Patent Value Assessment Based on Entity Recognition

With the acceleration of global innovation, patents, regarded as important indicators of technological progress and innovation, have increased exponentially. This rapid increase poses significant challenges in patent assessment, particularly due to high labor costs and the growing complexity of evaluation processes. Existing methods often rely on word vectors to extract textual information from patents; however, these approaches frequently fail to capture critical semantic details, thereby undermining the accuracy of patent value assessments. To address this issue, our study introduces entity recognition techniques to effectively extract and integrate textual features from patent specifications into the assessment framework, enabling more accurate identification of high-value patents. Using patents in the new energy vehicle sector as a case study, we selected representative features across three dimensions, technological, legal, and economic, to construct a comprehensive patent valuation system. Subsequently, seven types of entities were identified, such as invention structure, vehicle components, and effect-related entities. Furthermore, to overcome the limitation of the Conditional Random Field (CRF) model, which does not account for backward label dependencies, we proposed an enhanced Bidirectional CRF (BICRF) model to better recognize these entities in patent abstracts. The extracted entities were then incorporated as additional features into the assessment system. Finally, we evaluated patent value using multiple machine learning models, including support vector machines and random forests. Results revealed that incorporating entity-based features significantly reduced prediction errors in both training and testing datasets, demonstrating their effectiveness in enhancing predictive accuracy and model generalization.