Building a Question-Answering System to Extract Information From PDF Files Using BERT Transformers

The comprehension of complex PDFs such as research documents, clinical reports, and scientific manuals is a time-consuming task. Previous studies have demonstrated significant success in building question-answering systems to provide contextually relevant answers to user queries. However, addressing puzzling questions within a single end-to-end trained ML model remains a rigorous task. Such systems require a huge amount of labeled training data to train the base models for specific tasks. The creation of such data sets is still a challenge for complicated documents like the annual reports of big tech companies. This research paper addresses this challenge by focusing on the construction of a question-answering system tailored for PDF files, specifically targeting domains such as finance, bio-medicine, and scientific literature. Curated data sets for the PDF from the chosen Domains were created manually for the evaluation. Pre-trained Bidirectional Encoder Representations from Transformers (BERT) Models from the Hugging Face library were utilized for the chosen domains and evaluated with an F1 score. A score of 44\% was achieved for the BERT Large.