Extracting a human feeling from a text (a natural language processing task called Sentiment Analysis) using a recurrent neural network together with Keras library

In this paper, it is proposed to understand how the computer is able to extract a simple human
feeling of "liked" or "disliked" from a text. Basically the computer will learn to correctly place a movie
review in one of the two categories of positive or negative. We’ll see how, starting with input values and
output values called labels, the computer begins to learn and correctly recognize the output value (in this case
the 0 or 1 digit, zero representing a negative feeling and the one a positive feeling) through a model built on
the technique called supervised learning. So the proposed objective is to guess the human feeling (translated
by the number 0 or number 1) which is in fact the output value of the model, at a new value of the input, once
this model has been known. In this exercise we will use Keras API built on TensorFlow, a set of movie
reviews taken from IMDB and a recurring neural network RNN with LSTM (Long-Short Term Memory)
cells to preserve the memory of the words that were previously encountered. Keras comes with a set of
50,000 movie reviews that were already pre-processed (this will be explained below). By feeding the neural
network with these tens of thousands of texts (25,000 texts for training followed by another 25,000 texts for
test), the model built by Keras (using relationships of the words), manages to guess with a good accuracy, the
positive or negative human feeling, in other words the polarity of the text. The applications for sentiment analysis are endless starting from social media monitoring and VOC, tweets and facebook posts analyzes, to
the business analysis by text analysis.