Using Word N-Grams as Features in Arabic Text Classification - Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing - page 37

( ) is the probability of class c i ,
(, ) is the joint probability of class c i and the occurrence of term t,
( | ) is the probability of t given c i .

We used Boolean, TF-IDF, and lookup table convolution (LTC) to choose the
way the selected features would be weighted and represented numerically to the
CAL. Their mathematical representations are shown in Equations (5), (6), and (7),
respectively, where is the numerical weighting of selected feature j in text t :

1, ℎ ℎ

0, ℎ ℎ (5)

= ( ) log

())

(6)

(())

()

(7)

=

∑ (())

()

where:
() is a word in the text ,
is the total number of texts in the dataset,
is total number of words in the text,
() is the frequency of the word in the text and () is the number of
texts that the word occurs in.

An n × m matrix is then constructed, where n is the number of features, and m
is the number of texts in the training dataset. Each cell in this matrix is the weight
of feature j in text t . The selected features from the training dataset are then
extracted from that dataset and represented in the same way as they were in the
training dataset.

Table 2 Experimental parameters

Parameters Description

Dataset SPA: 6 classes, 1,526 texts

Training size 70%

Preprocessing Removing Arabic diacritics, numbers, Latin

characters, and stop list words; normalizing
Hamza and Taa Marbutah

Testing Size 30%

FTs Single word, 2-gram, 3-gram, 4-gram

FS methods DF, CHI, IG, GSS

No. of Terms Selected High-ranked terms (50, 100, 150, 200)

Threshold Minimum DF = 10

FR schemas Boolean, TF-IDF, LTC

CALs NB, KNN, SVM

Number of experiments 576