Attention　networks　have　had　great　success　in　sentiment　analysis　in　the　last　few　years,　encouraging　growing　numbers　investigating　what　aspects　of　sentimental　information　they　can　learn.　However,　most　recent　studies　focus　on　the　variation　of　representations　of　word　vectors　and　network　outputs.　In　this　paper,　we　focus　on　analyzing　attention　mechanisms　in　the　case　of　sentiment　analysis.　We　hypothesize　that　in　the　humor　level　evaluation　task,　attention　mechanisms　pay　more　attention　to　funny　words　when　asserting　humor　level.　By　exploring　the　attention　distribution　patterns,　we　find　that　attention　patterns　are　likely　to　distribute　more　attention　to　contextual　information　rather　than　only　to　some　words.　We　further　show　that　during　the　training　process,　the　attention　patterns　assign　more　and　more　weight　on　the　noun,　adjective,　or　verb　at　first　but　give　less　in　the　few　latter　epoch,　which　means　that　attention　mechanisms　does　pay　more　attention　in　the　early　epochs　to　words　which　have　obvious　ostensible　meaning　but　distribute　more　weight　on　contextual　information　in　the　end.　©　2022　IEEE.