Human-optimized machine translation solutions to reach audiences faster

Machine translation (MT) is the automated and rapid process of translating text by a trained ‘engine’. In order to achieve human-level quality, MT is often followed by review and post-editing (PE) by expert in-country linguists. The key components of MT include:

• Input text – the source that needs to be translated
• The engine – algorithms, models, and linguistic rules that analyze the input text and generate an equivalent output in the target language
• Bilingual data – large amounts of parallel texts in two or more languages used to train the machine translation system

There are five key types of Machine translation:

• Rules-based machine translation (RBMT) uses a collection of “rules” governing the construction of language. This type of MT was the first commercially available but it’s no longer widely in use.
• Statistical machine translation (SMT) uses algorithms to produce millions of possible permutations and selects the text that appears to be the best translation. Like rules-based MT, statistical MT is no longer widely used.
• Hybrid machine translation (HMT) combines rule-based and statistical MT to improve translation quality. Using translation memory (TM), hybrid MT can deliver consistent quality.
• Neural machine translation (NMT) uses artificial neural networks to learn from training data. This type of MT predicts the most likely word sequence. It is more accurate and fluent than SMT because it learns the relationships between words and phrases. Also, NMT can translate rare or low-resource languages more effectively.
• LMMs (Large Language Models), like ChatGPT, Bard, and PALM2 are pre-trained on massive amounts of text data, allowing them to acquire language knowledge and patterns from diverse sources in all languages. They create content based on predictions, and so can generate text that sounds human-like and conversational. They are more fluent but less accurate and predictable than NMT.

These two terms are used interchangeably, but “AI translation” is a newer term. (The first machine translation engine was developed in the late 1940s). All forms of machine translation involve artificial intelligence. Another term you’ll see is “AI-enabled translation workflows”. MT and AI translation tools are used by linguists as a cost-effective, time-saving translation solution for brands looking to expand globally.

These two terms are often used interchangeably, but they differ in some ways:

• Machine translation provides the initial translation from the source language into the target language and stops there. One tool, an MT engine, is used here.
• Automatic or automated translation refers to the entire workflow and delivery of machine-translated outputs to client-facing platforms like support chatbots. Automated translation involves various tools to automate tasks via specific triggers within a translation workflow.

Whether or not machine translation is a smart option for your business depends on a few factors related to volumes, quality, budget, and timelines such as:

• Language requirements. AI translation does not handle all languages equally well, so businesses must understand and plan for that.
• The type of content. MT is best for straightforward, low-visibility, non-specialized text but customization makes many content types possible.
• Amount of content. MT is most viable when there are high volumes of content.
• Budget available and size of investment. Some MT options are free (open source) and others require significant investment (proprietary tools).
• Time to deploy. You can use an LLM right away but if you want to customize an NMT engine it will take time to prepare data and train the engine.
• Quality requirements. Low-priority/low-visibility content can be well-suited for MT because it doesn’t need to be of perfect or human-level quality to be fit for purpose.
• Security. Relying on public MT services like Google Translate poses security risks. Also, LMM models use any data you put in to further train the AI, meaning your content is no longer under your control. Similarly, you do not control and own the content that is newly produced.

Acclaro can help you take a look at your program needs and make the best decision on whether to use MT or not. We can then help you build a customized program that will meet your business goals.

Traditionally, MT has been best suited for content where quality was not the highest priority, such as product reviews, FAQs, user-generated content, and knowledge bases. Also, it is considered viable and effective for high volumes of content that would otherwise go untranslated. MT has historically not been considered optimal for highly branded, creative, highly-visible content like websites and ad campaigns. However, with the recent advancements in AI and MT that improve quality and fluency, machine translation engines, when trained, can be used effectively on a wide variety of content types.

The short answer is yes, especially when machine engine learning and a human review and editing step are deployed as part of the process. However, machine translation is typically not suited for very nuanced content like advertising slogans and other marketing content that is designed to convert.

When you are considering machine translation, you need to be aware that there can be pitfalls, including in the areas of:

• Complexity and accuracy: Machine translation may not handle complex and specialized content like legal and medical well. Human translation is more reliable and worth the extra investment in domains such as these where mistakes can be exceptionally damaging
• Quality: The level of fluency and linguistic quality may not be as good as you need it to be for your content, and the amount of post-editing required may eliminate any cost/time benefit. Machine translation has improved significantly over the past few years but there are still some limitations and challenges with more nuanced content like idioms, slang, and humor.
• Bias: When machines are trained on past (older) translated content that uses biased language, then that bias could be carried over into the translated text. Also, bias and hallucinations occur with Large Language Model (LLM) translations.
• Data privacy: Because some engines do not protect data, you need to consider where data or sensitive information is stored, shared, or accessed.
• Ethics: Some consider that the use of MT in education and journalism is not ethical.

