Abstract

This work contains hybrid Convolutional Neural Network and Recurrent Neural Network (CNN-RNN)model for cross-domain sentiment analysis using WEBEmo dataset which was presented for text and emoticon data. Data is divided into a 70:30 proportion for Train and Test. The preprocessing stage breaks down text to individual tokens (word/n-grams), vectorizes the tokens into numerical representation using word embedding like: Word2Vec, GloVe and pads to same sequence length in a batch to the Neural Network. The CNN layer applies convolutional filters to preprocessed data to extract local features, and uses ReLU activation and max-pooling to decrease the dimensionality. This model has a similar encoder that converts the input, but then sends it over an RNN layer that uses Long Short-Term Memory (LSTM) units to learn from the sequence. The model is trained iteratively with backpropagation to optimize the weights. Data logging is used, which stores the data during training every 7% of the training data and 3% of the testing data to check the performance. Our experimental results show that the CNN-RNN model outperforms the state-of-the-art approaches, consistently yielding higher accuracy (up to 90.88% at 10% data), precision (up to 89.65%) and F-Score (up to 90.77%) across all but one subset of data. The hybrid model performed well and its strong performance would suggest that combining spatial and temporal features would be helpful in sentiment analysis for a wide array of domains.