The　task　of　cross-domain　sentiment　classification　is　to　analyze　the　sentiment　orientation　of　the　target　domain　lacking　labeled　data　using　the　source-domain　data　with　sentiment　labels.　A　hierarchical　attention　model　based　on　Wasserstein　distance　is　proposed　in　this　paper.　The　hierarchical　model　is　used　for　feature　extraction　by　combining　attention　mechanism,　and　Wasserstein　distance　is　used　as　the　domain　difference　metric　to　automatically　capture　the　domain-sharing　features　through　adversarial　training.　Further　auxiliary　task　is　constructed　to　capture　the　domain-special　features　cooccurring　with　domain-sharing　features.　These　two　kinds　of　features　are　united　to　complete　the　cross-domain　sentiment　classification　task.　The　experimental　results　on　Amazon　datasets　demonstrate　that　the　proposed　model　achieves　a　higher　accuracy　and　a　better　stability　on　different　cross-domain　pairs.　©　2019,　Science　Press.　All　right　reserved.