Context and Length Effects in Portuguese Alternative Text Generation dataset

Context and Length Effects in Portuguese Alternative Text Generation This repository contains the experimental materials and results from the study on automatic alternative text generation for images in Brazilian Portuguese using Multimodal Large Language Models (MLLMs). Repository Structure - /htmls — HTML content of the news pages used for contextual extraction.- /images — Images used in the experiment (one per page).- descriptions — Descriptions generated by Gemini and GPT under all evaluated configurations.- scores — Textual and visual similarity scores (BERTScore, CLIPScore, RefCLIPScore). Experimental Setup - 73 journalistic images collected from a Brazilian news portal.- Models evaluated: - Gemini (gemini-2.5-flash-preview-05-20) - GPT (gpt-o4-mini)- Configurations: - With and without HTML context - 15-word limit - 25-word limit - No word limit Each image was evaluated under 12 configurations (2 models × 2 context conditions × 3 length settings). Evaluation Metrics - BERTScore — Semantic similarity with journalist-provided alt text - CLIPScore — Text-image similarity - RefCLIPScore — Hybrid similarity (text + image reference)All metrics were computed using CPU-based processing.