Aiming　at　the　problem　of　sparse　Chinese　text　features　and　mixing　of　long　and　short　texts,　which　lead　to　the　difficulty　of　extracting　word　vector　features　and　the　single　convolution　kernel　of　traditional　neural　network　and　redundant　parameters,　the　ACL-RoBERTa-CNN　text　classification　model　uses　a　contrast　learning　method　to　learn　a　uniformly　distributed　vector　representation　in　order　to　achieve　the　effect　of　regular　expression　of　space.　Input　the　same　sentence　into　dropout　twice　as　＂positive　pairs＂,　replacing　the　traditional　data　enhancement　method.　Use　the　contrast　learned　RoBERTa　pre-training　model　to　train　the　word　vector,　send　the　word　vector　to　the　CNN　layer,　use　different　size　convolution　kernels　to　capture　the　information　of　different　length　words　in　each　piece　of　data,　finally　combine　with　Softmax　the　classifier　classifies　the　extracted　features.　The　experimental　results　on　two　public　data　sets　show　that　ACL-RoBERTa-CNN　classification　performance　is　better　than　TextCNN,　TextRNN,　LSTM-ATT,　RoBERTa-LSTM,　RoBERTa-CNN　and　other　deep　learning　text　classification　models.