Abstract: Convolutional Neural Networks (CNN) is used to predict the sentiment polarities of English words, and based on these polarities, a method is designed to quantify textual sentiment tendency. First, the fastText tool is utilized to train the word embedding model which transforms words to fixed-length dense word vectors. Then, taking the word vectors as input, one-dimensional CNN model is constructed and multiple model frameworks are also designed with different depths. Finally, CNN model is used to compute the mean of the sentiment polarity of the words in an English text, and the mean would be the quantized value of the text sentiment tendency. Experimental results show that the proposed CNN model improves the accuracy of predicting word sentiment when compared with some traditional machine learning models, and the proposed quantification method of text sentiment also presents a good consistence with the subjective results.