Automatically　extracting　application　features　from　their　descriptions　has　become　an　efficient　way　to　understand　user　requirements　and　accumulate　related　domain　knowledge.　Existing　approaches　typically　extract　features　based　on　their　syntactic　patterns,　which　may　lead　to　lots　of　false　positives　as　the　sentences　to　be　processed　do　not　contain　any　features.　In　this　paper,　we　first　propose　a　POS-weighted　sentence　classifier　based　on　advanced　word　embedding　techniques　to　filter　non-feature-containing　sentences　before　feature　extraction.　Specifically,　we　assign　different　POS　tags　with　different　weights　according　to　their　importance　in　sentences.　Secondly,　we　defined　a　group　of　patterns　with　dual　constraints　of　POS　and　dependency　relationship,　then　match　phrases　from　each　feature-containing　sentence　to　obtain　features.　To　evaluate　the　performance　of　our　classifier,　we　rigorously　produced　a　dataset　with　corresponding　annotations.　The　result　shows　that　our　classifier　can　successfully　filter　out　79%　of　non-feature-containing　sentences.　Applying　our　method　to　eight　applications,　it　outperforms　the　state-of-the-art　approach　in　precision,　recall,　and　f-measure.　©　2021　Owner/Author.