Information was retrieved from patent documents using a robust framework centered on the advanced DeBERTa model in this study. Emphasizing the crucial interplay between algorithmic selection and patent language intricacies, the developed method involved four epochs of model training. Significant enhancements were found through the Pearson correlation analysis. The proposed novel method improved model understanding. Cross-validation was conducted to ensure a balanced evaluation, and the contemplation of an ensemble approach was used for the model's refinement. String-matching algorithms validated the analysis of this research showing the model's efficacy in capturing relationships within patents. The research results allowed for algorithmic tailoring and applications in patent search. By fine-tuning, transferring learning exploration, and addressing biases, continual innovation in patent analysis and information retrieval is continued based on multi-modal integration.