Resources and analytics for opinion mining and recommender systems, with application to Arabic

Abstract

This dissertation aims at exploring artificial intelligence solutions that help humans in their everyday life decisions such as where to stay, which doctors to consult, or which movie to watch. In particular, the dissertation goal is to address challenges and advance systems for recommender systems and opinion mining with special emphasis on Arabic. Opinion mining systems aim at extracting sentiment from data initiated by people such as data published in social networks. Recommender systems on the other hand use other people’s opinions and experiences to provide recommendations for users’ personal decisions. In this dissertation, several challenges are addressed to provide advances in resources and algorithms for recommender systems and opinion mining, in particular for Arabic. The dissertation includes a first of kind comprehensive survey for opinion mining in Arabic with a unique system and deployment perspective covering all relevant components including advances in NLP software tools, lexical and corpora resources, machine learning models, applications, and a roadmap for future development. Furthermore, to advance the field of automated opinion mining in Arabic, the first challenge was to develop lexical resources that are critical for the semantic interpretation of language. This problem is formulated as link prediction with the goal of linking large-scale Arabic lexical resources to previously developed English WordNet resources. Multiple natural language processing techniques are used including lemmatization, stemming, shallow and semantic similarity measures, feature extraction based on machine translation tables, and semantic word embeddings. For recommender systems, the challenges are in addressing sparsity of user-item rating matrix for historical data, and the cold start problem for cases with no relevant history. To address these challenges, we propose two different models using collaborative filtering. The first approach is formulated as an optimization problem that aims at finding the