News　and　social　media　messages　usually　contain　subjective　opinions　conflicting　with　the　needs　of　readers　who　want　to　receive　objective　information　through　public　channels.　To　this　end,　the　detection　of　subjectively　biased　sentences　has　become　an　important　research　issue.　However,　existing　subjective　bias　detection　approaches　lack　considering　the　syntactic　structure　and　topical　context　of　biased　descriptions.　In　this　paper,　we　propose　a　Subjective　bIas　deTection　mEthod　(SITE)　that　comprehensively　fuses　multiple　bias-relevant　information.　Specifically,　we　first　investigate　the　modification　and　lexical　features　of　biased　sentences,　based　on　which　we　formulate　a　set　of　rules　to　characterize　biased　sentences.　Then,　we　extract　the　semantic　features　of　sentences　using　the　BERT　model,　based　on　which　we　further　mine　topic　features　by　clustering　semantically　similar　sentences.　Finally,　we　comprehensively　characterize　biased　sentences　by　fusing　such　features　and　train　a　classification　model　to　detect　biased　sentences　in　social　media.　We　conducted　a　series　of　experiments　on　a　public　dataset,　the　results　of　which　show　that　SITE　can　detect　biased　sentences　with　86.2%　accuracy,　outperforming　baseline　methods.　©　2023　Knowledge　Systems　Institute　Graduate　School.　All　rights　reserved.