Abstract:  Aspect-level sentiment analysis is a prominent research task in the field of natural language processing. Aiming to analyze the sentiment tendencies of different aspects of texts, to address the issues of insufficient interaction between aspect words and context, and to deal with low classification accuracy of existing aspect-level sentiment classification models, an ASMFF(Aspect-level Sentiment classification method is proposed based on Multi-interaction Feature Fusion). Firstly, the context and aspect words are distinctly labeled and fed into the BERT(Bidirectional Encoder Representations from Transformers) coding layer for text feature vector extraction. Secondly, the text feature vectors are fed into AOA (Attention Over Attention) and IAN (Interactive Attention Networks) networks to extract the interactive attention feature vectors. Finally, the two interactive feature vectors obtained are fused and learned, and probability calculation, loss back propagation, and parameter updating are carried out using the cross-entropy loss function. Experimental results on three publicly available datasets, Laptop, Restaurant, and Twitter, show that the classification accuracy of the ASMFF model is 80. 25%,84.38%, and 75.29%, respectively, which is a significant improvement over the baseline model. 

Key words: aspect-level sentiment classification, natural language processing, interactive attention networks, multi-interaction feature fusion