics.　Third,　we　designed　a　GRU　model　based　on　improved　mixed　word　embedding　to　improve　the　quality　of　text　classification.　Experiments　conducted　on　multiple　text　classification　datasets　demonstrate　that　our　methods　can　effectively　improve　the　text　classification　accuracy.　©　2019　IEEE.