With the development of professional online review systems, current sentiment analysis research is more focused on classifying the sentiment polarity of different aspects of the same subject, which also helps consumers to make decisions. To address the problems of multiple meanings of words, inadequate contextual semantic understanding and incomplete feature extraction in online reviews, this paper proposes an aspect-level sentiment analysis model based on BERT-BiSTM (Bidirectional Encoder Representation from Transformers and Bi-directional Long Short-Term Memory). The BERT word vectorisation can make use of the information in both directions in the text and can better learn the semantic information of the text. The results of the BERT model are fed into the BiLSTM model to fully perform feature mining for aspect extraction and sentiment polarity classification. The experimental results show that the model in this paper works well on the tasks of aspect extraction and sentiment polarity classification on the four datasets Semeval2014_restaurant (Res14), Semeval2014_laptop (Lap14), Semeval2015_restaurant (Res15) and Semeval2016_restaurant (Res16).