Abstract: Identifying the sentiment polarity of movie customer group reviews may inspire the platform to optimize movie recommendation algorithms and improve services and provide suggestions for consumers' movie choices. A MABM sentiment polarity recognition model is proposed based on the multi-head self-attention bidirectional long-short term mechanism. Using the online review data of the well-known movie review website Douban Review as the corpus, text mining tools are used to pre-process the data. Then dividing the data into training and test sets according to 8:2, and using 10 machine learning models and 4 deep learning models as the control group, the effectiveness and robustness of the MABM model are evaluated by cross-validation and test set validation comparison. In the two sets of comparison experiments, it is found that the deep neural network model predicts overall better than the machine learning model, and the MABM model used as the control set has the best classification results. The MABM model can effectively identify the sentiment polarity of consumer reviews, providing management insights and algorithm improvement suggestions for movie recommendation platforms.