This document discusses sentiment analysis and techniques for predicting sentiment intensity as valence-arousal values. It introduces convolutional neural networks (CNNs) for sentiment analysis that represent words as dense vectors and learn relationships between words and sentiment through convolutional and pooling layers. CNN models outperform lexicon-based methods in predicting valence-arousal ratings, with mean squared error rates ranging from 0.61 to 2.25 across datasets. Transfer learning is proposed to improve performance by pre-training a CNN on sentiment classification tasks and fine-tuning it for valence-arousal prediction.