Adaptable machine translation: what it is and why it matters

Heather Rossi 13 Mar 2023 6 mins
Adaptable mt language weaver rws
Maria is a chemicals manufacturing researcher working on a variety of interesting projects that require her to communicate with colleagues in different countries. She often uses free online translation tools for her research projects, but as her work becomes more technical and confidential, she has concerns about the accuracy and security of these solutions. While these tools have improved in quality over the years, most store content as data to improve their models, making them unsuitable for confidential or sensitive information. 
Additionally, even in the presence of a “user feedback” function, the machine translation models in these solutions are designed to perform well across most content types, which means that they are not built to handle specialized content. Maria needs to consider alternative solutions, due to the nature of her work. And it’s not just Maria…organizations and enterprises have these same considerations in mind for their employees who encounter multilingual data in their work, too. Security requirements can be met by choosing a machine translation provider that encrypts data, has flexible deployment models, and/or doesn’t mine your data, but what about output quality and adaptability?

What is adaptable MT and how does it work?

Most modern MT providers use Neural Machine Translation (NMT) which uses deep learning to translate content from one language to another. It’s a type of artificial intelligence that learns to predict the correct output sentence given the input sentence by adjusting its internal parameters; this process allows the network to gradually improve its translation accuracy as it is exposed to more and more training data. This means that NMT is inherently adaptable, but the degree to which the end user or the end organization has control over adaptation will depend on both your MT provider and what data you have available to you in order to customize the output. Here are some of the common mechanisms available on the market:
  • Customization as a service or self-service. Some MT providers will create custom models for their clients as a service. In this process, a data scientist will typically create a custom NMT model that has been trained with large bilingual data sets curated specifically for a customer or an industry. More recently, some MT platforms allow for businesses to take control of customization themselves, through a variety of self-service tools. Custom models created by the MT provider means that adaptation was performed by experts, while self-service gives you more direct control over the output, assuming you have the bilingual data already at hand (typically in the form of a Translation Memory). This control will be especially relevant if you have content you simply cannot release to a 3rd party.
  • Glossaries and feedback. The most thorough way to adapt MT is to engage the AI's deep-learning capabilities through training with large data sets, but not everyone has access to the bilingual data required. Another option is to use smaller data sets such as a glossary or approved revisions to the MT's output by your employees. A glossary (also known as a termlist, terminology or dictionary) is a defined, matched set of terms or phrases that should always be translated a certain way. Company or industry specific acronyms are great examples of glossary entries. These glossaries, or dictionaries, can be applied on top of the NMT output to instantly change content without any additional user intervention. If during the process of translation, users have feedback that they want to see implemented on future translations, some MT providers allow organizations or end users to submit these edits to improve the model. 
  • User-initiated or continuous adaptation. As an end user or an organization that has well-structured bilingual data sets, you should consider how you want the models to learn from new data. Will you always want an administrator or a linguist to oversee the model retraining or do you want the models to update themselves with new data sets as they are added? There are benefits to both: continuous adaptation is a great option for organizations that have well-established workflows for capturing good clean bilingual data (e.g., most localization or translation vendors). Once you set the structure of what models should be updated with which Translation Memories, then the NMT model will organically train as you translate more content, creating an automated virtuous quality circle. But for end users who don’t have these workflows set, are just learning about adaptation, or want to continually test out model quality based on varying input parameters, then you would want to have users initiate the adaptation each time.
The combinations available to you will ultimately depend on your MT solution.

Why might you want adaptable MT?

