A　new　information　extraction　method　which　is　based　on　Boosting　algorithm　is　provided.　It　can　automatically　generate　a　rule　based　on　a　training　instance.　This　rule　is　applied　to　training　set　and　change　the　probability　distribution　on　the　weights　of　positive　examples.　Next　instance　will　be　selected　from　training　set　based　on　this　distribution.　A　constraint　named　mode-match　which　can　describe　words　that　do　not　accord　with　lexical　rules　is　provided　too.　As　experiments　show,　for　the　texts　with　simple　characters,　both　recall　and　precision　can　be　achieved　to　100%.　Even　for　the　texts　with　complex　characters,　the　evaluation　of　F1　can　be　achieved　to　80%.