Generative AI: Using LLM's To Produce Audience-Tailored Translations In Place Of Classical NMT's One-Size-Fits-All

Last month we explored the ability of Large Language Models (LLMs) to produce higher-quality translations than traditional Neural Machine Translation (NMT) solutions. They are also more accurate at translating culturally recent phrases such as "dropping" a song. Yet, perhaps one of their greatest untapped powers that has never really existed before in mainstream commercial machine translation is their ability to create audience-tailored translations in place of classical NMT's one-size-fits-all translations. In other words, historical NMT solutions were designed to output a single "best" translation for each input, designed to most faithfully capture the meaning of the source text. In contrast, human translation often involves carefully considering the target audience of the translation, incorporating summarization and synthesis to produce a tailored translation that is customized for the target audience. For example, a passage from a medical text consisting of highly technical and arcane medical terminology would be translated as-is into the target language with the highest possible fidelity for a medical audience. If the same text was being translated for a general-audience website, the translator might summarize and distill the language down while translating it. For the first time in MT's seven decades, LLMs offer the unique opportunity to recreate this kind of audience customization in the mainstream translation process.

What might audience tailoring look like in a typical machine translation environment? Translation does not just encompass converting a passage from one language to another. It can also involve changing its "voice" if we think of voice as a form of linguistic transformation. To demonstrate audience tailoring, what if we take a simple English sentence "The Eagles released a new song yesterday" and change it it slang, stilted and non-native English? The prompts below were run using ChatGPT:

Translate into popular English slang: "The Eagles released a new song yesterday."
	The Eagles dropped a sick new track yesterday.
	The Eagles dropped a fresh banger yesterday.
	Yo, The Eagles just dropped a fire new track yesterday.
Translate into stilted English: "The Eagles released a new song yesterday."
	Yesterday witnessed the release of a fresh melody by the esteemed musical ensemble, The Eagles.
Rewrite in the voice of a non-native English speaker: "The Eagles released a new song yesterday."
	Yesterday, The Eagles, they make the release of a new song.
Rewrite in the voice of a Spanish speaker who learned English: "The Eagles released a new song yesterday."
	Yesterday, The Eagles, they released a new song.
Rewrite in the voice of a German speaker who learned English: "The Eagles released a new song yesterday."
	Yesterday, The Eagles, they have released a new song.

Slang "translation" is an especially fascinating area of transformation in that even the best copywriters aren't always up to date with the latest cultural constructions and metaphors, with LLMs offering the unique ability to "translate" formal corporate jargon into informal speech aimed at specific platforms. For example, a bland corporate product announcement could be more readily translated into social-friendly speech.

What about traditional translation from one language to another? Let's try translating the sentence from above into Spanish using Google, Bing and then a variety of ChatGPT prompts:

Sentence: The Eagles released a new song yesterday.

Google: The Eagles lanzó una nueva canción ayer.
Bing: Los Eagles lanzaron una nueva canción ayer.

ChatGPT: 
Translate into Spanish: "The Eagles released a new song yesterday."
	Los Eagles lanzaron una nueva canción ayer.
Translate into contemporary social media Spanish: "The Eagles released a new song yesterday."
	¡Los Eagles lanzaron una nueva canción ayer! #MúsicaNueva #Eagles #Estreno
Translate into Spanish slang: "The Eagles released a new song yesterday."
	Los Eagles soltaron un nuevo temazo ayer.
Translate into formal Spanish: "The Eagles released a new song yesterday."
	Los Eagles lanzaron una nueva canción ayer.
Translate into high school Spanish: "The Eagles released a new song yesterday."
	Ayer, los Eagles lanzaron una nueva canción.

What about a more complicated mixed summarization-translation task? Let's ask ChatGPT to construct a 50-word technical blurb about heart attacks:

Write a 50-word highly technical article about myocardial infarction aimed at medical professionals.
Myocardial infarction (MI) results from coronary artery occlusion, causing ischemia and subsequent myocardial cell necrosis. Prompt diagnosis via ECG, biomarkers, and imaging modalities is crucial. Early reperfusion therapy, along with adjunctive pharmacotherapy, plays a pivotal role in improving patient outcomes.

And repeat but in Spanish:

Write a 50-word highly technical article in Spanish about myocardial infarction aimed at medical professionals.
El infarto de miocardio (IM) se produce por la oclusión de una arteria coronaria, causando isquemia y necrosis celular miocárdica. El diagnóstico oportuno mediante ECG, biomarcadores y modalidades de imagen es crucial. La terapia de reperfusión temprana, junto con la farmacoterapia adyuvante, juega un papel fundamental en la mejora de los resultados del paciente.

