Machine translation (MT) is used for rapid processing of large volumes of texts. It does not simply translate texts but also provides a means of conveying a perfectly resonating message for the intended audience. Just like a chef at a renowned restaurant carefully selects ingredients and techniques to suit each guest's palate, machine translation must meet the requirements and the cultural nuances of the target audience.
Today's technologies enable machine translation to be adjusted to fit people’s specific needs through customization. In this article, we will take a closer look at the process of customizing MT engines and how businesses can benefit from it.
What are the reasons Machine Translation (MT) may sometimes fail to produce satisfactory results?
Machine translation is effective for general text processing . However, when presented with specialized texts that contain technical, legal, or economic terms, the translation is likely to be inaccurate. This is because the same word can have multiple meanings in different contexts. For example, the verb ‘to crack’ in the IT field may be translated into Russian as ‘to hack’, and the noun ‘bug’ could refer to an insect or a coding error. Machine translation struggles to capture the meaning embedded in a specific situation. This is a significant challenge for companies, since they frequently concentrate their activities within specific domains and use their own terminology. This is why the process of machine translation customization is necessary.
The Evolution of MT Customization
To truly appreciate the benefits of custom machine translation , it's important to know how it has changed over time. In the past, creating custom neural MT engines needed a lot of resources and technical skills, which meant that companies either had to spend a lot of money or depend on external partners.
In 2017, MT providers began making customization more accessible to language enthusiasts and developers. A crucial moment came in 2018 when Google introduced AutoML, aimed at democratizing the customizing process. Google CEO Sundar Pichai emphasized that AutoML would enable a wider range of developers to design tailored neural networks.
Today, the situation has changed. There are many customizable machine translation (MT) engines available, along with basic ones that allow some customization. This makes MT solutions easier for users to access.
Machine Translation Customization defined
Machine translation customization is a process of adapting machine translation engines to meet specific user needs, contexts, and preferences. It helps to improve the quality of translation in specialized fields or for specific tasks, making the translation more precise, relevant, and tailored to the needs of its users.
Let's draw an analogy. General machine translation is like a student at an economics university. They have an understanding of the field, can perform basic tasks, but lack deep comprehension of all processes. Custom machine translation is like an experienced business analyst, capable of adapting to various situations and client requirements. This is a more accurate and efficient way of translation.
What does MT Customization require?
MT customization involves several key components:
- First, the company provides its own glossary of terms along with their translations.
- Secondly, a Do Not Translate (DNT) list is required. This list must include the company's name, names of products or services, and words that may have different interpretations depending on the region.
The customization process requires the submission of a list of terms that instructs the MT engine on how to translate them accurately or whether to avoid translation entirely, thus reducing the necessity for additional editing.
Machine Translation (MT) training is a more advanced level of MT customization
While customizing machine translation models focuses on enhancing an existing model, MT training involves creating a new one and training it based on specific settings. Successful training of machine translation engines requires providing at least 15,000 high-quality unique bilingual segments. It’s a significantly more expensive process that can be beneficial in the long run, although it may not be suitable for every company.
Two types of data are used for training the MT engine: extensive linguistic corpora and translation memory.
Linguistic corpora consists of a multitude of specially prepared texts presented in two or more languages. They can cover various genres and styles , from literary works to colloquial speech.
Translation memory (TM) is a repository where texts are stored in segments, containing both the original and target language versions. The software continually compares new segments with those already stored in memory and suggests using existing translations.
How long does the MT training process take?
The model is trained on the same data multiple times, with each iteration using the results of the previous step to enhance the next one. The process itself can take from several days to several weeks. After completing the training, the model's quality is evaluated using various metrics and also manually by experts. This helps determine how effectively the model performs its task and whether it requires further tuning. The success of the training depends on the settings, the quality and diversity of the training data, and the quality of managing the training process.
Which option to choose: MT customization or MT training?
Let's take an example: you need a machine translation system for legal documents. In this case there are two options:
MT customization: You're starting with a pre-trained machine translation system that already has some translation capabilities. Customization involves adapting it to legal documents. You can adjust the model settings, add a legal dictionary, and include a list of your own terms to improve the accuracy of the results. Since you're working with an existing model, customization can be significantly faster than training a whole new system.
MT training: You're essentially teaching the MT system a new language pair. This requires a massive amount of data, it can be expensive and time-consuming. Training involves complex algorithms that look at the data and learn how to translate better. This needs powerful computers like strong GPUs, which use a lot of electricity. Figuring out the best way to train, including the setup and settings, takes a lot of trying different things and tweaking them. It's a job for experts and can take a long time.
In simpler terms, MT training is like building a whole new house, while MT customization is like renovating an existing one. The choice depends on your goals and resources. If your company lacks sufficient data for training, human and financial resources it's better to opt for MT customization. Ongoing costs for maintaining the glossary over time usually come out cheaper than the expenses associated with MT training.
Machine Translation Customization: A Strategic Choice for Businesses (Conclusion)
MT customization is an essential tool that will enable companies to obtain more accurate and relevant translations, thereby enhancing communication with clients, increasing operational efficiency, and improving the overall company image. The evolution of MT technology has made customization more accessible, allowing businesses to choose between adapting existing systems or investing in new training processes.
However, it is essential to realistically assess the company's capabilities when choosing between the two options. Because any imprudent innovation may not only fail to make a profit but also lead to losses. Ultimately, the decision between MT customization and training depends on a company’s specific goals, available data, and budget.
