AI translation, or machine translation, has advanced enough to be used in healthcare. However, it still has shortcomings. Weighing the capabilities and limits of AI translation is vital to language conversion decisions in healthcare and other fields.
There are several aspects of AI translation to understand in order to digest the implications of the proposed additions to Section 1557 of the Affordable Care Act. Likewise, getting familiar with the established types of editing for translated text or speech shows how Section 1557 changes may affect organizations. For example, choosing the best machine translation engine and level of post-editing for healthcare translation projects can save time and costs while improving care.
Background: Machine Translation of Languages
AI translation has improved to the point that, under some conditions, a machine translation engine can translate large swaths of communication. Digital translation sometimes saves costs and time. Nevertheless, machine translation engines’ abilities are still limited. Humans remain vital to the translation process. However, if translation quality continues to increase, language professionals may serve the role of “pre-editor” and “post-editor” when a machine translation engine converts the communication into the target language.
What Are the Four Types of Machine Translation?
1) Statistical Machine Translation or SMT
Statistical machine translation is a tactic that uses technology to digest large amounts of bilingual text. Then, it uses statistical models to produce a translation. It’s probability-driven translation. Statistical machine translations use examples instead of the earlier rule-based approaches.
However, SMT doesn’t consider syntax (word arrangement) well enough to be consistent. It usually requires translation post-editing in the target language. Also, some think SMT’s translation pipeline involved too much updating and fixing.
2) Rule-Based Machine Translation or RBMT
RBMT is the oldest type of translation. It uses grammatical rules to translate. Deep post-editing is usually necessary unless the structure is basic and the context is straightforward.
3) Hybrid Machine Translation or HMT
Hybrid translation integrates rule-based translation and statistical machine translation. Two different machines may be used in the translation workflow. It uses translation memories.
4) Neural Machine Translation or NMT
Neural machine translation uses neural networks to represent the process of translation. It also forgoes discrete symbolic representations in favor of continuous representations. It doesn’t have a pipeline but still uses statistics. NMT can be specialized in a certain field, like engineering.
The best machine translation to replace human translation
Neural machine translation shows the most promise. However, it may not be necessary when the context and grammar are simple. Older, more accessible methods like statistical machine translation may be sufficient. Some language pairs, like Portuguese and Spanish, are easier than those with different alphabets and grammar systems.
What is machine translation post-editing?
Machine translation post-editing addresses the shortcomings of translations carried out by AI. Specifically, machine language conversion still has limitations in accuracy. Also, achieving the word choice and natural flow of a native speaker is not easy for a machine. Therefore, a human edits the machine-translated text afterward. After the translating program has finished, this review by a human translator or a bilingual leads to the term “machine translation post-editing.” Having a professional translator or interpreter review the text is the best option. Businesses and organizations are standardizing the editing processes after a machine converts text (or speech). Right now, they are termed: “Light post editing” and “Full post editing.”
“Light Post Editing” Compared to “Full Post Editing”
Thus far, the business world has delineated two types of editing that people carry out on AI-translated communication. They are “light post-editing” and full “post editing.” Both are types of “machine translation post-editing.” While the names indicate how much time and effort is spent after the translation, the differences are more complex. The goals and larger picture of the text often determine whether to use light or full editing.
Light post-editing
Light post-editing makes the text intelligible and free of glaring errors. It still may sound far from a native speaker’s communication at times. Lighter editing is fine for projects destined internally within an organization. When speed and cost are priorities, and exact accuracy won’t affect decisions, then a light review by a post editor will suffice. It’s also better for shorter and simpler communication. When the grammar and syntax are simple, or the language pairs are close (like Swedish and Norwegian), a human translation may not be necessary. Sometimes, it’s acceptable for manuals or technical instructions.
Lighter machine translation post-editing shouldn’t be relied on for client-facing communication. In other words, the text may not instill enough confidence and trust. Potential customers or others looking to engage with an organization should trust its authority and ability to deliver based on their communication. Texts with nuances and longer, convoluted structures or clauses might still be unclear after a post editor finishes at an LPE level. In addition, regional particulars or slang usually need deeper editing, especially for distant language pairs.
