One　of　key　problems　in　implementing　a　dynamic　interface　between　human　and　agents　is　how　to　do　semantic　mapping　from　natural　language　questions　to　OWL.　The　paper　views　the　task　as　a　two-class　classification　problem.　A　pair　of　question　variable　and　OWL　element　is　a　sample.　Two　classes　of　＇Matched＇　and　＇Unmatched＇　explain　two　relations　between　the　question　variable　and　the　OWL　element　in　a　given　sample.　Building　appropriate　semantic　mapping　is　the　same　as　classifying　the　sample　to　a　＇Matched＇　class　by　an　effective　machine　learning　method　and　a　trained　model.　Two　types　of　features　of　samples　are　selected.　Syntactical　features　denote　the　syntactical　structure　of　a　given　sample.　Semantic　features　present　multiple　relations　between　the　question　variable　and　the　OWL　element　in　one　sample.　Preliminary　experimental　results　show　that　the　sum　precision　of　the　learning-based　model　is　better　than　that　of　the　constraints-based　method.　©　Springer-Verlag　Berlin　Heidelberg　2007.