Incorporating Crowdsourced Annotator Distributions into Ensemble Modeling to Improve Classification Trustworthiness for Ancient Greek Papyri

TL;DR

This work tackles ground-truth uncertainty in crowdsourced ancient Greek papyrus classification by combining a novel Normalized Distribution of Annotations (NDA) with a two-branch ResNet ensemble trained under different labeling schemes. The CXE-ResNet uses one-hot labels, while the KLD-ResNet leverages NDA in a KL-divergence loss, and their outputs are fused with a $k$-NN in a stacked generalization framework to improve accuracy beyond either model alone. Entropy analyses quantify classification uncertainty and reveal that misclassifications correlate with higher output entropy, while NDA enables richer signal in the learning process. The approach yields an ensemble accuracy of $95.6\%$, demonstrates the utility of NDA for handling annotator-disagreement, and offers practical guidance for pruning or auditing noisy images in large crowdsourced datasets.

Abstract

Performing classification on noisy, crowdsourced image datasets can prove challenging even for the best neural networks. Two issues which complicate the problem on such datasets are class imbalance and ground-truth uncertainty in labeling. The AL-ALL and AL-PUB datasets - consisting of tightly cropped, individual characters from images of ancient Greek papyri - are strongly affected by both issues. The application of ensemble modeling to such datasets can help identify images where the ground-truth is questionable and quantify the trustworthiness of those samples. As such, we apply stacked generalization consisting of nearly identical ResNets with different loss functions: one utilizing sparse cross-entropy (CXE) and the other Kullback-Liebler Divergence (KLD). Both networks use labels drawn from a crowd-sourced consensus. This consensus is derived from a Normalized Distribution of Annotations (NDA) based on all annotations for a given character in the dataset. For the second network, the KLD is calculated with respect to the NDA. For our ensemble model, we apply a k-nearest neighbors model to the outputs of the CXE and KLD networks. Individually, the ResNet models have approximately 93% accuracy, while the ensemble model achieves an accuracy of > 95%, increasing the classification trustworthiness. We also perform an analysis of the Shannon entropy of the various models' output distributions to measure classification uncertainty. Our results suggest that entropy is useful for predicting model misclassifications.

Figures (12)

• Figure 1: An Example of an Oxyrhynchus Papyrus: Fragment. Note the extensively damaged condition of the manuscript.
• Figure 2: Examples of Character Images from the AL-PUB Dataset
• Figure 3: Characters From Damaged Papyri
• Figure 4: Normalized Distribution of Annotations (NDA) for the image Shown in Figure \ref{['fig:gamma']}.
• Figure 5: AL-PUB Gamma ($\Gamma \gamma$) Example Z_POxy.v0015.n1805.a.01_135790_216_Gamma_5-12.jpg.