Deep Learning
Deep learning (DL) has changed the course of artificial intelligence (AI). It has solved numerous complex
problems that have plagued the AI society for several years. In fact, deep learning models are more
advanced variants of artificial neural networks (ANNs) with multiple layers, if either linear or non-linear
with different weights that connect each layer to its top and bottom layers. DL models' ability to learn
hierarchical characteristics from various types of data, such as numerical, image, text, and audio, makes
them strong in attempting to solve recognition, regression, semi-supervised, and unsupervised problems
[1]
Deep learning applies a multilayer strategy to the neural network's hidden layers. Classical machine
learning techniques define and extract features either manually or using feature selection methods.
Deep learning models, on the other hand, learn and extract features automatically, resulting in higher
accuracy and performance. In overall, the hyper parameters of classifier models are also automatically
measured. Figure 1 depicts the differences between traditional machine learning (Support Vector
Machine (SVM), Bayesian networks, or decision trees) and deep learning in sentiment polarity
classification. Deep learning and artificial neural networks are currently the best solutions to many
problems in image and speech recognition, and also natural language processing. [2]
Machine Learning
Pre-processing Feature Extraction Modeling
One-hot Naïve
Token Lemma Stem Count SVM
Encode Bayes Positive
Output
TEXT
stop Sentiment
POS NER TF_IDF
Logistic
regression
word
Negative
... ... ...
Deep Learning
Dense
Embedding
Hidden Layer
Pre-processing
Sentiment Positive
Token Lemma Stem
TEXT
POS stop
NER Negative
word
Figure 1 Figure 1 Differences between two classification approaches of sentiment polarity, machine learning (top), and deep
learning (bottom). Part of Speech (POS); Named Entity Recognition (NER); Term Frequency-Inverse Document Frequency (TF-
IDF). [2]
In the following, we will illustrate the stricter of four learning algorithms used to analyze Arabic reviews’
sentiments, starting with RNN ending with BiLSTM: