A Text Sentiment Analysis Model Based on BiLSTM-Conv1D Deep Neural Network

This paper aims to perform sentiment classification and prediction using deep learning techniques, focusing on accurately identifying emotions within text data. Initially, three different models are evaluated for their effectiveness. The model based on BiLSTM is employed to detect emotions by leveraging the bidirectional dependencies present in sequential data. Additionally, the BiLSTM model is enhanced by incorporating pre-trained GloVe word embeddings, which improves the model's ability to capture the underlying semantics of words. Then, the model combining BiLSTM with a one-dimensional convolutional neural network (Conv1D) extracts local features by convolution and enhances the training efficiency. Finally, our experimental results show that classification model combining BiLSTM with Conv1D achieves an optimal balance of training time and accuracy, with a validation set accuracy of 93.92% and a significantly shorter training time, which makes it suitable for efficient sentiment categorization prediction in contexts with limited computational resources.