Complaint Classification Using Hybrid-Attention GRU Neural Network

Recently, a growing number of customers tend to complain about the services of different enterprises on the Internet to express their dissatisfaction. The correct classification of complaint texts is fairly important for enterprises to improve the efficiency of transaction processing. However, the existing literature lacks research on complaint texts. Most previous approaches of text classification fail to take advantage of the information of specific characters and negative emotions in complaint texts. Besides, some grammatical and semantic errors caused by violent mood swings of customers are another challenge. To address the problems, a novel model based on hybrid-attention GRU neural network (HATT-GRU) is proposed for complaint classification. The model constructs text vectors at character level, and it is able to extract sentiment features in complaint texts. Then a hybrid-attention mechanism is proposed to learn the importance of each character and sentiment feature, so that the model can focus on the features that contribute more to text classification. Finally, experiments are conducted on two complaint datasets from different industries. Experiments show that our model can achieve state-of-the-art results on both Chinese and English datasets compared to several text classification baselines.