Iterative Evidence Searching over Long Structured Documents for Question Answering

Anonymous

Abstract: We propose a simple yet effective model, DOCHOPPER, for selecting evidence from long structured documents to answer complex questions. Similar to multi-hop question-answering (QA) systems, at each step, DOCHOPPER iteratively uses a query $q$ to extract information from a document, and combines this information with $q$ to produce the next query. However, in contrast to most previous multi-hop QA systems, DOCHOPPER is able to extract either short or long sections of the document, thus emulating a multi-step process of “navigating” through a long document to answer a question. To enable this novel behavior, DOCHOPPER does not combine document information with $q$ by concatenating text to the text of $q$, but by combining a compact neural representation of $q$ with a compact neural representation of a (potentially large) hierarchical part of the document. We evaluate DOCHOPPER on three different tasks that require reading long structured documents and finding multiple pieces of evidence, and show DOCHOPPER outperforms Transformer models for plain text input. Additionally, DOCHOPPER is efficient at inference time, being 10–250 times faster than baselines.

Paper Type: long