Effective Approaches and Challenges in Hindi-English Neural Machine Translation
Aakrit Singhal · International journal of high school research · 2022
With the rise in the use of Indian languages, visualizing and improving translation tasks is essential in order to narrow the gap between the comprehension of these languages globally.In this paper, approaches of sequence-to-sequence models including attention mechanism and long short-term memory are explored and compared with other state-of-the-art models such as transformers.Previous work has explored multiple Indian languages including Bengali, Hindi, Marathi, Telugu, and Kannada.Due to the overall increase cases of use of the Hindi Language, a comparison between models is done through a custom preprocessing of a Hindi-English dataset and normalizing the data to the required needs.While some systems are focused on specific end-to-end language translations, some are large-scale systems with translation capabilities of multiple languages.The paper proposes that adding attention to the basic encoder-decoder architecture can not only improve the translation tasks with higher accuracies, but also reduce training times and make it possible to train with larger datasets.Additionally, certain challenges with Hindi translation such as word sense disambiguation are also explored by looking at examples of previous work and some of the results produced by them.