AI translators struggle with nuances of language

Jan. 17—The effectiveness and usefulness of artificial intelligence when it comes to translating language remains a work in progress and depends on the goal of the translation.

Experts note AI isn't necessarily great at understanding nuances between languages and often misses the mark when it comes to literary translations. That's not to say AI can't be an effective tool in shrinking gaps in languages, they said, but as with AI's use in countless lanes, some guiderails and personal touch are required.

Some aspects of AI use large language models to "learn" the patterns of human-written works, and have become a chatbot that seems to do it all: generate ideas, write essays and even translate languages. ChatGPT was developed using a large corpus including everything from Shakespeare to the New York Times to Wikipedia, according to Joe Wilck, associate professor of practice in analytics & operations management at the Freeman College of Management at Bucknell University.

Because the writings used to develop ChatGPT also included other languages, the tech can perform translations between languages, depending on the amount of data it has "learned" about the languages.

"ChatGPT is able to deduce what you ask it and formulate words and sentences similar to what auto-complete does," Wilck said. "Google Translator is best if you're working with some obscure language because ChatGPT is not going to be trained on as many writings of those languages."

Bucknell Mathematics Professor Keegan Kang said ChatGPT is better than other, older technologies like Google Translator because it catches more nuances, with the caveat that you are mostly translating to English.

"Let's say you are translating German to French, it's possible that Google Translator might be better," Kang said.

Wilck said the accuracy of AI translations depends on the language you ask it about and whether you seek a literal, word-for-word translation or a translation that captures nuances.

"If you do English to Spanish and you use slang or phrases that mean something different than their literal definition, it can understand the slang in English and can convert to Spanish capturing the essence of what you are translating," Wilck said.

Nuances aside, if you are simply looking for a word-for-word translation, Wilck said technology like Google Translator is just as good, if not better than AI and ChatGPT.

"Google Translator is good, particularly with direct questions and is handy to have on your phone," Wilck said. "Chat GPT is better if you are using slang in the conversation."

Literary translations

Lynn Palermo is a retired professor from Susquehanna University. Her expertise is in literary and academic translations from French to English.

Palermo's concerns with AI and translation go beyond accuracy and lean toward ethics. Palermo said no professional would simply run a piece of text through ChatGPT and accept it.

"I went on ChatGPT and sent through an essay I've been working on from French to English," Palermo said. "At first glance, it looks pretty good, but then you start looking more closely and there are certain sentence structures in French that I almost always change to the corresponding English sentence translation. ChatGPT kept those French sentence structures."

Working with the AI technology was frustrating for Palermo. She said her first draft through ChatGPT consisted of mistakes she wouldn't have made had she done it herself. "Instead of my being an active translator, I'm a reactive editor," Palermo said.

This brought up another train of thought for Palermo: If AI is doing the translation, who is the author of that work? "Does that make me an author or an editor?" Palermo pondered. "Say I were translating a novel using ChatGPT, I ask 'who is taking credit for that translation?' The computer generated that translation."

Last summer, Kang used AI to translate French novels into English. Kang said there are several pitfalls to using AI to do the sort of lengthy translations he was working on.

"In French, there is 'tu,' which is informal and 'vu' which is formal. Both mean 'you' in English," Kang said. "Direct translation misses that kind of context and nuances. That really means knowing what the situation is and adding additional words."

AI's limited character count also resulted in some lost references in Kang's work, he said.

"Usually, references are not captured across different chapters," Kang said. "ChatGPT had a memory of 6,000 characters as of last year, and would forget about what had been translated before."

Phrases or idioms that relate to a certain culture may slip through the cracks when it comes to AI. Kang said there are phrases that have to do with French history that would probably not make sense to English-speaking people given a direct translation.

Palermo said AI cannot read the context and, therefore, misses small, but essential communications.

According to Kang, the average person will have difficulty deciphering whether something was machine-translated or not. "If you're looking to win literary prizes, AI doesn't quite cut it yet, but for something you just want to read to relax and forget about, AI does the job," Kang said.

Palermo, though, was adamant in her belief that AI would not make much of a dent in the field of literary translation.

"Literary translation is becoming more and more recognized as a form of creative writing," Palermo said. "It will always depend on the translator knowing better. If people are using it and just run it through and accept that, it'll be disastrous."

In her conversations with the American Translators Association (ATA), Palermo said the organization said they consider AI to be another tool. "They have courses on how to use it and how not to use it and I think it will continue to be part of our education," she said.

The ATA released a statement on AI in November and said while the tool can be helpful in low-stakes situations, it is insufficient in creative ones.

"While automated translation has proven helpful — and even welcome — as a tool for certain low-stakes, time-sensitive tasks, it is less so for others. This is especially true for creative endeavors, where human ingenuity is key, in light of the reliance of AI tools on previously produced ideas and information," the association said. "Language experts are likewise essential to multilingual communications whenever the stakes are high, helping non-specialists determine how and when to deploy technology."