notesum.ai

Published at November 22

Information Extraction from Heterogenous Documents without Ground Truth Labels using Synthetic Label Generation and Knowledge Distillation

cs.CL

Released Date: November 22, 2024

Authors: Aniket Bhattacharyya¹, Anurag Tripathi¹

Aff.: ¹Amazon

Arxiv: http://arxiv.org/abs/2411.14957v1

Refer to caption

Model	Merchant name	Amount	Date	CORD, M=1	CORD, M=24	Inference speed	Annual cost
Sonnet Zero-Shot (TAIL)	50%	70%	76%	94%	93%	0.3	$1.25X
LLaVA zero shot TAIL prompt	24%	28%	58%	48%	42%	1.5	$0.13X
LLaVA-noisy labels	46%	63%	68%	-	-	1.5	$0.13X
LLaVA-exact labels	-	-	-	82%	-	1.5	0.13X
LayoutLMV3-Tail	41%	58%	-	42%	-	0.8	0.2X
LLaVA-Net	52%	70%	83%	84%	69%	1.5	$0.2X