Full post-editing
This editing requires a deep understanding of the languages. It also needs to consider the communication’s goals deeply. Full post-editing analyzes the entire text to eliminate more complicated errors: consistency, goals for style, persuasion, tone and other nuances.
A post editor will ensure consistency throughout the text. Jargon and technical terms, various stylistic elements and formatting will stay the same. Localization elements like measuring systems are standardized, too.
Overall, full post-editing is better for authoritative, persuasive or otherwise nuanced communication. Full post-editing moves nuanced text or speech closer to a native speaker’s language style. Outside readers or listeners will have more faith in the communication after FPE.
One concern is that it may not have saved time if the machine translation engine produces odd or error-ridden text that bogs down the post editor with needed corrections. A human translator may work faster. It’s important to weigh the machine translations’ distance from native communication along with editing costs versus human translation costs.
In some sectors, like law or medicine, it’s necessary to consider liability. Risks associated with foregoing the needed amount of editing, in terms of machine translation post-editing, can be damaging. Having language professionals assess, edit and proofread is important.
Saving Costs through Pre editing
Instead of conducting extensive machine translation post-editing, another type of editing sometimes reduces the effort and costs of post-editing. Pre-editing communication before translation makes sense when it will be translated into several languages. Some organizations only allow pre-edited text to undergo AI translation.
Pre-editing involves changing or removing parts of a text that are likely to result in the machine translation engine making errors. Nipping problems or complexity in the bud is the goal. It can save costs and time. Fixing, simplifying and standardizing the text ahead of time can save needed corrections that might be multiplied by the number of language pairs.
Text that is likely to result in translation errors may have first-language errors or be ambiguous. Long sentences are another problem. By reducing text to simpler structures ahead of time, the machine need not try to follow the threads of additional clauses. It doesn’t have to try to reproduce them in the target language. For texts with jargon, technical terms and other particularities, standardizing the terms can help the translation quality.
Machine Translation in Healthcare
There are settings where machine translation can solve problems when used correctly. Healthcare environments often drive discussions about translation and interpretation. Using AI translation in healthcare shows the potential utility, convenience and cost-effectiveness in reaching a target language.
However, the potential for medical problems or negligence due to language conversion errors makes it wise to regulate digital translation usage. The US government has begun addressing how human translation and AI translation should be used in healthcare. It also speaks to the need for machine translation post-editing. Specifically, a proposed rule to Section 1557 of the Affordable Care Act tries to guarantee that human linguists review important medical translations and interpretations.
What is Section 1557 of the Affordable Care Act?
Section 1557 prevents discrimination based on a patient’s personal characteristics in a healthcare environment. Patients should not suffer discrimination based on who they are. Bear in mind that this protection also extends to patients’ caregivers who may prefer other languages.
Specifically, Section 1557 “prohibits discrimination based on race, color, national origin, age, disability, or sex (including pregnancy, sexual orientation, gender identity, and sex characteristics) in covered health programs or activities.”
Who does Section 1557 protect?
Section 1557 protects patients and people helping them, like their family members. It seeks to prevent discrimination against patients. It also tries to prevent gaps in care, worse care or inappropriate care.
Section 1557 comes into play for translation and interpretation for Limited English Proficient (LEP) patients in the U.S. It includes people who would prefer conducting their healthcare in another language. Section 1557 aims to ensure patients receive the same access to healthcare as a native English speaker.
Therefore, for medical interpretation and translation projects, healthcare providers must seek to provide high-quality translations. Reaching human translators and interpreters quickly has always been difficult and can be expensive. Some language conversions may be conducted by machine translation engines with thorough post-editing.
What are the new policies about machine translation in healthcare?
