Predicting the sentiment in sentences based on words: An exploratory study on ANEW and ANET

Current practices for sentiment prediction from text mostly involve words-in-a-bag approach that utilize techniques such as support vector machines or naïve Bayes. In this study, ANET (Affective Norms for English Text) sentence ratings of pleasure and arousal are compared with ANEW (Affective Norms for English Words) word ratings using regression and single layer neural networks. The sentences in ANET are decomposed into their words to obtain valence and arousal ratings from ANEW. A stop list is formed for non-words as well as words that are not found in ANEW. Then we studied whether the sentence sentiment reflected in terms of valence and arousal can be predicted from the sentiment of words in the sentence. Using linear regression, we found that approximately 35% of the variance in ANET valence and arousal ratings can be explained by ANEW valence and arousal ratings. Furthermore, Pearson correlation coefficient for ANEW and ANET ratings are similar for both valence and arousal, and close to 0.6. We also trained neural networks to investigate if non-linear approximations improved prediction of sentence sentiments from the constituent words. Out of several feedforward neural network configurations, a network with 200 hidden layer nodes turned out to be capable of identifying sentence sentiments accurately: the words' valence and arousal values explained 88% of the variance in the sentences' valence ratings and 91% of the variance in the sentences' arousal ratings. This preliminary study indicates that a proper choice of neural network might be adequate to estimate sentiments of sentences from sentiments of words.