Meteorological　disaster　warning　information　plays　an　important　role　in　our　life.　However,　if　the　meteorological　department　mistakenly　sends　incorrect　meteorological　disaster　warning　information,　it　will　have　catastrophic　consequences.　Meteorological　disaster　warning　information　is　often　sent　in　the　form　of　text.　Therefore,　the　study　of　meteorological　disaster　warning　texts　is　very　important.　Different　from　many　studies　on　English　texts,　we　focus　on　Chinese　meteorological　disaster　warning　texts.　This　article　proposes　a　new　method　combining　neural　kernel　mapping　support　vector　machine(SVM)　and　multi-head　self-attention　mechanism　to　improve　the　accuracy　of　predicting　Chinese　meteorological　disaster　warning　texts.　Our　method　takes　multi-head　self-attention　mechanism　as　the　neural　kernel　mapping　of　support　vector　machine.　In　addition,　in　order　to　solve　the　problem　of　training　difficulties　caused　by　insufficient　Chinese　meteorological　disaster　warning　texts　data,　this　paper　develops　an　automatic　semantic　annotation　system　for　Chinese　meteorological　disaster　warning　texts.　Based　on　correct　Chinese　meteorological　disaster　warning　texts,　the　system　can　automatically　generate　sample　data　of　four　types　of　errors(including　wrong　words,　repetitions,　missing　words,　and　reverse　order).　In　our　experiments,　we　use　3　self-made　Chinese　meteorological　disaster　warning　text　datasets　and　4　other　types　of　Chinese　text　datasets.　The　experimental　results　show　that　compared　with　other　methods,　our　method　is　not　only　effective　for　Chinese　meteorological　disaster　warning　texts,　but　also　has　certain　advantages　for　other　types　of　Chinese　text　data　sets.