Abstract
Abstract This research proposes a novel end-to-end framework for insurance policy summarization and classification by introducing three methodological enhancements: Enhanced TF-IDF (E-TF-IDF) for discriminative feature extraction, a Modified Botox Optimization Algorithm (M-BOA) for optimizing Bi-LSTM summarization, and an Enhanced Batch Normalization-based LeNet (EBN-LeNet) for robust policy classification. The process begins with preprocessing, where the input insurance policy document undergoes stemming, tokenization, Stopword removal, and keyword identification to refine the text. The preprocessed text is then fed into the feature extraction phase, where relevant features are extracted. A Bi-LSTM (Bi-directional Long Short-Term Memory) model is employed to summarize the text. To enhance its performance, the M-BOA is used to fine-tune the model’s weights under constraints like content coverage and redundancy reduction, which ensure better summarization quality. Next, augmentation is performed through processes like synonym replacement, deletion, swapping, and insertion to enhance the summarized text. From the augmented text, aspect term-based features are extracted and passed to the policy classification phase. Here, classification is performed using an EBN-LeNet model, which accurately categorizes insurance policies.