<p>Automatic categorization of textual documents into predefined topics is an important natural language processing (NLP) task. This work proposed a strong framework for topic classification, with many independent methods with the goal of potentially achieving the best results for topic classification. The first step in the system is data pre-processing. In this phase, pre-processing takes place, which includes converting file formats, removing stop words to reduce noise, and text normalization. Then, it was possible to vectorize the text for feature extraction using the Term Frequency-Inverse Document Frequency (TF-IDF) method, changing from unstructured documents to structured numerical documents, as well as capturing the importance of terms with respect to all documents. The core classification model is a Bidirectional Gated Recurrent Unit (BiGRU) neural network. This architecture is selected not only because it processes text data in forward and backward directions, but also because it extracts richer contextual information than unidirectional approaches. To enhance the BiGRU architecture, a new approach called the Modified Builder Optimization Algorithm (MBOA) is proposed. The MBOA was a new metaheuristic algorithm developed to optimize the model hyperparameters. The performance of the proposed framework has been tested on two benchmark datasets, including the BBC News dataset and the larger AG News dataset. The proposed MBOA-optimized BiGRU model demonstrated exceptional performance, establishing a new state-of-the-art on both datasets. On the BBC News dataset, the model achieved an accuracy of 98.62%, a precision of 98.71%, a recall of 98.59%, and an F1-Score of 98.65%. Similarly, on the more diverse AG News dataset, it attained an accuracy of 98.23%, a precision of 98.41%, a remarkable recall of 98.82%, and an F1-Score of 98.61%.</p>