A gated hierarchical lstms for target-based sentiment analysis

Abstract

The deep neural model combining attention mechanism has achieved remarkable success in the task of targetbased sentiment analysis. In current research, the attention mechanism is more broadly combined with LSTM(Long Short-Term Memory) networks, however, such neural network-based architectures generally rely on complex computation and only focus on the single target. We propose a gated hierarchical LSTMs(GH-LSTM) model of combining regional LSTM and sentence-level LSTM via a gated operation for the task of targetbased sentiment analysis. This approach can distinguish different polarities of sentiment of different targets in the same sentence through a regional LSTM, and is able to concentrate on the long dependency of target in the whole sentence via a sentence-level LSTM. The experimental results on multi-domain datasets of two languages from SemEval2016 indicate that, our approach yields better performance than SVM(Support Vector Machine) and several typical neural network models.