Netizens＇　complaint　reports　are　the　key　to　the　early　detection　and　treatment　of　environmental　water　pollution　events,　but　there　are　often　reports　that　are　malicious,　contain　exaggerations,　and　deliberately　exaggerate　the　facts.　The　manual　analysis　of　complaint　reports　is　complicated,　and　the　lack　of　sufficient　labeled　data　does　not　allow　for　an　effective　credibility　classification　model.　This　paper　proposes　a　deep　cross　domain　network,　which　learns　knowledge　from　the　source　domain　(microblog　texts)　and　applies　it　to　the　target　domain　(complaint　report　texts)　to　improve　the　complaint　report　reliability　classification　model＇s　performance.　First,　the　long　short　term　memory　extracts　the　domain-shared　features　of　the　source　and　target　domains,　the　source-private　features,　and　the　target-private　features.　Then,　the　self-attention　mechanism　fuses　the　domain-shared　features　and　the　domain-private　features.　Finally,　the　multilayer　perceptron　outputs　the　classification　result,　while　the　multi-kernel　maximum　mean　difference　is　used　to　calculate　the　distance　between　the　source　and　target　domains　as　part　of　the　loss　function.　Experiments　on　complaint　report　texts　and　microblog　texts　show　that　our　proposed　method　consistently　outperforms　other　nonknowledge　transfer　models　and　completes　the　credibility　analysis　of　water　environment　complaint　reports　well.　This　paper　provides　an　innovative　method　for　the　credibility　analysis　of　water　environment　complaint　reports.