Reference-Based Metrics Are Biased Against Blind and Low-Vision Users’ Image Description Preferences

This is the official Github repository for Kapur & Kreiss (2024): "Reference-Based Metrics Are Biased Against Blind and Low-Vision Users’ Image Description Preferences" published at the NLP for Positive Impact Workshop (NLP4PI) at EMNLP 2024. TODO link Arxiv

The datasets folder contains the context sensitive (context_sensitive.xlsx) and context insensitive (context_insensitive.xlsx) datasets used in this work (as well as the expanded variation of the context sensitive dataset used in Section 4.2 in context_sensitive_expanded.xlsx). This folder also contains the code to generate these datasets (add_correlations.py, dataset_generation.R) from released data in Kreiss et al. (2022): "Context Matters for Image Descriptions for Accessibility: Challenges for Referenceless Evaluation Metrics."

The analyses folder contains all code for correlation calculations and figures in our paper. Figures 1 and 2 were generated from main_with_description_length.R, and Figures 3 and 4 (appendix) were generated from robustness_and_appendix.R

Please email [email protected] with any inquiries.