Aspect-Based Sentiment Analysis with a Position-Aware Multi-head Attention Network

Aspect-based sentiment analysis aims to predict the sentiment polarities of the given aspects in its context. A gated recurrent unit (GRU) network can effectively obtain the temporal feature of the context, and the multi-head attention (MHA) mechanism can effectively obtain the vertical connection of the context. This paper proposes a position-aware multi-head attention network, which combines a GRU network with MHA. First, the input layer contains position information of aspect terms, word embedding, and aspect embedding. Second, we process the input information separately through the GRU network and MHA to obtain vertical space information of context and time-series information. Finally, we employ an interactive approach to learn aspect terms and context, generating a more efficient representation for aspect and context. Extensive experiments on the dataset of SemEval 2014 demonstrate the effectiveness of our proposed model.