Low-resource sentiment analysis faces problems such as imbalanced data distribution and low recognition rate of ambiguous emotions due to the scarcity of corpus. This is because the extraction of emotional features is constrained by data distribution and diversity, and it is easy to produce biases when modeling emotional corpus. To address these issues, we use the pre-trained XLM-R model to integrate the word embeddings of low-resource sentiment sentences into sentence embeddings. While using resampling and fast masking language modeling methods to balance the sampling probability of pre-trained data and local data, we also save computing costs. At the same time, since the relevance of sentence embeddings is lower than that of word embeddings, we adopt the strategies of word repetition and momentum contrast to improve the quality of positive and negative samples, and optimize the clustering effect of sentence embeddings in combination with the unbiased contrast learning paradigm. In addition, in the task of sentiment classification, this paper adds the ECAnet module to the traditional classification model DPCNN and uses the FocalLoss to improve the loss function, which enhances the ability to capture multidimensional features of key emotional information, thereby alleviating emotional confusion. We conducted experiments and validations on two Mongolian sentiment datasets, Mongolian and WCM, and three Chinese sentiment datasets, NLP&CC2013, NLP&CC2014, and WEC. The classification accuracy and F1 value both showed significant improvements, which fully demonstrated the effectiveness of the proposed method.