Dimensionality　reduction　is　an　important　technique　in　machine　learning　and　data　mining,　which　makes　the　processing　of　high　dimensional　data　faster.　An　efficient　method　for　dimensionality　reduction　can　find　a　low-dimension　feature　subset　extracting　the　most　relevant　information.　The　dimensionality　reduction　methods　based　on　neural　network　are　applied　to　all　kinds　of　data,　especially　computer　vision　data.　In　this　paper,　we　focus　on　the　text　data　with　high　sparse　and　high　dimension,　then　reduce　its　dimension　by　using　the　variational　auto-encoder.　The　performance　of　variational　auto-encoder　in　dimensionality　reduction　is　observed　by　comparison　test.　First,　unstructured　text　data　is　converted　to　computer-processable　vectors　using　term　frequencyCinverse　document　frequency.　Then　variational　auto-encoder　is　used　to　reduce　the　dimensionality.　Finally,　the　experiment　verifies　the　efficiency　of　variational　auto-encoder　by　comparing　seven　commonly　used　dimensionality　reduction　methods.