FAQ 's

Machine Translation systems analyse the input given in the source language and convert it into the target language. They use various rule-based computational and linguistic techniques or other statistical methods. Modern MT systems use neural networks to carry out this process. These methods use the deep learning model where the system has been trained on a particular dataset and uses the learning from these training models to translate the given input. This method is more accurate as it considers a big dataset before translating.

• Rule-Based Machine Translation (RBMT): It utilises linguistic rules and dictionaries created by human experts to carry out translations.
• Statistical Machine Translation(SMT): It utilises statistical methods and extracts data from large corpora of bilingual texts to predict translations.
• Neural Machine Translation (NMT): It utilises neural networks which aim to mimic the functioning of the human brain and deep learning models to generate translation.
• Hybrid Systems: These MT systems utilise elements of RBT, SMT and NMT to augment the strengths of each method.

While Machine Translation systems support multiple languages, the range of each system depends upon the languages they have been trained on. The lack of available resources for each language can hamper the MT systems from expanding. Lesser-known dialects and languages may not be widely available or have lower accuracy due to limited resources.

Machine Translation became a reality in the 1950s with the initial focus on rule-based systems which relied on linguistic rules to carry out translations. It evolved over the decades with the advent of statistical methods in the 1990s and neural machine translation methods in the 2010s. NMT brought significant improvements in translation quality and now represents the state-of-the-art in MT.

MT systems' functionality depends on the number of languages they have been trained in. Since Bhashini was developed specifically for Indian languages, it has been trained on large datasets of Indian languages and can provide more accurate translations than Google and Bing. Furthermore, it can focus on dialectical differences and cultural or contextual understanding not supported by other MT systems.

MT systems may not always produce accurate translations. They often lack common sense which causes them to give grammatically correct yet nonsensical answers at times. Especially for idioms and complex sentences, they may fail to give the correct output. They might also fail to encompass all cultural information in the translation and might produce culturally inappropriate outputs. Furthermore, due to limited resources and training, they might not produce outputs in non-standard dialects.

Machine Learning is a branch of Artificial Intelligence which works through the training of models. A machine is provided with a dataset and is coded to learn the patterns and trends in the data. Using them, it predicts outputs and makes decisions on the testing data. Similarly, machine learning techniques are used on models to train them on large datasets of language data. Using the derived trends and patterns, machine translation systems make predictions of target translation and give output.

Machine translation systems have become better at generating basic translations. However, they still struggle with fully grasping cultural nuances and contexts. Though neural machine translation systems are working towards comprehending contextual codes, they often lack the cultural awareness that human translations possess. As a result, they might produce grammatically correct yet culturally insensitive translations.

The future for Machine Translation systems looks promising with the advancements in technology. One can expect an increase in the accuracy and quality of the translations. Neural machine translation models aim to incorporate more contextual and cultural understanding in the output. Furthermore, one can expect MT systems to expand the number of languages they can translate to and from. Personalisation tools can also be expected which will allow users to generate domain-specific translations.