We performed sentiment analysis on article citation sentences corpora bearing three polarities viz. positive, negative, and neutral. Due to scarcity of negative citation sentences, the dataset suffers from a huge class imbalance issue. To tackle this, we proposed an ensemble feature engineering method for deep learning, which uses embedding of text and its dependency relationships. The performance of deep learning models was compared with a support vector machine and logistic regression approach using bag of words. Experimental results show that deep learning can be used effectively for an imbalanced dataset by applying the proposed ensemble features. Statistical significance test indicates that one-hot supervised LSTM is statistically not different from the baseline methods for two datasets, one developed by us and the other taken from literature.