Story Generation Using GAN, RNN and LSTM

This paper explores the domain of story generation and presents a novel approach that uses Generative Adversarial Networks (GANs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) networks. The objective is to generate realistic and engaging stories for children. The traditional language models are proficient in maintaining grammatical consistency but often fail to establish long-term coherence. This study addresses and explores the individual performances and capabilities of 3 distinct text generation models based on GAN, RNN & LSTM respectively. The project employed three individual models and trained them on the same dataset and the evaluation was conducted using METEOR scores, accuracy, and loss metrics. To address this, our study introduces the use of GANs to enhance the quality of synthetic text. The MaskGAN model gave the highest accuracy and decent output on the trained dataset followed by the RNN and LSTM models. This paper is a significant step forward for story generation, highlights the unique contributions of GANs, RNNs, and LSTMs, amplifies the consistency and quality of independently generated narratives, and provides a foundation for future comparative analyses. MaskGAN achieved the highest accuracy and excelled in generating realistic and high-quality narratives followed by RNNs which exhibited decent accuracy but faced challenges with longer narratives, while the beam search-enhanced LSTM improved narrative quality offering a promising solution for coherent story generation.