Recommender　system　usually　suffers　from　bias　problem.　To　deal　with　this　problem,　recommendation　debiasing　has　received　much　attention　recently.　In　line　with　this　trend,　our　work　aims　to　shed　light　on　a　newly　identified　bias　known　as　sentiment　bias　-　the　divergence　in　recommendation　performance　between　positive　users/items　and　negative　users/items.　By　investigating　this　aspect,　we　hope　to　contribute　to　understanding　and　mitigating　the　sentiment　bias.　Existing　efforts　usually　focus　on　utilizing　regularization　strategies　or　causal　interventions　for　eliminating　the　bias.　However,　these　methods　fall　short　of　decomposing　the　underlying　user　preference　or　sentiment　bias,　which　limits　their　ability　to　significantly　enhance　recommendation　performance.　In　this　study,　we　address　the　issue　of　sentiment　bias　by　incorporating　disentanglement　and　attention　networks　from　a　causal　perspective.　We　construct　a　causal　graph　to　capture　the　cause-effect　relationships　in　the　recommender　system,　specifically　considering　the　role　of　sentiment　polarity　expressed　in　review　text　as　a　confounding　factor　between　user/item　representations　and　observed　ratings.　To　mitigate　the　negative　impact　of　sentiment　bias,　we　propose　a　disentangled　framework　that　learns　representations　where　user　preferences　and　sentiment　bias　are　structurally　disentangled.　Furthermore,　we　use　an　attention　network　to　extract　the　expected　sentiment　of　the　candidate　items　from　the　users　historical　behaviors.　Extensive　experiments　conducted　on　five　benchmark　datasets　validate　the　effectiveness　of　disentanglement　in　removing　sentiment　bias.