Request Correction

MennaAly at NakbaArchiveClassifier Shared Task: Transfer Learning with ResNet for Historical Image Classification

This paper describes our submission to the NakbaArchiveClassifier shared task at Nakba-NLP 2026, co-located with LREC 2026. The task consists of binary image classification, where a model must classify historical images into one of two categories: destruction or not_destruction. We adopt a transfer learning approach based on pretrained residual networks, fine-tuned on the provided training data. To mitigate class imbalance, we incorporate weighted cross-entropy loss during optimization. In the development phase, our ResNet18 model achieved a peak macro F1-score of 0.8137 on the validation set. For the final phase, we trained on the combined training and validation data (1,599 labeled images) and generated predictions for the hidden test set of 402 images. Our final submission achieved a macro F1-score of 0.83228 with an accuracy of 0.84577 on the official evaluation set. These results underscore the effectiveness of lightweight transfer learning approaches for historical image analysis under limited-data conditions, demonstrating that compact residual architectures can achieve competitive performance without complex architectural modifications.