Research on a tailored deep learning method-based information extraction model for engineering document in the advanced electronic industry

With the increasing competition in the market, enterprises in the Advanced Electronic Industry (i.e. advanced PCB manufacturer) need to respond to customer needs more quickly. In the actual enterprise production service business, the customer will form an inquiry document with the production specifications and quality requirements, and the manufacturer shall make product quotations based on the technical information in the customer's inquiry document. Traditionally, inquiry documents need to be obtained by employees from emails or other channels, extract key information, and transform them into internal specification documents. The complex text nested within the documents is often interwoven with heterogeneous tables and figures, creating a highly challenging information extraction scenario. Addressing these pain points, this study proposes an end-to-end information extraction solution. It constructs the first dedicated dataset for the industry, integrating 9,000 real document samples to cover the core scenarios of the research. The solution adopts a hierarchical processing architecture, creates a layout analysis model based on deep learning algorithms to accurately locate information element regions, enables multimodal text recognition through OCR combined with large language models, and designs an adaptive table parsing algorithm to overcome the extraction limitations of traditional solutions. Experimental results demonstrate the effectiveness of this solution, indicating potentials for real-world implementation and deployment in the advanced electronic industry.