Hybrid Fractal Zeiler and Fergus Network for Text Classification Using Large Language Models

Due to the rapid advancement of the Internet, the utilization of text data has exponentially increased in recent times. But it poses various complexities for integrating the strategy of Text Classification for the categorization of assorted sets of texts scientifically. In Natural Language Processing (NLP), text classification has been widely used for public opinion analysis, the retrieval of text content, and the sentimental analysis of online comments. Furthermore, the existing schemes are limited by the potential lack of understanding of the full context of a text, dependency on large amounts of labeled data for training, difficulty in handling noisy or unstructured data, limited interpretability in how results are generated, and the risk of propagating inherent biases from the training data. To bridge this gap, this article develops a Fractal Zeiler and Fergus network (FractalZFNet) for text classification using LLM. Initially, the text data is preprocessed based on stop word removal and stemming and offers an output. Then, the Retrieval-Augmented Generation Large Language Model (RAG-LLM) is utilized to generate text. Further, the preprocessed and RAG-LLM models’ outputs are considered for feature extraction. Then, from the input text data, the word2vec feature is extracted. After that, the obtained feature vectors are given to the duplicate removal process, which is carried out utilizing holo-entropy. After that, feature selection is performed utilizing chord distance and classification of text is carried out using FractalZFNet. Here, the FractalZFNet is modeled by the integration of FractalNet and Zeiler and Fergus network (ZFNet). The evaluation results reveal that the FractalZFNet attained an accuracy of 91.687%, True Positive Rate (TPR) of 90.456%, and True Negative Rate (TNR) of 92.507%.