Label Confidence Weighted Learning for Target-level Sentence Simplification

TL;DR

This paper introduces Label Confidence Weighted Learning (LCWL), a novel training approach for encoder-decoder models that improves multi-level sentence simplification by incorporating label confidence weights, outperforming existing methods.

Contribution

The paper presents LCWL, a new confidence-weighting scheme for training encoder-decoder models in sentence simplification, demonstrating superior performance over state-of-the-art baselines.

Findings

LCWL outperforms unsupervised baselines on English grade-level simplification.

Fine-tuning LCWL with in-domain data enhances simplification quality.

Combining LCWL with Symmetric Cross Entropy yields better results than supervised methods.

Abstract

Multi-level sentence simplification generates simplified sentences with varying language proficiency levels. We propose Label Confidence Weighted Learning (LCWL), a novel approach that incorporates a label confidence weighting scheme in the training loss of the encoder-decoder model, setting it apart from existing confidence-weighting methods primarily designed for classification. Experimentation on English grade-level simplification dataset shows that LCWL outperforms state-of-the-art unsupervised baselines. Fine-tuning the LCWL model on in-domain data and combining with Symmetric Cross Entropy (SCE) consistently delivers better simplifications compared to strong supervised methods. Our results highlight the effectiveness of label confidence weighting techniques for text simplification tasks with encoder-decoder architectures.