Keyword　extraction　is　a　critical　technique　for　document　retrieval　and　text　mining,　Web　page　retrieval　and　document　clustering.　The　traditional　keyword　extraction　method　is　overly　dependent　on　word　frequency,　which　may　lead　to　the　limitations　of　the　keyword　extraction　in　short　sentences.　In　order　to　solve　this　problem,　we　propose　a　novel　word　embedding　generation　method　for　keyword　extraction,　which　trains　a　special　domain　word　embedding　to　extract　keywords　automatically　from　user-generated　query　words.　To　ensure　that　the　experimental　results　are　not　biased　by　the　above　test　sample,　we　train　the　word　embedding　with　the　Chinese　version　of　Wikipedia　for　contrast　experiment.　Compared　with　other　methods,　the　recall　rate　of　the　proposed　method　reaches　92.55%,　higher　than　the　other　current　methods.