Effective　sentiment　analysis　on　social　media　data　can　help　to　better　understand　the　public＇s　sentiment　and　opinion　tendencies.　Combining　multimodal　content　for　sentiment　classification　uses　the　correlation　information　of　data　between　modalities,　thereby　avoiding　the　situation　that　a　single　modality　does　not　fully　grasp　the　overall　sentiment.　This　paper　proposes　a　multimodal　sentiment　recognition　model　based　on　the　attention　mechanism.　Through　transfer　learning,　the　latest　pre-Trained　model　is　used　to　extract　preliminary　features　of　text　and　image,　and　the　attention　mechanism　is　deployed　to　achieve　further　feature　extraction　of　prominent　image　key　regions　and　text　keywords,　better　mining　the　internal　information　of　modalities　and　learning　the　interaction　between　modalities.　In　view　of　the　different　contribution　of　each　modal　to　sentiment　classification,　a　decision-level　fusion　method　is　proposed　to　design　fusion　rules　to　integrate　the　classification　results　of　each　modal　to　obtain　the　final　sentiment　recognition　result.　This　model　integrates　various　unimodal　features　well,　and　effectively　mines　the　emotional　information　expressed　in　Internet　social　media　comments.　This　method　is　experimentally　tested　on　the　Twitter　dataset,　and　the　results　show　that　the　classification　accuracy　of　sentiment　recognition　is　significantly　improved　compared　with　the　single-modal　method.　©　2022　ACM.