An improvement in the quality of Google Translate’s algorithm could aid in the development of evidence-based medical practices.
A study published in Annals of Internal Medicine finds that Google Translate provides article translations accurate enough to include in systematic reviews. Researchers used Google to translate articles in nine different languages and had non-fluent individuals abstract them.
When they compared those results with abstractions provided by native speakers, agreement between the two methods approached 90%.
Systematic reviews provide much of the basis for evidence-based decision-making in the healthcare system. The limitations of those reviews often include the exclusion of non-English articles, explains the study’s lead author, Jeffrey Jackson, M.D., of the Medical College of Wisconsin.
“The obvious problem is, how do you translate these articles if you’re not living in Washington, D.C., or someplace where you can find people who actually speak the language to look at your article?” says Jackson.
Most articles include raw data in a table format, which doesn’t require much sophisticated translation beyond understanding row and column titles. Before researchers get to that point, however, they have to decide whether the study’s quality is high enough to merit including its results in the systematic review.
If researchers can’t accurately translate the study’s methods and understand its limitations, they can’t determine its quality.
For big reviews encompassing hundreds of trials, excluding a handful on the basis of language probably isn’t an issue, according to Jackson.
“But often, these meta-analyses are on a relatively small number of studies,” Jackson says. “If it turns out you’re pooling five English trials, but you excluded two French trials, a Spanish trial, a German trial, a Chinese trial and an Indian trial, you may have excluded half the world’s literature.”
It’s also impossible to know how frequently that sort of situation arises, because articles that cite language limitations rarely say how many potentially eligible articles got excluded. Fortunately, regardless of the breadth of the problem, the team found a very reasonable solution: Jackson says the biggest trick was simply cleaning the text up properly to get a solid translation.
That amounted to cutting the text out of a PDF, dropping it into a word processor and getting rid of any funky non-text characters. Google Translate has a word limit, so Jackson fed the text in a paragraph at a time, making sure each chunk ended with a complete sentence.
“It really works—you got comprehensive sentences that, you know, sometimes the verb might not be exactly right, but there was no doubt about what was actually being said,” he says.
While the tool works well for research purposes, Jackson doesn’t see it taking the place of language interpreters in face-to-face interactions between doctors and patients.
“Hard data—like what was the average age of people in a study, how many weeks did they get medications, what were the entrance criteria—you can translate very easily with Google Translate. More nuanced stuff, though, not so much,” says Jackson.