COMPREHENSIVE STUDY OF SEQUENCE MODELS FOR NEXT-WORD PREDICTION (LSTM, GRU, AND BILSTM ON MULTI-DOMAIN TECHNICAL TEXT)

Abstract

Next-word prediction is a fundamental challenge in Natural Language Processing (NLP), powering intelligent writing assistants and conversational AI. Although large language models dominate this field, lightweight sequence-based neural approaches remain attractive because of their efficiency and interpretability. This study presents a comparative analysis of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (BiLSTM) models for next-word prediction using a curated technical text corpus spanning multiple domains, including machine learning, artificial intelligence, quantum computing, healthcare, and robotics. Experiments show that BiLSTM achieves the highest accuracy (88.93%) and lowest perplexity (3.99), outperforming vanilla LSTM (84.27%,4.30) and GRU (79.61%, 5.06) under identical settings. This study offers a reproducible benchmark and analysis of RNN-family architectures for next-word prediction in technical domains, highlighting their viability in focused deployment scenarios.