ADeCNN: An improved model for aspect-level sentiment analysis based on deformable CNN and attention

Aspect-level sentiment analysis aims at identifying the sentiment polarity of target in the context. In most of the previous sentiment analysis models, there usually exists the problem of insufficient extraction capability of local features and long-distance dependency features. To solve the above problem, in this paper, we propose an improved model (called ADeCNN) for aspect-level sentiment analysis, by incorporating the attention mechanism into the deformable CNN model. In ADeCNN, we use deformable convolutional layers and bi-directional long short-term memory network (Bi-LSTM), combined with sentence-level attention, to extract sentiment features, and to break through the limitations of the model's long-distance dependency feature extraction capability. We then use a gated end-to-end memory network (GMemN2N) to integrate the target into the sentiment feature extraction process, so as to obtain sentiment features. And finally, we obtain the corresponding sentiment analysis results through the classifier. In addition, in order to solve the problem that the same words have large differences in the polarity of sentiments expressed in different targets, the model is also constructed with the ability to generate different attention weights based on target to assist sentiment analysis, with the aim of further enhancing the correlation between the target and the words in the sentence. We setup experiments to demonstrate the functionality effectiveness and performance gains of ADeCNN, based on the SemEval 2014 Task4 and SemEval 2017 Task4 datasets. Extensive experimental results show that ADeCNN outperforms its competitors, producing an arresting increase of the classification accuracy on all the three datasets of Laptop, Restaurant, and Twitter.