Eliciting　security　requirements　in　early　stage　of　system　development　has　been　widely　recognized　as　an　efficient　way　for　minimizing　security　cost　and　avoiding　recurring　security　problems.　However,　in　many　projects,　security　requirements　are　not　explicitly　specified　but　rather　mixed　with　other　requirements,　requiring　precise　and　fast　identification　of　such　security　requirements.　Although　several　probability-based　approaches　have　been　proposed　to　tackle　this　problem,　they　are　either　imprecise　or　domain-dependent.　In　this　paper,　we　propose　a　tool　supported　method　to　efficiently　identify　security　requirements,　which　combines　linguistic　analysis　with　machine　learning　techniques.　In　particular,　we　apply　a　systematic　approach　to　identify　linguistic　features　of　security　requirements　based　on　existing　security　requirements　ontologies　and　linguistic　knowledge.　We　automatically　extract　such　features　from　textual　requirements,　which　are　then　used　to　train　security　requirements　classifiers　using　typical　machine　learning　techniques.　We　have　implemented　a　prototype　tool　to　support　our　approach,　and　have　systematically　evaluated　our　approach　based　on　three　realistic　requirements　specifications.　The　evaluation　results　show　that　our　approach　has　promising　potential　to　train　classifiers　that　can　classify　requirements　specifications　from　different　application　domains.