Towards Annotation Bias in Information Extraction Across Datasets

Anonymous

TL;DR: We conduct three probing experiments to study the annotation bias in IE and propose a multi-stage framework to mitigate it.

Abstract: Annotation bias is a negative phenomenon that can mislead models. However, annotation bias in information extraction appears not only across datasets from different domains but also within datasets sharing the same domain. We identify two types of annotation bias in IE: bias among information extraction datasets and bias between information extraction datasets and instruction tuning datasets. To systematically investigate annotation bias, we conduct three probing experiments to quantitatively analyze it and discover the limitations of unified information extraction and large language models in solving annotation bias. To mitigate annotation bias in information extraction, we propose a multi-stage framework consisting of annotation bias measurement, bias-aware fine-tuning, and task-specific bias mitigation. Experimental results demonstrate the effectiveness of our framework in addressing annotation bias.

Paper Type: long

Research Area: Information Extraction

Contribution Types: Model analysis & interpretability, Data analysis

Languages Studied: English