Extracting the level-emotional pairs from sentences through the emotion classification task is helpful to mine the emotions of power users on different levels from the data middle platform, which has important significance for personalized recommendation. In the multimodal field, existing methods usually use two independent models to complete two subtasks respectively, term extraction of entities such as goods and important people contained in sentences or entity level emotion classification of entities according to the given level terms to predict the user’s emotional orientation. The above methods have two problems: 1) using two independent models will lose the continuity of the underlying features between the two tasks, and cannot model the potential semantic associations between sentences; 2) level-level emotion classification can only predict the emotion of one level at a time, which does not match the throughput of level term extraction to extract multiple levels at the same time, and the sequential execution of two models makes the extraction of level­ emotional pairs inefficient. Secondly, the model uses sequence labeling to simultaneously output multiple levels contained in a sentence and their nonresponding sentiment categories, which improves the efficiency of level ­ sentiment pair extraction. In addition, part-of-speech is introduced in these two subtasks: the grammatical information contained in it is used to improve the performance of level term extraction; the opinion word information is obtained through part-of-speech to improve the performance of level-level sentiment classification.