Community　detection　is　an　important　topic　in　complex　network　analysis　which　can　explore　valuable　relationships　in　the　networks,　such　as　protein-protein　interactions,　advertisement　recommendations,　etc.　Recently,　the　structure　and　attributes　of　a　network　are　expected　to　be　integrated　to　obtain　a　more　accurate　community　division.　But　existing　community　detection　algorithms　based　on　multiobjective　optimization　evolutionary　algorithms　(MOEAs)　for　attributed　networks　have　two　common　problems.　First,　their　encoding　strategies　completely　depend　on　the　network　structure,　which　limits　their　use　of　attribute　information　in　search.　Second,　the　calculation　of　the　attribute　objective　function　is　time-consuming.　In　this　paper,　we　propose　a　novel　algorithm　that　combines　the　nodes＇　embedding　vectors　generated　by　a　Skip-Gram　model　with　an　attention-based　multiobjective　optimization　evolutionary　algorithm　to　discover　overlapping　communities　on　networks　with　attributes.　With　the　help　of　embedding　vectors,　the　attention-based　encoding　strategy　can　overcome　the　problem　of　the　limited　searching　capability　of　traditional　MOEAs＇　encoding　schemes　that　depend　only　on　a　network　structure,　and　an　attribute　objective　function　based　on　embedding　vectors　is　designed　which　can　be　calculated　in　linear　time　to　improve　the　computational　efficiency.　The　statistical　results　in　artificial　and　real-world　networks　demonstrate　the　feasibility　and　effectiveness　of　the　proposed　method.