PERFORMANCE EVALUATION OF AN ADOPTED SENTIMENT ANALYSIS MODEL FOR ARABIC COMMENTS FROM THE FACEBOOK

摘要

Nowadays, the resources of social media are important for sharing data, news, and opinions. Users of social media can write their tweets, posts, and comments to express their feedback about some services and products. Sentiment analysis is one of the approaches for analyzing users` opinions to extract useful information. This research work analyzes and investigates a sentiment analysis model. The model contains four phases mainly: document/dataset collection, preprocessing operations, scoring and sentiment classification, and evaluation. The dataset collection is concerned with collecting Arabic documents or comments from social media like Facebook. The preprocessing operations involve tokenization, rejection of stopwords, normalization, and stemming. Scoring and sentiment classification are concerned with many important themes mainly: checking negation, handling intensifiers, identifying emotions and sentiment classification. The evaluation phase evaluates the performance of the sentiment analysis model. Moreover, the sentiment analysis model is supported by a set of Arabic lexical resources such as list of Arabic stopwords, list of positive and negative emotions, list of positive and negative modifiers, list of affixes of the light stemmer, and others. The sentiment analysis model helps classifying the users` comments to either positive or negative or neutral sentiments (Sentiment Polarity). The adopted sentiment analysis model is presented to identify sentiments in the Modern Standard Arabic (MSA). The model also can investigate and identify sentiments in informal Arabic (colloquial) where most of social media users are using. Some measurable criteria such as precision, recall, accuracy, and error-rate are adopted to evaluate the performance of the sentiment analysis model. Several experiments are done adopting three important themes of Arabic words mainly: negations, emotions, and intensifiers. The model behavior is changed and affected by using such themes either individually or combined. The model performance is also affected by using the type of Arabic sentence and Arabic language style. Finally, the sentiment analysis model behaves well and presents good accuracy values. The accuracy values of the predicted positive comments are 98.2%, 91.8%, and 85.8% while the values are 93.2%, 92.6%, and 70.1% for the negative comments respectively for MSA, Mixed Arabic, and informal Arabic styles.