PERSONALIZED CONTENT RECOMMENDATION SYSTEM FOR ARABIC REVIEWS USING DEEP LEARNING AND REINFORCEMENT LEARNING

Abstract

This study presents a hybrid, personalized content-level recommendation framework tailored to the Arabic language reviews in hotels thus solving the twofold problem of morphological richness and dialectal difference of the language. The aim of the present study is to enhance recommendation accuracy by integrating between Aspect-Based Sentiment Analysis (ABSA), Sequential deep learning and reinforcement learning in dynamic Personalization framework. Attentive classifier in the form of gated recurrent unit (GRU) is suggested to filter sentiment analysis based on the finest-grained level of granularity. Sentiment vectors outputted are provided to a Deep Q-Learning (DQL) agent where the recommendation policy would be iteratively adjusted based on user feedback. Systematic empirical assessment of a sentence level sentiment classification model was performed on the SemEval 2016 Task 5 Arabic, which has 2,291 annotated hotel reviews. The accuracy of the classifier reached 86.5% and its AUC rate of 0.91, which was enough to confirm effective discrimination against the polarity regardless of the linguistic peculiarities of the language of input. Precision (87.1%) and recall (86.6%) were observed to be high using a confusion matrix. The explicit reinforcement learning metrics were not presented, but the qualitative review allowed concluding that the deep Q-Network (DQN) agent was able to correct its recommendations in respect of user sentiment patterns. The current study proposes a combined model that further develops the modern Arabic recommender system by integrating the linguistic-awareness methods with real time personalization approaches and thus develops a solid research framework in this regard that can be used in the future to study the Arabic-language adaptive recommendation of content framework.