Sentiment Analysis for the Low-Resourced Latinised Arabic "Arabizi"

The expansion of digital communication mediums from private mobile messaging into the public through social media presented an opportunity for the data science research and industry to mine the generated big data for artificial information extraction. A popular information extraction task is sentiment analysis, which aims at extracting polarity opinions, positive, negative, or neutral, from the written natural language. This science helped organisations better understand the public’s opinion towards events, news, public figures, and products. However, sentiment analysis has advanced for the English language ahead of Arabic. While sentiment analysis for Arabic is developing in the literature of Natural Language Processing (NLP), a popular variety of Arabic, Arabizi, has been overlooked for sentiment analysis advancements. Arabizi is an informal transcription of the spoken dialectal Arabic in Latin script used for social texting. It is known to be common among the Arab youth, yet it is overlooked in efforts on Arabic sentiment analysis for its linguistic complexities. As to Arabic, Arabizi is rich in inflectional morphology, but also codeswitched with English or French, and distinctively transcribed without adhering to a standard orthography. The rich morphology, inconsistent orthography, and codeswitching challenges are compounded together to have a multiplied effect on the lexical sparsity of the language, where each Arabizi word becomes eligible to be spelled in many ways, that, in addition to the mixing of other languages within the same textual context. The resulting high degree of lexical sparsity defies the very basics of sentiment analysis, classification of positive and negative words. Arabizi is even faced with a severe shortage of data resources that are required to set out any sentiment analysis approach. In this thesis, we tackle this gap by conducting research on sentiment analysis for Arabizi. We addressed the sparsity challenge by harvesting Arabizi data from multi-lingual social media text using deep learning to build Arabizi resources for sentiment analysis. We developed six new morphologically and orthographically rich Arabizi sentiment lexicons and set the baseline for Arabizi sentiment analysis on social media