ChatGPT Solves Some Translation Conundrums

The machine translation (MT) abilities of OpenAI’s ChatGPT represent a big leap forward from existing automatic translation—and as a former professional translator and simultaneous interpreter in several languages who has longed worked in this area, your columnist is in a position to know.

I recently studied the quality of ChatGPT translation for accuracy, innovation, grammar, and professional terminology, including general, legal and technology fields; the main goal of this comparison was to determine where ChatGPT falls within the quality spectrum of machine translation services for business use. The results were groundbreaking. OpenAI has solved translation issues that have haunted MT since its inception.

It appears to perform translations in all major languages, as well as several indigenous languages, such as Cherokee and Kinyarwanda. Significant further analysis of other languages is required for a full endorsement, but below are the results of my study of ChaGPT text translation from English into Spanish, French, and German.

THE UPSIDE

A high number of translations go through English at some point in the language conversion process. Since automatic translation was first invented, there has been a crying need for software to understand what English speakers mean to convey by their use of personal pronouns in order to generate proper translation.

The Devil’s in the Personal Pronouns

This article addresses personal pronouns in the linguistic and grammatical sense—not meant in any relation to social commentary.

Almost all languages have some type of gender differentiation in the construct of the language grammar itself. English has only six personal pronouns typically recognized by translation software (I, you, he, she, we, they). Most other languages have 8 to 10:

1. I
2. you-singular-formal
3. you-singular-informal
4. he, she
5. we-masculine
6. we-feminine
7. you all-plural-masculine
8. you all-plural-feminine
9. they-masculine
10. they-feminine

Languages like Japanese have even more due to their variety of honorific modes of address.

The baked-in language gender limitation is actually an influence and limitation to social evolution.

Why Are Personal Pronouns Important in Machine Translation?

In other languages, almost every word in the sentence changes according to the gender, plurality, and mode of address: the verbs, adjectives, articles, even prepositions. One wrong pronoun makes the entire translation fall apart. Yet OpenAI appears to have solved this serious issue, automatically translating surrounding words appropriately.

I myself have previously created hack solutions that pushed traditional MT to generate the proper pronouns during translation, yet now with ChatGPT, the proper pronoun will be automatically inserted as long as the subject that the pronoun represents is somewhere in the English text. If the subject is Maria, the “you” that ChatGPT will insert will be “you-singular-informal”; whereas if the subject is Mrs. Smith or a title like doctor or president, then “you-single-formal” will be inserted. Not all languages or countries use the informal “you,” but ChatGPT is easy to guide toward the correct mode of address for the audience. By contrast, the DeepL.com translator has a useful API setting for formal/informal.

Processing Paragraphs

With traditional approaches to automatic translation, pronouns that refer to previous nouns (words like it, that, them, these, those, who, which, etc.) can often be mistranslated because the translation is processing one sentence at a time, with therefore no memory when a pronoun refers to something found in previous text. ChatGPT is revolutionary in the translation world in that it processes in paragraphs, not single sentences, so it retains that memory. Without it, the wrong pronoun will probably be inserted, and thus all the surrounding words will be incorrectly translated as well, confusing the reader.

ChatGPT Comprehends the Intent of Text

Programming a computer to comprehend what a human person is trying to say is insanely difficult, especially because people wander in their thoughts, change direction mid-sentence, and forget what they wanted to say, particularly in non-professionally written text. Comprehension and handling of uncontrolled natural speech and composition is what has made DeepL Pro such a shining star in MT. ChatGPT also has that understanding.

THE DOWNSIDE

Vocabulary

The second half of software comprehension is understanding professional vocabulary. A good example of this would be the legal expression “attorney-in-fact.” Many services mistranslate attorney-in-fact as “someone who is, in fact, an attorney.” However, the attorney-in-fact is often not an attorney; the expression has a very special translation. Traditional translation services may insert a “guess word” if they don’t know the exact translation, much like the well-known AI examples of false guessing in sports articles and legal briefs. An advantage to ChatGPT translation is that there appears to be little or no guessing by the software; therefore, you can quickly spot words that were not translated by ChatGPT because those words remain in the original language. This is good, as it raises a flag to hire professional assistance to improve the translation. Yet notably, many languages actually use English tech words, having no equivalents in the formal version of their own language, so flags won’t always work with technology material.

Usability Issues

ChatGPT delivers high-caliber translation results; however, its major drawback is a painful lack of usability. Companies like Microsoft, with its Translator (Azure), and DeepL Pro have developed outstanding APIs as well as copy-paste web page processing fields that make performing translations easy for any company. ChatGPT and its derivatives, on the other hand, offer command line one-paragraph-at-a-time experiences in which the user must type, for instance, “Translate from English to German:” followed by the text to be translated, ending in double quotes. If you want to make an employee suffer, assigning that person to translate your website in this manner will achieve your objective. (I’ve also found two third-party applications that use the “command line” type translation feature of ChatGPT: one called GameBuster and another called AI-Pro. GameBuster has issues but does not provide customer service. AI-Pro.org, although new, is trying hard to create usable applications for the general public, and its customer service is phenomenal.) DeepL Pro and Microsoft Translator are infinitely easier to use and will offer the above translation improvements that OpenAI has rolled out.

* * *

This author looks enthusiastically forward to the day when the entire world will finally have outstanding automatic translation at their fingertips, easy to use and accurate. As a bonus, maybe it will translate our industry as “speech technology” rather than as “technology of talk.” 

Sue Reager specializes in across-language speech communication, applications, and context engines. Her innovations are licensed by Cisco Systems, Intel, and telecoms worldwide. For the prior 20-plus years Reager was responsible for translating software and media in Europe, Africa, South America, and the United States.