Arabic Text Generation Using GANs Models

Abstract

Text generation, a fundamental task in Deep Learning and Natural Language Processing, involves crafting new outputs based on inputted text or images, with farreaching applications in robotics, assistive technologies, storytelling, and more. The Generative Adversarial Network (GAN), a dual-neural network system, engages in a zero-sum competition where a discriminative network assesses the authenticity of generated data from a generative network. This adversarial training encourages the discriminator to maximize the probability of accurately classifying both real and generated data, while the generator seeks to minimize it, establishing a dynamic equilibrium. Recent advancements in GANs models highlight their efficacy in English text generation. By training on a text dataset, the discriminator learns intricate word and sentence formulations, empowering the generator to produce text in the desired language. This thesis takes a novel approach, applying SeqGAN, TextGAN, and RankGAN models to an Arabic dataset extracted from artistic and cultural news articles, meticulously split into thousands of sentences. The use of TextBox tools, specifically tailored for GANs models, is instrumental in fine-tuning these models for accurate Arabic sentence generation. Rigorous evaluations of each GAN model's output are conducted, paving the way for a comprehensive comparative analysis, for offering nuanced insights and recommendations into their performance.