Aspect-level sentiment classification based on location and hybrid multi attention mechanism

摘要

The main task of aspect-level sentiment classification is to judge the sentiment polarity of a sentence under a given aspect word. Existing methods such as recurrent neural networks or attention mechanisms cannot make full use of location information. Moreover, when the sentence length is longer and the grammatical structure is more complicated, the above methods may cause the loss of important information and it is difficult to dig deeper semantic emotional features. Therefore, we propose a hybrid network model, and this model designs a shallow and deep two-layer network, and constructs a positional attention mechanism and an interactive multi-head attention mechanism in the corresponding network to capture multi-level emotional characteristics. The experimental results show that, in most case, the model performs better than the relevant baseline model on Restaurant and Laptop of the SemEval 2014 and ACL14 Twitter, and can effectively identify different aspects of emotional polarity.