We propose an algorithm to extract the sentiment from a NL text.
Combining the Neural Networks, characterized by high predictive power and harder interpretation, with more informative models,
allows to predict a sentence sentiment while quantifying it with a numeric value. Using an objective quantity improves the interpretation of the results.
After showing how to properly reduce the dimensionality of the textual data with a data-cleaning phase, we show how to combine: WordEmbedding, K-Means clustering, SentiWordNet, and the Naïve Bayes* model.
We show then how to use such a model for unlabelled data while preserving the interpretability and good performance.