CO-ATTENTION NETWORK AND LOW-RANK BILINEAR POOLING FOR ASPECT BASED SENTIMENT ANALYSIS

Aspect Based Sentiment Analysis (ABSA) is an important and challenging task in language understanding. It aims to assign the correct polarity to a given sentence considering the entity on which an opinion is expressed. Extant neural networks usually employ Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), or Attention networks to address the so-called "target-sensitive sentiment" problem, referring to the fact that sentence polarity is decided by aspect information and surrounding contexts jointly. However, those models usually are complicated and will incur tremendous training cost. Instead of using sophisticated sequential networks, we present a novel co-attention based network to capture the correlation between aspect and contexts. We evaluate our model on three public datasets and the results demonstrate a strong evidence of improved accuracy (up to 2.32% absolute improvement) and efficiency (model converges at least 4 times faster).