Bullet　screen　is　a　new　type　of　comment　appearing　on　video　websites.　When　the　video　is　played,　it　will　be　displayed　in　real　time.　Through　sentimental　analysis　of　the　screen,　we　can　quickly　understand　the　sentimental　types　of　the　current　video.　However,　it　is　unrealistic　to　deal　with　all　the　bullet　screens　of　the　whole　video　at　one　time　with　traditional　methods.　This　paper　proposes　a　method　of　aggregating　all　text　content　in　video　to　represent　very　long　video　content　to　achieve　video　feature　extraction.　The　method　proposed　in　this　paper　is　more　effective　than　PCA　and　scalable　methods　in　reducing　dimension　and　achieving　aggregation.　©　2023　SPIE.