Robust PDF Document Conversion using Recurrent Neural Networks

The number of published PDF documents in both the academic and commercial world has increased exponentially recent decades. There is a growing need to make their rich content discoverable information retrieval tools. Achieving high-quality semantic searches demands that document's structural components such as title, section headers, paragraphs, (nested) lists, tables figures (including captions) are properly identified. Unfortunately, format known not conserve because it simply represents document stream low-level printing commands, which one or more characters placed bounding box with particular styling. In this paper, we present novel approach structure recovery using recurrent neural networks process data representation directly, instead relying on visual re-interpretation rendered page, been proposed previous literature. We demonstrate how sequence commands can be used input into network learn classify each command according its function page. This three advantages: First, distinguish among fine-grained labels (typically 10-20 opposed 1-5 methods), results accurate detailed resolution. Second, take account text flow across pages naturally compared methods concatenate sequential pages. Last, our method needs less memory computationally expensive than methods. allows us deploy models production environments at much lower cost. Through extensive architectural search combination advanced feature engineering, were able implement model yields weighted average F1 score 97% 17 distinct labels. best achieved currently served Corpus Conversion Service (CCS), was presented KDD18. enhances capabilities CCS significantly, eliminates for human annotated label ground-truth every unseen layout. proved particularly useful when applied huge corpus articles related COVID-19.