Sunday 06 April 2025
Researchers have made a significant breakthrough in developing a new training method for visual question answering (VQA) models, specifically designed for languages with limited resources such as Vietnamese. VQA is a task that involves processing both textual and visual information to answer questions about images. Currently, VQA models are mostly trained on large datasets from English-speaking countries, which can lead to poor performance when applied to other languages.
The new training method, called curriculum learning, involves gradually adjusting the difficulty level of the dataset during training. This is achieved by selecting samples with varying levels of complexity and incorporating them into the model’s training process. The idea is that the model learns to adapt to different types of questions and images, ultimately improving its ability to generalize to unseen data.
The researchers used a combination of raw and augmented textual features to train their VQA model. Augmented samples were created by paraphrasing the original question, which helps the model learn more easily by enriching the input representation. The model was then trained on a dataset specifically designed for Vietnamese VQA, called ViVQA, which consists of 15,000 question-answer pairs derived from images.
The results show that the proposed method significantly improves the performance of baseline models, achieving an accuracy of 55.4% compared to 54.3% achieved by the baseline model. The study also demonstrates the effectiveness of using multiple paraphrased questions to enrich the input representation, which leads to improved generalization and versatility in handling the inherent variability of the Vietnamese language.
One of the key benefits of this approach is that it can be applied to other low-resource languages with minimal modifications. This makes it a promising solution for addressing the lack of high-quality datasets and pre-trained models available for many languages. Furthermore, the use of curriculum learning can help reduce the reliance on manual annotation and human evaluation, making the training process more efficient.
The researchers also explored different epoch update strategies to adjust the threshold value for selecting samples during training. They found that a linear decay strategy, where the threshold value decreases from 1.0 to 0.8 over the training epochs, achieved the best performance.
In addition to its potential applications in VQA, this study highlights the importance of developing effective training methods for low-resource languages. As technology continues to advance and become more widely available, it is essential that we develop solutions that can be applied universally, without relying on large datasets from English-speaking countries.
Cite this article: “Boosting Vietnamese VQA Performance with Curriculum Learning and Paraphrased Embeddings”, The Science Archive, 2025.
Visual Question Answering, Vietnamese Language, Curriculum Learning, Vqa Models, Training Method, Dataset, Image Processing, Natural Language Processing, Machine Learning, Low-Resource Languages.
