Assessing Caption Accuracy through Consensus and Semantic Evaluation

In this article, the authors explore the effectiveness of adversarial attacks on neural machine translation (NMT) systems. They introduce a new metric, ROUGE-A, which measures the ability of generated captions to accurately describe images. The authors also propose a method called SGA, which improves the performance of NMT systems by generating more accurate and diverse captions.
The authors evaluate the effectiveness of SGA using several metrics, including BLEU-4, METEOR, ROUGE-L, CIDEr, and SPICE. They find that SGA outperforms other state-of-the-art methods in these evaluations, indicating its superiority in generating high-quality captions. However, the authors also demonstrate the vulnerability of NMT systems to adversarial attacks, which can significantly degrade their performance.
To address this issue, the authors propose a technique called visual grounding, which involves identifying and locating objects or regions in an image as per language descriptions. They argue that visual grounding is essential for improving the transferability of NMT systems and enhancing their robustness against adversarial attacks.
Overall, the article provides valuable insights into the challenges and limitations of NMT systems and proposes several solutions to enhance their performance and robustness. The authors demonstrate the effectiveness of their proposed methods through rigorous evaluations and demonstrate the potential of NMT systems for improving image captioning accuracy and robustness.