Asking questions on handwritten document collections

This work addresses the problem of Question Answering (QA) on handwritten document collections. Unlike typical QA and Visual (VQA) formulations where answer is a short text, we aim to locate snippet lies. The proposed approach works without recognizing text in documents. We argue that recognition-free suitable for documents historical collections robust recognition often difficult. At same time, human users, image snippets containing answers act as valid alternative textual answers. uses an off-the-shelf deep embedding network which can project both words word images into common sub-space. bridges visual domains helps us retrieve potentially question. evaluate results two new datasets: (i) HW-SQuAD: synthetic, counterpart SQuAD1.0 dataset (ii) BenthamQA: smaller set pairs defined from popular Bentham manuscripts collection. also present thorough analysis compared recognition-based recognized using OCR. Datasets presented this are available download at docvqa.org.