Narrow Convolutional Neural Network for Arabic Dialects Polarity Classification

Abstract

The complexities and tangles of Arabic dialect in orthography and morphology typically make the sentimental analysis quite challenging. Moreover, most of the classification approaches have addressed this problem based on hand-crafted features. Since the Arabic language has multi-dialects and the language has no word-based order, the extraction process and the classification tasks are more difficult and time consuming. Deep neural network approaches applied to the Arabic language colloquial are very limited. These deep learning approaches typically comprise a structure that is very complex for small quantities of data. The structures are based on wide convolutional networks that are not capable of capturing the entire semantic and sentiment features for Arabic dialects. In this paper, a narrow structure of the convolutional neural network (CNN) has been proposed in order to obtain the tweets representations and classify the Arabic tweets into five, three and two polarities. Sensitivity analysis has been conducted to evaluate the impact of various combination structural properties, such as the number of convolutional filters, pooling size, and filter size on the classification performances. The proposed Arabic narrow convolutional neural network (NCNN) has captured the entire semantic and sentiment information contained in the tweet by maximizing the features of the detector's range. The NCNN performances were estimated to be at its optimum when structured by three convolutional layers, each one followed by the max pooling layer. The model has been developed without using lexicon resources and lexical features or augmented the dataset with extra training data. The narrow model is the first baseline model for Arabic dialects sentiment classifications for a sentence level as it is the first narrow CNN model addressing the Arabic Dialect tweets. NCNN model achieved the lowest macro average mean absolute error (MAE M ) for five polarity and higher Macro average recall (P) for three and two polarities on the SemEval-2017 Arabic dialect Twitter datasets when compared to the other state-of-the-art approaches.INDEX TERMS Arabic tweets sentiment analysis, multilayers convolutional neural network, narrow convolutional neural network.

A Transformer-Based Neural Machine Translation Model for Arabic Dialects That Utilizes Subword Units

“…Additionally, the model often handles source and target sentences of variable length. Furthermore, sub-word units approach can be applied on various Arabic NLP tasks such as sentiment analysis [ 51 ] and text summarization.…”

Investigating the impact of pre-processing techniques and pre-trained word embeddings in detecting Arabic health information on social media

“…As for traditional methods, the majority of Arabic works have emphasized some pre-processing techniques, such as stemming, but none of the studies discussed determined the impact of normalizing Arabic letters or removing diacritics. Some claimed these techniques negatively affect the classifier performance [ 44 , 45 ] but did not elaborate on or provide evidence for their assertions. Furthermore, there have been no studies to date on the detection of Arabic health-related tweets on Twitter.…”

“…Kashida, also known as tatweel , is a decorative element in Arabic writing used to justify or stretch the text with a phonetic value [ 88 ]. We found two studies [ 36 , 45 ] that removed Kashida.…”

A Systematic Review for Sentiment Analysis of Arabic Dialect Texts Researches

“…While in [28][29][30] hybrid approach was utilized for Sentiment Analysis of Arabic Dialects in which both machine learning algorithms and lexicon-based approach were used. Deep Learning models as well increasingly utilized recently for Sentiment Analysis of Arabic Dialects as shown in [3,[31][32][33][34][35][36][37]. Moreover, the authors of [38,39] used Deep Learning models and Machine Learning algorithms (classifiers) to compare the Sentiment classification results.…”

“…Figure 4 presented the same finding. Twitter attracted researchers due to many reasons such as the tweets are formed in short sentences, Twitter API that makes it easy to export Web-based tool (lexicon-based approaches or machine Learning approaches) [31] S4 Deep learning and ensemble implementations [10] S5 Hybrid lexicon approach (unsupervised and supervised technique) [32] S6 Deep learning [44] S7 Hybrid approach (machine learning and semantic orientation) [13] S8 SVM classifier and NB classifier [33] S9 Deep learning (CNN and LSTM) [45] S10 Aspect-based sentiment analysis [14] S11 Machine learning algorithms [27] S12 Hybrid model (corpus-based and lexicon-based models) [15] S13 Machine learning algorithms [38] S14 Machine learning algorithms and deep learning (CNN) [39] S15 Machine learning algorithms and deep learning (SVM and RNN) [16] S16 Machine learning algorithms (SVM, MNB, SGD, KNN, LR, PA) [34] S17 Deep learning (CNN) [17] S18 Machine learning algorithms (SVM, NB, DT, KNN) [28] S19 Hybrid approach (lexicon-based and machine learning) [18] S20 Machine learning algorithms [29] S21 Hybrid approach (lexicon-based and machine learning (SVM)) [46] S22 Machine learning algorithms (BNB, MNB, NSVC, LSVC, SGD, RGD, LR) [47] S23 Machine learning algorithms (SVM, NB, KNN, LR, MLP) [35] S24 Deep learning (CNN and LSTM) [36] S25 Deep learning (narrow CNN) [19] S26 Machine learning algorithms (SVM, NB, BNB, MNB, SGD, LR) [30] S27 Hybrid approach (lexicon-based and machine learning (SVM and NB)) [20] S28 Machine learning algorithms (SVM and NB) [26] S29 Hybrid model (corpus-based and lexicon-based models) [37] S30 Deep learning (CNN and LSTM) [21] S31 Machine learning algorithms (SVM, NB, and KNN) [22] S32 Machine learning algorithms (SVM, BNB, MLP) [48] S33 Shallow neural network (syntax-ignorant n-grams embeddi...…”