Improving Opinion-based Question Answering Systems Through Label Error Detection and Overwrite

TL;DR

This paper introduces LEDO, a model-agnostic, efficient framework for detecting and correcting label errors in opinion-based question answering systems, significantly improving model accuracy without complex computations.

Contribution

The paper presents LEDO, a novel, computationally efficient, and generalizable method for label error detection and correction applicable across multiple tasks and datasets.

Findings

Achieved 1.1% MRR gain in retrieval models

Improved PR AUC by 1.5% in machine reading comprehension

Increased Average Precision by 0.9% in ranking models

Abstract

Label error is a ubiquitous problem in annotated data. Large amounts of label error substantially degrades the quality of deep learning models. Existing methods to tackle the label error problem largely focus on the classification task, and either rely on task specific architecture or require non-trivial additional computations, which is undesirable or even unattainable for industry usage. In this paper, we propose LEDO: a model-agnostic and computationally efficient framework for Label Error Detection and Overwrite. LEDO is based on Monte Carlo Dropout combined with uncertainty metrics, and can be easily generalized to multiple tasks and data sets. Applying LEDO to an industry opinion-based question answering system demonstrates it is effective at improving accuracy in all the core models. Specifically, LEDO brings 1.1% MRR gain for the retrieval model, 1.5% PR AUC improvement for the machine reading comprehension model, and 0.9% rise in the Average Precision for the ranker, on top of the strong baselines with a large-scale social media dataset. Importantly, LEDO is computationally efficient compared to methods that require loss function change, and cost-effective as the resulting data can be used in the same continuous training pipeline for production. Further analysis shows that these gains come from an improved decision boundary after cleaning the label errors existed in the training data.