AI - Can we trust automated correction?
The 1980s was the invention of the Rubik's Cube, the release of The Empire Strikes Back and the royal wedding between Charles and Princess Diana... but did you know that it was also the arrival — more discreetly — of WordCheck, the first spellchecker?
In those early days, spellchecking was simply a matter of comparing the words in a text with a database. If the word did not exist in the database, then it was underlined. There were no grammar checks or suggestions for correction, but it did help to highlight a number of typos.
Over forty years later, where are we?
Comparing today’s automated proofreaders
Let's compare the proofreaders of three writing programs, Microsoft Word, Google Docs and Talers.
We will test them on three different types of mistakes:
Spelling
Grammar
Style (pleonasms, repetitions etc.)
(Word Online)
(Google Docs)
(Talers)
As you can see, all of them detect and highlight spelling mistakes. Word (1/2) and Talers are able to spot grammar errors while Google Docs doesn’t.
Our own corrector goes beyond standard expectations by pointing out a stylistic mistake. Thus proving an alternative in order to make it easier to read overall.
However, this type of errors are the most difficult to detect because it needs a global understanding of the language and the context of the text itself. Furthermore, style is very different from one person to another and having harsh standards would snuff out creativity and diversity.
A new era with artificial intelligence
Today's proofreaders use a system of handwritten rules to check whether a sentence is grammatically correct and does not contain any obvious style errors.
This method has the advantage of being simple and efficient. On the other hand, one would have to create a tremendous amount of rules if one really wanted a proofreader to detect all mistakes as well as a human does.
For example, phonetic errors like “Bring me sum!” are still ignored by automatic correctors.
The recent boost in artificial intelligence brings a new generation of automatic proofreaders, even more efficient, but also much slower than the current ones.
Like a human, this artificial intelligence is able to understand the context, the tone and the “environment” of a text.
If we ask GPT-3 to correct the text “Bring me sum!”, the artificial intelligence correctly returns the following sentence: “Bring me some!”
Artificial intelligence brings a new level of proofreading, but they can be biased depending on the sources that have been used to “tune” them (source). This can lead to errors related to the cultural or linguistic context of the user. Languages such as English or Spanish, which have many variations depending on the country in which they are spoken, are perfect examples.
Conclusion
More than 40 years after "WordCheck", what did we learn? Can we trust automated correction?
The answer is nuanced, although today we can delegate an increasingly important role to automated proofreaders. It is always necessary to have a correction done by a person who can detect the tone and context clues therefore fine-tuning it if necessary.
Nevertheless, understanding the limits of autocorrection will allow to better exploit it, and to make it a complementary tool to improve the quality of one's writing as a whole.
At Talers, we recommend the following workflow:
Use an automated proofreader to spot common mistakes,
From time to time do a review with a qualified AI-tool,
When your text is ready, effectuate one last check with a pair of friendly human eyes.
