Sequential Multi-Kernel Convolutional Recurrent Network for Sentiment Classification

Abstract

The emergence of deep learning as a commanding technique for learning heterogeneous layers of feature representations have consequently substituted traditional machine learning algorithms which are generally poor in analyzing compound sentences. Additionally, convolutional and recurrent neural networks have auspiciously yielded state-of-the-art results in sentiment classification and Natural Language Processing (NLP). In this paper, a deep sentiment representation model through the combination of multiple Convolutional Neural Networks (CNN) kernels with Long Short-Term Memory (LSTM) is proposed for sentiment classification. Our model gains word vector representation using pre-trained Global Vectors for Word Representation (GloVe) embeddings, thereafter used as input to the CNN layer which extracts higher local text representations. Finally, Bidirectional LSTM (biLSTM) generates sentiment classification of sentence representation based on context dependent features. Our combined approach of CNN and biLSTM was experimented using the Stanford Large Movie Review Dataset (IMDB) and Stanford Sentiment Treebank Dataset (SSTB) for binary classification. The evaluation achieves outstanding results in outperforming several existing approaches with 90.4% accuracy on the Stanford Sentiment Treebank dataset and 94.8% accuracy on the Stanford Large Movie Review dataset. These results are achieved with a drastic reduction of model parameters and without a pooling layer in the CNN architecture, helping to retain local and structural information in comparison to other existing deep neural network frameworks. (Abstract)