Turkish Image-Captioning Benchmark on MS COCO 2014

Progress in image captioning and machine translation depends on large, high-quality labelled datasets - and for Turkish, those barely existed. English benchmarks like MS COCO have driven years of research, but Turkish-language vision-language work had no comparable resource to train and evaluate against.

Closing that gap meant more than machine translation: it required a benchmark of human-verified Turkish captions large enough to support serious research.

At Istanbul Bilgi University I extended the MS COCO 2014 dataset into a benchmark for Turkish image-captioning and machine-translation research. The dataset adds 616,767 human-verified Turkish captions covering 123,287 images, giving the research community a reusable, quality-controlled resource in a language that previously lacked one.

To validate the benchmark, I trained five image-captioning models on it, spanning a range of architectures from CNN+LSTM models using a ResNet backbone through to Meshed-Memory Transformers. Comparing these models on the new data demonstrated that the benchmark supports meaningful evaluation across both classic and state-of-the-art approaches.