Machine Translation vs AI-Powered Translation

By: Megan Chloe Animashaun, Translation Team Lead

The core purpose of Machine Translation (MT) is to automate the “translation of text from one natural language […] to another.” [1] This automation revolutionised how we transform textual content, increasing the speed of deliveries, the productivity of linguists, and the overall cost-effectiveness of translations. Progress in Machine Learning (ML) and Artificial Intelligence (AI) has since prompted a new evolution in translation technology, causing much debate about the good and the bad.

In this, we will provide you with the lowdown on Machine Translation and Artificial Intelligence so that you can consider what is best for your next translation project.

A Brief History of Machine Translation

The first and most dominant use of MTs was in the late 1970s with Rule-based Machine Translation (RBMT) systems. RBMT analysed the structure and rules of languages using corpora such as dictionaries to produce translations. However, given the complexity and nuances of languages, you can imagine how limited RBMT was!

RBMT was followed by Statistical Machine Translation (SMT) in the 1990s. SMT analysed bilingual corpus to discover patterns and design statistical models. The statistical model was about generating a translation based on the most probable translation for a given source. In 2006, Google Translate emerged as one of the most prominent online SMTs. The rise of Google Translate allowed other online machine translation software competitors to appear in the market, such as Microsoft Translator, in 2009.

Modern Machine Translation and Artificial Intelligence

From the 2010s onwards, the MT that dominated the industry was Neural Machine Translation (NMT). Using the algorithms of Deep Learning NMT’s neural network models “make decisions like the human brain” [2]. These NMTs “require large-scale training corpus” [3], which can undergo fine-tuning to produce accurate and fluent translations. Like the human brain, NMT models undergo ‘continual learning’, meaning if introduced to a new corpus, it will ‘learn’ and memorise its translation for future use.

As research in Machine Learning (ML) has continued and innovative systems have developed, such as GPT-4 and OpenAI, we are now at the dawn of Generative AI (GenAI) Translation. Like NMT models, GenAI Translation models must be trained on a large body of data and benefit from continual learning. However, the main difference between the two is that NMT will generate translations using existing translations in its corpus, whereas GenAI Translation will generate translations from scratch using its corpus.

Can GenAI be used for Translation?

Although GenAI is new to the scene, it can certainly be used for translation. However, it is worth noting that not all types of content are recommended to be translated by an AI-powered system, particularly official documents, branded content, and highly creative content.

For everything else, an AI-powered translation certainly offers a range of benefits:

• Contextual Improvements—As algorithms improve, GenAI translation models are getting better at generating translations that match the language’s nuances and better at recognising the original language’s context.
• Feedback Loops for Continual Learning—AI enhances continual learning, as the algorithms initially used for NMT have improved in recent years. Some AI-powered translations learn not only as new content is added but also offer an AI-powered feedback loop that the AI can use to learn from and fine-tune its dataset.
• Business Benefits—Using AI Translation can offer several business benefits. First, AI can help cut the cost of translation. Second, it increases scalability. Unlike MT, an AI-powered translation can deal with more content and more languages. Lastly, AI can process content much quicker, decreasing translation turnover time.

However, as with anything new to the party, there are some things to be mindful of when opting to use an AI-powered translation:

• Unintentional Plagiarism and Biases – Since AI draws its output from its dataset, it can unintentionally resemble existing content and introduce unintentional biases. Hence why AI is not recommended to be used alone for official, branded, or creative content. However, these issues may be reduced through AI detection tools and fine-tuning.
• SEO Challenges—AI may struggle due to its lack of human-like creativity for content that requires SEO. It is important to remember that AI, for the time being, cannot think outside of its dataset, however expansive it might be.
• Data and Privacy Concerns – Since AI is trained on all types of information and any information you might input, many have raised concerns over protecting data from being stored in AI.

To limit the challenges of an AI-powered translation, you can opt to train your own model with your company’s previous translations so that you are more likely to get the desired output, rather than using a generalised AI. On the other hand, you may be more inclined to use an NMT with AI capabilities. For example, if you wish your NMT enhanced continual learning, you may opt for an NMT that has introduced AI algorithms into their feedback loop.

The ‘Human-In-Loop’

Regardless of whether you choose to go down the Machine Translation approach, AI approach, or both, having a ‘human-in-loop’ is always recommended. This ensures that all translations undergo a quality check by an experienced translator who is native to that language. Some may call this the balancing act between Machine Learning and human expertise. While the assistance of translation technology is forever welcomed, quality is always of utmost importance at Ultimate Languages!

That said, if you have a project in mind and need consultancy on MT and AI for translation, please get in touch! You can speak to one of our experts, who can help guide your project to success.

Get In Touch

[1] Russell, S. and Norvig, P. (2010). Artificial Intelligence A Modern Approach Third Edition. [online] Available at: https://people.engr.tamu.edu/guni/csce421/files/AI_Russell_Norvig.pdf.

[2] IBM (2023). What Are Neural Networks? [online] www.ibm.com. Available at: https://www.ibm.com/topics/neural-networks.

[3] Cao, Y., Wei, H.-R., Chen, B. and Wan, X. (2021). Continual Learning for Neural Machine Translation. [online] pp.3964–3974. Available at: https://aclanthology.org/2021.naacl-main.310.pdf [Accessed 8 Sep