Abstract

This paper describes a novel approach of multitask learning for an end-to-end optimization technique for document classification. The application motivation comes from the need to extract "Situation Frames (SF)" from a document within the context of DARPA’s LORELEI program targeting humanitarian assistance and disaster relief. We show the benefit of our approach for extracting SF: which includes extracting document types and then linking them to entities. We jointly train a hierarchical document classifier and an auto-encoder using a shared word-level bottleneck layers. Our architecture can exploit additional monolingual corpora in addition to labelled data for classification, thus helping it to generalize over a bigger vocabulary. We evaluate these approaches over standard datasets for this task. Our methods show improvements for both document type prediction and entity linking.

Date

May 4, 2020

Authors

Karan Singla, Shrikanth Narayanan

Conference

ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

Pages

8149-8153

Publisher

IEEE