Boubaker, IkrameAssas, AimeneLaboudi, Zakaria2022-11-142022-11-142022http://hdl.handle.net/123456789/14292In recent years, social media has become an integral part of people's lives, who share their memories, experiences, feelings and interests through it. This resulted in the generation of substantial volumes of data in many formats, in particular the textual form that represents the largest portion of these data. In order to extract valuable information from social media textual contents, several methods for natural language processing tasks have been developed, especially through deep learning techniques. In this respect, various researches ranging from machine learning to deep learning methods have been proposed to study depression detection through twitter data analysis. Although these researches could achieve good results, they usually rely on the Bi-LSTM model since it has proven to be efficient when tested on text processing purposes. Moreover, such approaches do not deploy the trained models, in order to make predictions using other data that were not used during the training phase. In an attempt to overcome these limitations, we propose a well-defined methodology for building efficient models that allow predicting depression through Tweets analysis. The idea consists in putting all variants of the RNN model under test in order to highlight the strengths and weaknesses of each variant. Once the training phase accomplished, we proceed to a deployment phase in which the well-performing variant is used to analyze the sentiment of Twitter users in UK and US during COVID-19 period toward depressive behaviors. The proposal is validated through several experiments and comparisons according to some evaluation metrics. Overall, the obtained results are satisfactory and encouraging.enSentiment analysisDeep learningTwitter dataDepressionOther