For some uses of MT, a lack of adaptability doesn't matter. For example, many businesses first look at buying an enterprise MT solution because they want a secure option for a variety of internal translation requirements, like Maria’s organization. Perhaps a paralegal needs to get a quick sense of whether certain foreign-language documents are relevant to a case. Or employees whose native language isn't the company's working language want to translate policies or internal communications to understand them better. For these cases, what matters is that the baseline MT quality is good enough, and that the solution offers the security and deployment flexibility that the business needs. Adaptability may be only a 'nice to have.'
But there are scenarios where the ability to adapt your MT solution to your particular needs would be a significant advantage, specifically:
  • When translating for an external audience (with or without human post-editing). Without any adaptation, the right MT solution could be appropriate for a variety of customer-facing translation use cases. But there will typically be many instances where you'll want to enhance the baseline MT quality to reflect your company's preferred terminology (including what not to translate), as well as your style and tone of voice. By adapting MT to create your own custom models, you can also enhance the productivity of post-editors – human translators who review MT output and amend it if necessary – if you're using them as part of your translation process.
  • When you need precise translation of very specific content. What do you do if you need high accuracy in translation and the content in question is too specific for your generic MT models? For example, what if you need a nuanced understanding of legal, scientific or financial content with very specific terminology? Adaptable MT will give you the ability to more accurately translate this content with your existing SMEs and datasets.
This is why, even if adaptability doesn't seem important for your initial use case when buying an MT solution, it's still worth considering. Knowing that your chosen solution has good adaptability options gives you the flexibility to expand your use of MT with greater confidence in future.

Adaptable MT models aren't all created equal

We've already seen that adaptable MT solutions may differ in the level of adaptability they offer, so you'll want to look for a solution that offers you the approaches you prefer. Are you happy with customization-as-a-service or do you want self-service options? If the latter, do you think you'll always want adaptation to be manually initiated or might you want to automate the process? What sources of adaptation are important to your organization, so you can be sure you can use them all?
Security is another key consideration – especially since it's usually the starting point for enterprise MT investments. Be sure to engage your organization's security teams and complete a security review of the MT vendor prior to sending them sensitive data for customization, or uploading such data to their self-service adaptability platform.
Self-service capabilities can also differ in important ways beyond those already covered:

  • Adaptation for individual or group. Some self-service adaptable MT solutions can only adapt to single users – which is ideal for individual translators, but probably not optimal for enterprise adaptation. Say, for example, Adam and Annie are colleagues who both have important input on how their business is translating some content. Adaptable MT designed for the individual translator will not include Adam’s feedback into Annie’s future translations. Translation teams or translators may like keeping their edits in a closed environment but most enterprise organizations will want an adaptable MT solution that can facilitate feedback from anyone and – once input is approved – reflect it back to all users. That way, the whole organization benefits from the contributions of everyone, and gets more consistent and relevant MT as a result.
  • Domain flexibility. Can you create as many domain-specific models as you want, so that you can (for example) let different sub-brands, business units or departments within your organization reflect their own needs?
  • Training data caps. Is there a limit to how many translation units you can use for MT training? If there is, is it large enough to cater to the volume you have available, or will it severely limit your ability to leverage all your available translation assets?
  • Ease and speed of use. How user-friendly are the adaptability features? The best self-service solutions empower users with an intuitive user interface for the options they support, whether that be applying overlays, training the model, rolling out each custom iteration (domain and language-pair), or automating the process.
As long as your MT solution gives you the opportunity to create custom models for your chosen domains and language pairs, the beauty of adaptable MT is that it can keep getting better. Whether adaptation is performed by the MT provider or via a self-service platform or whether it’s continuous or user-initiated, you can create a virtuous cycle of quality improvement through continual MT training. Given today’s advances in AI deep learning coupled with rapidly unlocking novel use cases of machine translation, adaptive capabilities represent a real opportunity for many organizations to maximize the overall value of their MT investment.
To explore the concept of adaptable MT and its benefits further, see how we implement adaptable MT within Language Weaver in this webinar recording.
Heather Rossi

Heather Rossi

Senior Solutions Consultant, Language Weaver
Heather is a machine-learning consultant focused on the practical use of linguistic AI in a variety of industries. She specializes in use cases where multilingual content needs to be translated and comprehended.
All from Heather Rossi