Neuro-Evolution-Based Language Model for Text Generation

In the dynamic field of natural language processing, the enhancement of text generation models presents a complex challenge, compounded by the intricate architectures and substantial parameters of contemporary neural networks. This study introduces a groundbreaking method that applies Genetic Algorithms to evolve the architecture of Long Short-Term Memory networks (LSTM), specifically tailored for text generation tasks. Our approach employs a sophisticated gene encoding mechanism that captures the diverse LSTM network configurations and the optimal depth of the network required for generating coherent and contextually relevant text. Our method systematically refines and evolves network architectures through iterative selection, crossover, and mutation processes, uncovering the most effective configurations for text generation. This evolutionary process is designed to yield an LSTM network architecture with enhanced performance in text generation.