Dynamically Generated Question Answering Evidence using Efficient Context-preserving Subdivision

Avi Bleiweiss

2022

Abstract

Recently published datasets for open-domain question answering follow question elicitation from a fairly small snippet of Wikipedia content. Often centered around an article section, the evidence is further subdivided into context-unaware passages of uniform token-lengths to found the basic retrieval units. In this study we hypothesized that splitting a section perceived as an opaque text fragment may hinder quality of answer span predictions. We propose to dynamically draw content corresponding to an article-section url from the most updated online Wikipedia rather than from an archived snapshot. Hence approaching space complexity of O(1), downward from O(n) for a dataset that is fully populated with static context. We then parse the url bound content and feed our neural retriever with a list of paragraph-like html elements that preserve context boundaries naturally. Using knowledge distillation from a sustainable language model pretrained on the large SQuAD 2.0 dataset to the state-of-the-art QuAC domain, shows that during inference our natural context split recovered answer span predictions by 7.5 F1 and 4.1 EM points over a synthetic distribution of fixed-length passages.

Paper Citation

in Harvard Style

Bleiweiss A. (2022). Dynamically Generated Question Answering Evidence using Efficient Context-preserving Subdivision. In Proceedings of the 14th International Conference on Agents and Artificial Intelligence - Volume 3: ICAART, ISBN 978-989-758-547-0, pages 276-284. DOI: 10.5220/0010815700003116

in Bibtex Style

@conference{icaart22,
author={Avi Bleiweiss},
title={Dynamically Generated Question Answering Evidence using Efficient Context-preserving Subdivision},
booktitle={Proceedings of the 14th International Conference on Agents and Artificial Intelligence - Volume 3: ICAART,},
year={2022},
pages={276-284},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0010815700003116},
isbn={978-989-758-547-0},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 14th International Conference on Agents and Artificial Intelligence - Volume 3: ICAART,
TI - Dynamically Generated Question Answering Evidence using Efficient Context-preserving Subdivision
SN - 978-989-758-547-0
AU - Bleiweiss A.
PY - 2022
SP - 276
EP - 284
DO - 10.5220/0010815700003116