Abstract: Our client had ~89k PDFs in its internal systems, over 24k of which were uncategorized. We estimate that time spent searching for unorganized documents has cost our client tens of thousands of dollars per year. Our goal was to utilize machine learning to label as many documents as possible while meeting agreed-upon precision requirements. We deployed a natural language processing (NLP) pipeline that involved text extraction, data preprocessing, feature engineering, and model selection, training, tuning, and inference. We ultimately labeled 71% of our client’s unlabeled documents, versus our target of 72%.
Keywords: natural language processing (NLP), transfer learning, fine-tuning, optical character recognition (OCR), text extraction, classification, transformer.
Watch the team present this project in the session recording here.