Aspect extraction is the basis for aspect-based sentiment analysis, aiming to find out the target of opinions from reviews. The existing neural network for aspect extraction model based on sequence labeling has a poor effect on the extraction of long aspect terms. To solve this problem, we propose a new aspect extraction framework, which uses a three-layer convolutional neural network (CNN) to learn multi-layer semantic features from re-views, and then uses the non-local attention mechanism to obtain dependence feature be-tween different words. Moreover, conditional random field (CRF) has been used to reduce the probability of label conversion errors. CNN filters are good at learning local features without considering long-distance dependencies, while the non-local attention mechanism can strengthen the long-distance dependencies of words and ensure the integrity of long aspect terms. The proposed model is tested on two datasets of SemEval and compared with some baseline models. The experimental results show that the performance of the model is superior.