Text merging algorithm for Thai captioning system

The increasing number of Deaf and Hard of Hearing individuals has amplified the need for quality captions, especially in real-time captioning. Previous studies have successfully explored various methods to enhance caption quality, including alignment-based and neural network approaches. While alignment-based methods represent a conventional approach, neural network models are inherently more complex yet offer superior performance. However, these neural models demand large datasets, posing challenges for languages such as Thai with limited datasets available. In this study, we propose two captions improvement methods. The Simple2In1, a simple but effective method to improve the quality of Thai captions on a small-scale Thai dataset. This method utilizes a pre-trained mT5 model to generate a single sequence from two sequences from two different captioning systems. The MASs-ATTEN, this method utilizes multiple aligned sequences to improve the quality of the first method further. Despite a small dataset and limited input sequences, our methods show potential in improving six evaluation metrics, surpassing all baseline models.