Even though MT may provide you with cost and time benefits there are many other factors to weigh. The bottom line is that you should work with an experienced MT translation partner who can help decide if machine translation is the best option for your business.

There are significant benefits to the strategic implementation of MT for your organization including:

• Improved speed: Machine translation is automatic and near-instantaneous. Even when humans are involved, automating some parts can improve their efficiency significantly.
• Reduced costs: While enterprise MT programs are not free, an MT or MT+PE process can typically save up to 40% with customized engines in certain domains and for certain content types.
• Increased volumes of translated content: Some organizations have so much content that humans alone could never handle it. In this case MT allows you to translate content that would otherwise never be available to global customers or employees in their own language.

Most of the time, the best translation process involves both machine and human efforts. The balance may favor one or the other depending on the purpose and intent of the content being translated. It is a continuum. For example:

• For content with low visibility and that otherwise would not get translated at all, machine translation alone with little or no human involvement might be viable. Knowledge bases, reviews, or user-generated content are some examples.
• For content that is not highly nuanced, a human-plus-machine translation approach is best. Examples include blogs, service manuals, and other general/informational documentation, some web content, and some technical/scientific texts when not highly specialized.
• For content that has an emotional impact, we recommend that businesses use transcreation, which is an all-human process in which the content is adapted completely for a new market. In this process, idioms, slang, references, and emotional content are all recrafted to fit the target market. Web pages, ad campaigns, and branded content (taglines, slogans) are examples of this.

Global brands across industries such as eCommerce, technology/software, financial services, media & entertainment, travel & hospitality, and non-profits leverage machine translation as part of their global growth and localization strategies. Translation companies use machine translation and AI tools to help increase the productivity and efficiency of human translators and localization specialists.

MT does not replace humans, though their roles may evolve. Humans are still needed in roles such as:

• Post-editors, who make the changes needed post-translation so the text is accurate and fluent
• Solution architects and MT technologists, who help clients decide whether, when and how to use AI translation, and then set up the workflow
• Language data specialists, who know how to clean and use data to train MT software
• Quality engineers, who drive measurement programs and assess the quality of MT output so re-training can occur (using traditional TER/BLEU /METER measurement frameworks)
• Transcreators and in-country copywriters, who adapt or author creative or highly emotive texts that are beyond the capability of MT
• Language prompt engineers, who provide the LLM with the right instructions to capably handle the translation per the specifications.

Post-editing (PE) is the process by which a professional bilingual post-editor reviews and refines raw machine-translated content. This language professional is different from a translator, since their knowledge includes an awareness of what types of errors are possible in the machine translation process and how to edit to different levels of quality.
While the level of PE must be defined for each program, in the translation industry you will often hear about two types of post-editing:

• Light Post-Editing (LPE): Here the goal is to make minimal changes in order to ensure the translation is clear and accurate. Businesses use LPE for projects where speed and budget outrank quality.
• Full Post-Editing (FPE): This is a more thorough approach, considering not just accuracy but also tone and terminology. It takes longer and costs more, but it guarantees more fluid translations, higher cultural relevance, and better linguistic quality. Choose this approach when you need your translated content to sound as if it were written by a native speaker.

Post-editing by a trained bilingual linguist should be used after the engine has been fully trained but the quality still does not meet the requirements of the business. You can choose to post-edit all content, or a subset/sampling.

Relying on MT services like Google Translate poses security risks in a few ways:

• LMM models use any data you provide to it to further train the AI, which means that you no longer control your content. Also, any content that the AI produces is not your own data either.
• When you use an online MT service like Google Translate, your text is sent to the service provider’s servers for processing. If that content contains sensitive, proprietary, personal, or confidential information, there is a risk that sensitive or confidential information could be exposed.
• Depending on the MT service provider and its data storage practices, there may be concerns related to data residency and compliance with data protection regulations. Users and organizations need to be aware of where their data is processed and stored.

Machine translation (MT) has varying degrees of effectiveness depending on the language pairs involved. Generally, MT performs well for widely spoken languages with abundant training data, while it may face challenges with less common languages or language pairs. Languages that MT handles well include:

• Major world languages like English, Spanish, French, German, Chinese, Japanese, and Russian. These languages often have extensive bilingual corpora available for training, resulting in more accurate translations.
• “High resource” languages, meaning languages with a large volume of online content, such as articles, books, and websites, which contribute to better-trained models, improving the quality of MT. MT struggles when there is limited training data.

MT struggles with language that:

• Have complex grammar structures, extensive inflections, or multiple dialects. Examples include languages like Finnish, Hungarian, and Arabic. These are called ‘highly inflected languages.
• Have a limited online presence. If a language lacks diverse digital content, training MT models becomes challenging. This is often the case for some indigenous or minority languages.

While advancements in machine translation have significantly improved performance across languages, challenges persist in achieving consistently high-quality translations across all languages and language pairs.