Patients and providers sometimes turn to translation engines in healthcare settings. When professional translators and interpreters aren’t available, healthcare professionals must move care along. The government recognizes both the utility and risks of using raw machine translations in healthcare.
Given the persisting limitations of machine translation, DHHS has proposed adding a rule to Section 1557 of the Affordable Care Act. This rule would mandate that a human review every machine translation in order to prevent medical problems caused by improper translations.
Specifically, this proposed rule reads:
“If a covered entity uses machine translation when the underlying text is critical to the rights, benefits, or meaningful access of a limited English proficient individual, when accuracy is essential, or when the source documents or materials contain complex, nonliteral or technical language, the translation must be reviewed by a qualified human translator.”
It’s clear that requiring post-editing in healthcare environments has become a priority for DHHS. The potential for mistranslations and confusion is too risky in medical situations to extract human judgment. Digital translation has not evolved enough to forgo post-editing.
While mandating that human translation at every juncture would prioritize high-quality translations, in the past, DHHS has stated it does not want to stifle the development of AI translation. Of course, DHHS sees that the convenience of machine translations for simple, short, or straightforward communication is effective and inexpensive or free. Time can save lives and preserve health.
In the proposed rule in August of 2022, DHHS did not specify what qualifications a human translator needs for post-editing. However, human discretion about translations remains important, not just because of medical jargon. Human judgment is also still needed because of the complicated usage of words in medical situations.
What do changes in machine translation in 1557 mean for healthcare organizations?
Healthcare organizations will need to pay attention not just to interpretation and translation quality but also to prepare for additional steps if they decide not to rely on human translation. They should be ready to either contract a post editor or work with a linguist or a company that provides post-editing machine translation.
Overall, having a language access plan in place is important for healthcare organizations. Taking into account the limited capacities of a machine translation engine is important. Proactively lining up a post editor or a company offering post-editing should be part of any language procedure involving only machine translation.
Some healthcare entities hope to make their own in-house medical linguists to cover more territory or hire more linguists. For providers who contract interpreters and translators on a case-by-case basis, they may save costs by having AI complete some legwork beforehand. Some hospitals, networks, practices and other medical organizations already contract language service companies to provide comprehensive language solutions. Others may need to start doing so.
In any type of healthcare entity facing translation projects, ensuring that editors review raw machine translations before they are utilized is becoming vital. This can avoid healthcare problems and liability concerns. Hopefully, DHHS will specify who could serve as a qualified editor. In that case, there may be an increase in standardization of the post-editing process.
AI Translation & Healthcare Summary
Understanding the strengths and limitations of a machine translation engine compared to human translation is important for many sectors. Healthcare is an important setting for the application of digital translation. There are four types of AI translation. These include statistical machine translation, rule-based MT, hybrid MT and neural MT. The different levels of post-editing, called light or full post-editing, address the shortcomings of raw machine translation. Deciding between full or light machine translation post-editing depends on the purpose, audience, communication style, and required speed. Pre-editing can prevent grammatical errors and other shortcomings when the communication is converted into several target languages.
AI translations can assist with speed, convenience and costs in healthcare. However, the potential for translation errors resulting in medical mistakes or negligence has caused DHHS to address machine translation engines’ usage in healthcare. Their policies fall under Section 1557 of the Affordable Care Act, which prohibits discrimination. Those who speak other languages should receive the same access to healthcare as native English speakers in the US.
DHHS has proposed a rule stating that AI translations must be reviewed by a “qualified human translator” when accuracy is vital or when the text is not straightforward. This required machine translation post-editing would stand for communication involving nonliteral, technical or difficult language.
If this proposed rule is enacted with this same verbiage, some healthcare entities will need to establish procedures for machine translation post-editing. These procedures will differ depending on whether an organization has in-house medical linguists, contracts them or works with a company that provides medical translation and interpretation.
Post-editing machine translation will likely become a growing specialty for interpreters and translators. Post editor would be a less demanding role than converting language. In addition, further specialization into medical pre and post-editing specialties would be a natural evolution.