Note the lack of creativity here, with ChatGPT generating a Spanish text that is nearly identical to the English version.

Google Translate translates this as:

Myocardial infarction (MI) is produced by the occlusion of a coronary artery, causing ischemia and myocardial cell necrosis. Timely diagnosis using ECG, biomarkers, and imaging modalities is crucial. Early reperfusion therapy, along with adjuvant pharmacotherapy, plays a critical role in improving patient outcomes.

What about ChatGPT?

Translate the following text into English: ""
Myocardial infarction (MI) occurs due to the occlusion of a coronary artery, causing ischemia and myocardial cell necrosis. Timely diagnosis through ECG, biomarkers, and imaging modalities is crucial. Early reperfusion therapy, along with adjunctive pharmacotherapy, plays a fundamental role in improving patient outcomes.

The two differ only slightly, with Google being more faithful to the source text with "is produced by" than ChatGPT's "occurs due to" while ChatGPT is more accurate with "plays a fundamental role" than Google's "critical role". The only substantive difference is that Google correctly translates the technical term "adjuvant" whereas ChatGPT changes it to "adjunctive". In this case, the two terms are interchangeable, with PubMed preferring the former by 50 results to 197, while Bing prefers the latter with 49,800 to 39,500 results. Based on the expected training data for ChatGPT, this suggests that Google Translate prioritized faithfulness, while ChatGPT preferred fluency (the most common phrasing in open web content).

But what about audience tailoring? Here we take the Spanish blurb from above ("El infarto de miocardio (IM) se produce por la oclusión de una arteria coronaria, causando isquemia y necrosis celular miocárdica. El diagnóstico oportuno mediante ECG, biomarcadores y modalidades de imagen es crucial. La terapia de reperfusión temprana, junto con la farmacoterapia adyuvante, juega un papel fundamental en la mejora de los resultados del paciente.") and ask ChatGPT to translate for different audiences:

Translate the following text into English: ""
Myocardial infarction (MI) occurs due to the occlusion of a coronary artery, causing ischemia and myocardial cell necrosis. Timely diagnosis through ECG, biomarkers, and imaging modalities is crucial. Early reperfusion therapy, along with adjunctive pharmacotherapy, plays a fundamental role in improving patient outcomes.

Translate the following text into English for Twitter: ""
Myocardial infarction (MI) occurs due to the occlusion of a coronary artery, causing ischemia and myocardial cell necrosis. Timely diagnosis through ECG, biomarkers, and imaging modalities is crucial. Early reperfusion therapy, along with adjunctive pharmacotherapy, plays a fundamental role in improving patient outcomes. #HeartHealth #MyocardialInfarction #TimelyDiagnosis

Translate the following text into simple English for non-native English speakers: ""
Myocardial infarction (MI) happens when a coronary artery gets blocked, causing lack of blood flow and cell death in the heart muscle. It is important to diagnose it quickly using tests like ECG, biomarkers, and imaging. Early reperfusion treatment and additional medications are crucial for improving patient outcomes.

Translate the following text into simple third-grade English: ""
A heart attack happens when a blood vessel in the heart gets blocked, which stops the blood from flowing and causes the heart muscle cells to die. It's really important to diagnose it quickly using tests like ECG, biomarkers, and special pictures of the heart. Treatment that opens up the blocked blood vessel and medicine that helps the heart work better are very important for helping the patient get better.

Translate the following text into English a kindergartener could understand: ""
A heart attack is when a blood vessel in the heart gets blocked and hurts the heart. Doctors use special tests like ECG, biomarkers, and pictures of the heart to find out what's wrong. They give special medicine and treatments to make the heart feel better and help the person get better.

Here we can see ChatGPT successively reducing the complexity of the text and adopting age-specific cliches like "make the X feel better" and "special pictures of X".

Yet at the same time, we can also see the very real limitations of audience tailoring in the absence of actual understanding, with ChatGPT passing through terms like "ECG" and "biomarkers" even in its kindergartener-focused summary, which are above the vocabulary of most kindergarten-aged children. This is a byproduct of the completion-based nature of LLMs in that they don't actually have an underlying semantic understanding of concepts, they merely encode word correlations that map words to their surrounding context probabilities. Here it is unlikely that ChatGPT encountered sufficient examples of such complex medical terms being explained to children in its training data to learn an appropriate mapping to age-appropriate terminology. A human translator would construct an age-specific metaphor to translate, but LLM's use of correlations rather than actual semantic models severely constrains their ability to do so, capturing their key limitations in creative tasks like audience-tailored translation.