Information Technology Reference
In-Depth Information
Table 1
SPA dataset basic statistics
News
Classes
No.
of
Texts
No. of Unique
Words
2-grams
3-grams
4-grams
Cultural
258
12,521
33,499
40,311
42,137
Economic
250
9,158
26,305
32,363
34,184
General
255
12,639
32,336
38,384
39,976
Political
250
9,720
26,667
32,136
33,278
Social
258
11,818
33,914
42,279
44,999
Sports
255
8,641
24,425
31,339
33,885
Total
1,526
36,497
154,240
208,390
224,903
2.2
Experimental Setup
As mentioned in Section 1, several consecutive steps must be followed in creating
the classification model. For data preprocessing, we removed stop words, which
consist primarily of function words of Arabic, such as prepositions and pronouns
[5]. In addition, we removed numbers, Latin characters, and diacritics and
normalized different forms of Hamza (إ ،أ) to (ا) and Taa Marbutah (ة) to (ﻩ). We
randomly chose 70% of the data for training and the remaining 30% for testing.
We selected single words, 2-grams, 3-grams, and 4-grams as features and chose
four of the most used FS methods: DF, chi-squared (CHI), information gain (IG),
and Galavotti, Sebastiani, Simi (GSS). The objective of FS is to order the features
according to their importance to a given class, with the most appropriate features
that distinguish the class from the other classes in the training dataset in the
highest ranks. The highest-ranked features are then selected. This selection
process yields reduced feature spaces, positively affecting CA and speed. The
mathematical representations of DF, CHI, IG, and GSS are shown in Equations
(1), (2), (3), and (4), respectively.
(,
)=(,)
(1)
,
=
|
|⋅ [
,
⋅
,
,
⋅
(
,
)]
(
)⋅
⋅
⋅(
)
(2)
(
,
)
=
∑
(
,
)
.
(,)
().()
+
∑
,
(
,
)
(3)
.()
(,
)=(,
).,
−.(,
).,
(4)
where:
m
is the number of classes,
Tr
is the total number of texts in the training dataset,
c
denotes class,
t
is a term, which could be a 1-, 2-, 3-, or 4-word level gram,
(,
)
is the number of documents in class
c
i
that contain term
t
at least once,
Search WWH ::
Custom Search