The　existing　end-to-end　Aspect-Based　Sentiment　Analysis　(ABSA)　algorithms　focus　on　feature　extraction　by　a　single　model,　which　leads　to　the　loss　of　the　important　local　or　global　information.　In　order　to　capture　both　local　and　global　information　of　sentences,　an　end-to-end　ABSA　method　based　on　features　fusion　is　proposed.　Firstly,　the　pre-trained　model　BERT　is　applied　to　obtain　word　vectors;　secondly,　Iterated　Dilated　Convolutions　Neural　Networks　(IDCNN)　and　Bi-directional　Long　Short-Term　Memory　(BiLSTM)　with　Self-Attention　mechanism　(BLSA)　are　adopted　to　capture　local　and　global　features　of　sentences,　and　the　generated　local　and　context　dependency　vectors　are　fused　to　yield　feature　vectors.　Finally,　Conditional　Random　Fields　(CRF)　is　applied　to　predict　aspect　words　and　sentiment　polarity　simultaneously.　On　Laptop14　and　Restaurant　datasets,　our　model’s　F1　scores　increased　by　0.51%,　3.11%　respectively　compared　with　the　best　model　in　the　comparison　experiment,　and　0.74%,　0.78%　respectively　compared　with　the　single　model　with　the　best　effect　in　the　ablation　experiment.　We　removed　each　important　module　in　turn　in　subsequent　experiments　and　compared　it　with　our　model.　The　experimental　results　demonstrate　the　effectiveness　of　this　method　in　aspect　word　recognition　and　its　better　generalization　ability.　©　2023,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.