Attention-based CNN and BiLSTM hybrid Model for Aspect-level Sentiment Classification

Inhalt:
Aspect-based sentiment analysis(ABSA) is an important task in the field of text classification to analyze the subjective information of texts. Most existing ABSA models ignore the importance of represent the aspect and context separately and only use a single type of neural network to represent aspect and context. The most used are convolutional neural network(CNN) and Long Short-Term Model(LSTM). CNN can effectively extract phrase-level features, but it is easy to ignore the contextual relevance of the text. LSTM can capture the global structure information of the text, but it has insufficient ability to distill the key information. Considering different feature analysis demands of context and aspect, We use their different characteristics to propose a model based on them and attention, which is more accurate and efficient. CNN is used to extract the feature of aspect words, and it is fused with processed by Bi-LSTM. The attention mechanism captures the sentiment information of aspect and context. This process improves the accuracy of sentiment classification. The experiments conducted on SemEval 2014 task4 datasets(Li et al. 2015; Tang, Qin, and Liu 2015) (Pontiki et al., 2014) demonstrate the efficiency and effectiveness of our model.