Semantic models for preconceived notions for opinion mining in Arabic -

Abstract

Opinion mining (OM), or sentiment analysis (SA), is the process of having computing machines automatically understand and interpret text to identify opinions expressed on certain subjects. OM has become interesting given the abundance of user-generated opinionated data on the Web, especially on social media websites (Facebook, Twitter, LinkedIn, etc.). OM has significant applications in Politics, Social Media, and Business. Providing insights into the public opinion shapes critical decisions in these fields such as influencing voters' directions in the currently active US elections, or having a business decide on a new product launch. OM belongs to a multi-disciplinary research intersecting the fields of machine learning, psychology, social media, and natural language processing (NLP), with NLP presenting the most significant challenge in having machines automatically understand the semantics in text. In this thesis, we explore solutions for opinion mining in Arabic (OMA) due to Arabic's importance as the 5th most-spoken language worldwide, and that recently became a key source of internet content with a 6,600percent growth in number of users compared to the year 2000. OMA's challenges include Arabic's lexical sparsity and ambiguity due to its rich morphology, where Arabic words are packed with significant amounts of information through complex concatenative and inflectional systems. Arabic has also a wide range of dialectal variations, as well as ambiguity caused by optional diacritization in Arabic scripts. Additionally, Arabic suffers from the lack of reliable large-scale sentiment lexical resources that can help training and evaluating accurate machine learning models. To address these challenges, we present solutions inspired from Psychology and Neuroscience. We present a meta-framework to automate the human cognitive processes while reading and interpreting sentiment. Furthermore, for the inference step, we develop new deep learning methods for OMA, inspired from neuroscience and state-of-the-art neu