This paper proposes a deep convolutional neural network called CharSCNN for sentiment analysis of short texts. The network combines character-level and sentence-level information to improve sentiment prediction. It is applied to two corpora: the Stanford Sentiment Treebank (SSTb) for movie reviews and the Stanford Twitter Sentiment corpus (STS) for Twitter messages. For SSTb, CharSCNN achieves 85.7% accuracy in binary classification and 48.3% in fine-grained classification. For STS, it achieves 86.4% accuracy. The network uses two convolutional layers to extract features from words and sentences. The first layer extracts character-level features, while the second layer extracts sentence-level features. Word-level embeddings are learned using unsupervised pre-training, and character-level embeddings are extracted using a convolutional approach. The network is trained using stochastic gradient descent and backpropagation. The results show that the proposed approach outperforms previous methods in sentiment analysis of short texts. The paper also discusses the effectiveness of character-level features in detecting negation and the importance of unsupervised pre-training for improving model performance. The main contributions include the use of convolutional neural networks for character-to-sentence-level feature extraction, the demonstration that a feed-forward neural network can be as effective as RNTN for sentiment analysis, and the achievement of new state-of-the-art results for the SSTb and STS corpora.This paper proposes a deep convolutional neural network called CharSCNN for sentiment analysis of short texts. The network combines character-level and sentence-level information to improve sentiment prediction. It is applied to two corpora: the Stanford Sentiment Treebank (SSTb) for movie reviews and the Stanford Twitter Sentiment corpus (STS) for Twitter messages. For SSTb, CharSCNN achieves 85.7% accuracy in binary classification and 48.3% in fine-grained classification. For STS, it achieves 86.4% accuracy. The network uses two convolutional layers to extract features from words and sentences. The first layer extracts character-level features, while the second layer extracts sentence-level features. Word-level embeddings are learned using unsupervised pre-training, and character-level embeddings are extracted using a convolutional approach. The network is trained using stochastic gradient descent and backpropagation. The results show that the proposed approach outperforms previous methods in sentiment analysis of short texts. The paper also discusses the effectiveness of character-level features in detecting negation and the importance of unsupervised pre-training for improving model performance. The main contributions include the use of convolutional neural networks for character-to-sentence-level feature extraction, the demonstration that a feed-forward neural network can be as effective as RNTN for sentiment analysis, and the achievement of new state-of-the-art results for the SSTb and STS corpora.