An Entity-Based Main Path Analysis Method to Trace Knowledge Evolution at Micro-Level

Main Path Analysis (MAP) method is a significant method for knowledge flow extraction from citation networks. Traditional MPA methods treat documents as network vertices, while neglecting the more granular information within the document, this neglect limits an in-depth understanding of knowledge development. To remedy the weakness, this study leverages deep learning algorithm on MPA method to facilitate an entity-based pathfinding method, thus to improve the interpretability of MPA method. This study introduces a four-step process to implement the proposed method: (1) Data preprocessing to structure the citation network for analysis. (2) Knowledge entity extraction using BERT-BiLSTM-CRF for identifying significant entities. (3) Main path search at the document level with a cluster-based approach for path identification. (4) Entity relationship identification across documents using a BERT-based model with a three-level masking strategy. This study aims to transform literature-based citation networks into detailed entity-based networks, enabling finer-grained knowledge flow extraction. Finally, to demonstrate the advantages of the new method, extensive experiments are conducted on a patent dataset pertaining to thin film head in computer hardware. Experimental results show that our method is capable of discovering more fine-grained knowledge flows from important sub-fields, and improving the interpretability of candidate paths as well.