Research On Tigrigna News Classification

In recent days,the volume of data is growing exponentially due to the advancement of web technologies and the internet.Although such an enormous amount of data is valuable and most of this information is textual documents,it becomes a problem for users to select a meaningful data of their need unless the information is organized and managed properly.One such task to handle this problem is known as text classification(TC).It aims to automatically classify text documents into predefined classes by learning the characteristics of the categories of previously labeled documents,generally using machine learning algorithms.In the literature of TC,research studies are mainly focused on the popular languages such as English,Chinese,Japanese,and many European languages,while there is an ever-increasing need for the low resourced languages such as Tigrigna(a Semitic language).Text classification for Tigrigna text is a challenging task because it is complex in its grammar and morphology,and it has low linguistic resources.The main objective of this research is to eventually get an effective text classification model for the Tigrigna news text using supervised learning techniques.Because of the absence of Tigrigna corpus for text classification,it is crucial to initiate text classification research by developing new corpora.As a result,a new corpus constists of 5036 labeled Tigrigna news articles was developed.The constructed corpus was divided into a training dataset and testing dataset with a split ratio of 80 to 20 respectively.Additionally,Tigrigna has lack of NLP tools and resources for preprocessing text documents.Therefore,during the preprocessing phase,we constructed three algorithms.1)Tigrigna text normalizer(a routine that removes punctuation marks,non-Tigrigna words,and normalizes the cliticized,hyphenated and short written word forms);2)Tigrigna Stopwords;and 3)Tigrigna Stemmer.We made these contributions by exploring and analyzing the grammatical structure of the Tigrigna language.After preprocessing,three various document vector representation strategies were applied to extract features and transform documents to a numeric feature vector in space for the next processing phase.The first vector representation strategy was using the traditional Term Frequency-Inverse Document Frequency(TFIDF).To show how word embeddings can be applied in Tigrigna text classification,a model for Tigrigna words was built by training word2vec on the entire corpus.The resulting word embeddings model was then used to transform each news article into a vector space using the TFIDF weighted averaged word vectors strategy.Similarly,two paragraph embeddings were built by training the two architectures of doc2vec model,namely,the Paragraph Vector-Distributed Bag of Words(PV-DBOW)and Paragraph Vector-Distributed Memory(PV-DM)on the training corpus.We then constructed document embeddings by concatenating the two paragraph vectors using PV-DBOW+PV-DM strategy.The subsequent processing phase,classification task was performed with supervised classifiers(k-nearest neighbors(KNN),support vector machine(SVM),Multi-Layer Perceptron(MLP),and Random Forest(RF)).Each classifier was trained with the three vector representation strategies on the training dataset and evaluated on the test dataset.The experiments showed that SVM with TFIDF weighting scored the highest among the four supervised classifiers with an overall accuracy of 93.65%followed by MLP with an accuracy of 93.45%.The result also pointed out that the nonlinear SVM and MLP showed better performance based on paragraph vectors(PV-DBOW+PV-DM)representation with an average accuracy of 92.16%and 91.07%and based on word2vec(TFIDF weighted averaged word2vec)with an average score of 91.96%and 91.87%respectively.On the other hand,TFIDF yielded the lowest accuracy score of 77.98%by RF.However,the TFIDF weighted averaged word2vec has dramatically improved the performance of RF with an average accuracy measure of 91.87%(enhanced by 13.89%).Moreover,all classifiers performed higher accuracy scores using TFIDF weighted averaged word2vec approach.Therefore,based on these findings,the results of the word embeddings are reliable to use in categorizing Tigrigna text documents that can effectively expand the semantic features for words and documents,which allows us to perform accurate categorization of Tigrigna news texts.