MEMORY INTERACTIVE NETWORK FOR ASPECT-BASED SENTIMENT ANALYSIS
At present,most aspect-based sentiment analysis solutions are based on the embedding of display aspect words and their position weighted expansion,but this kind of method is no longer applicable in implicit aspect words and long text scenes.Therefore,this paper proposes a memory interaction network.The long text was split into multiple short sentences,and a multi-[CLS]input structure was constructed.The aspect phrase vector,each short sentence vector and the[CLS]vector of each short sentence were obtained from BERT.After multiple attention interactions,deep emotion classification features were obtained.The results show that the Macro F1 of this model reaches 70.40％,and the F1 value of each category is higher than other models.
万方数据
维普
