![](https://csdnimg.cn/release/download_crawler_static/15130762/bg5.jpg)
608 Jialin et al.: A New Fine-grain SMS Corpus and Its
Corresponding Classifier Using Probabilistic Topic Model
algorithms (Naive Bayes, SVM, k-NN, etc.). Fifteen results achieved by combinations of
classifiers and tokenizes were shown in their work in order to give a baseline. They also
verified that SVM is the best classifier in SMS spam filtering task. In resent three years,
some representative works appeared [5,6][16][20,21][27]. They paid more attentions on the
new machine learning techniques.
Even though most researchers have been devoting to use or to improve kinds of standard
classifiers for identifying SMS spam. However, the limited length of SMS messages leads to
lack of enough features. These seriously degrade the performance of the traditional
classifiers (such as SVM, K-NN, and Bayes). The traditional classifiers usually use vector
space model (VSM) to represent text which is apt to long text. It is easily to cause
similarity-drift when it is used for SMS messages. Moreover, it is well-known that VSM is
lack of contextual and semantic information.
Latent Semantic Analysis (LSA) was first to bring semantic dimensionality between the
text and words [26]. Then Probabilistic Latent Semantic Analysis (PLSA) is the further
improvement of the LSA [17]. In PLSA, a document is regarded as a mixture of topics, while
a topic is a probability distribution over words. Latent Dirichlet Allocation (LDA) was
proposed in the first time [5], which added Dirichlet priors in the distributions. LDA is a
more completely generative model and achieves great successes in text mining and other
artificial intelligence domains [4]. Many LDA variations have been generated in recent years,
especially using for social media mining [18]. In order to alleviate sparse problem, some
researchers utilize external knowledge base or corpuses to augment the information [29]. In
addition, Author-Topic Model (ATM) [24] and Twitter-LDA are famous models which take
different strategies to aggregate the short texts into long texts [30]. Although many
researchers are eager to use topic model for social media mining, only Modupe, Olugbara,
and Ojo (2013) directly used Author-LDA it in SMS spam filtering in their work. This is
most related to our work, but author information is usually missing in SMS dataset.
Therefore, this approach has limitation and isn’t used commonly.
3. The New SMS Corpus
Reliable dataset is significant in any scientific research. However, in the studies of mobile
spam filtering, there are fewer corpuses available. Because of legal and privacy reasons, it is
difficult to collect SMS messages from mobile phone users or mobile network operators.
3.1 Collection of the new SMS corpus
In order to fill the gap of the scarce dataset in study of SMS spam filtering and benefit all
the researchers in this field, we have collected a new fine categorized SMS corpus
4
which is
unique and the largest one as far as we know. The new SMS corpus is a set of tagged
Chinese real messages. It contains 8,874 legitimate and 4,204 illegitimate, a total of 13,078
SMS messages. In particularly, we had completed manual classifying for the illegitimate