Identification of domain-relevant patents via weakly supervised deep learning

Abstract

Patent identification is the process of finding patents relevant to a specific technical topic, especially in the early stages of research and development (R&D) projects. Accurately identifying relevant patents helps scientists, researchers, and industry maximize IP value, anticipate challenges, and gain insights into technological trends, competition, and future innovation opportunities. Conventional approaches, like keyword searches or retrieval based on patent classification codes, frequently result in low precision, while machine learning methods demand extensive manual annotation, posing a major bottleneck for domain-specific applications. In this work, we present a deep learning–based approach for domain-specific patent identification, with a focus on the plasma physics and cybersecurity domains. Our methodology employs a weak supervision paradigm to construct a high-quality training dataset by integrating multiple noisy labeling sources, including linguistic patterns, domain heuristics, and expert-defined rules. Using this synthesized training dataset, we fine-tune pre-trained transformer models, systematically optimizing hyperparameters to maximize performance. The resulting models can be deployed as automated patent identification systems tailored to specialized scientific and industrial contexts. We evaluate our models on previously unseen test sets using standard performance metrics. A comprehensive evaluation on unseen test set demonstrates that our approach achieves high accuracy and significantly outperforms a benchmark in-context learning approach based on large language models.