In response to the lack of context balance, insufficient introduction of emotional attention to network nodes, and inability to effectively extract the deep meaning of implicit emotional sentences in most current implicit sentiment analysis methods, this paper proposes a method that integrates network representation learning with attention mechanism for implicit sentiment analysis. The research first models and calculates basic vector representations such as word mixing sequences, part of speech, and sentence dependency relationships. Then, it focuses on calculating the feature values of network node attributes, and synchronously combines network representation learning and implicit sentiment analysis part of speech classification knowledge to comprehensively classify and weight emotional texts. Finally, convolutional neural networks are used to comprehensively calculate the correlation between implicit keywords, key sentences, and the entire text, output the probability score of sentiment polarity, and complete the analysis. The experiment was conducted in SemEval-2014 Task 4, and the results showed an improvement in F1 score compared to similar context based and deep learning models.