Contrastive Training Improves Zero-Shot Classification of

Introduction

• What is the name of the LayoutBERT paper?

  Contrastive Training Improves Zero-Shot Classification of Semi-structured Documents

  AWS AI Labs

• What are the main contributions of the LayoutBERT paper?

  • explore the zero shot classification setting
  • pairwise contrastive objective for pretraining and finetuning
  • unsupervised pretraining with pseudolabels

Method

• LayoutBERT pseudolables are subset of tokens extracted from a document

  

• LayoutBERT uses two encoders, a text encoder and a document encoder that incorporates visual information
• The LayoutBERT architecture encodes text and 2D positional embeddings
  • like LayoutLM with fewer positional embeddings and no visual information
• LayoutBERT finetuning is done with the class labels as the text

Results

• LayoutBERT zero shot performance results

  

• LayoutBERT supervised zero shot performance results: contrastive finetuning does the best

  

Conclusions

This paper provides a simple boost to baseline methods for structured document classification. It’s clear that the benefit of the method comes from the information included in the text descriptions of the labels. I would have liked to see the label content analyzed for the datasets that are evaluated.

Reference

https://arxiv.org/pdf/2210.05613.pdf