Biases　and　de-biasing　in　recommender　systems　have　received　increasing　attention　recently.　This　study　focuses　on　a　newly　identified　bias,　i.e.,　sentiment　bias,　which　is　defined　as　the　divergence　in　recommendation　performance　between　positive　users/items　and　negative　users/items.　Existing　methods　typically　employ　a　regularization　strategy　to　eliminate　the　bias.　However,　blindly　fitting　the　data　without　modifying　the　training　procedure　would　result　in　a　biased　model,　sacrificing　recommendation　performance.　In　this　study,　we　resolve　the　sentiment　bias　with　causal　reasoning.　We　develop　a　causal　graph　to　model　the　cause-effect　relationships　in　recommender　systems,　in　which　the　sentiment　polarity　presented　by　review　text　acts　as　a　confounder　between　user/item　representations　and　observed　ratings.　The　existence　of　confounders　inspires　us　to　go　beyond　conditional　probability　and　embrace　causal　inference.　To　that　aim,　we　use　causal　intervention　in　model　training　to　remove　the　negative　effect　of　sentiment　bias.　Furthermore,　during　model　inference,　we　adjust　the　prediction　score　to　produce　personalized　recommendations.　Extensive　experiments　on　five　benchmark　datasets　validate　that　the　deconfounded　training　can　remove　the　sentiment　bias　and　the　inference　adjustment　is　helpful　to　improve　recommendation　accuracy.　©　2022　ACM.