Artificial intelligence has brought significant and often surprising innovations to the field of communication, and also to the field of translation. Long-known tools such as the various machine translation systems are now joined by much more advanced systems, such as ChatGPT, a generative language model developed by OpenAI, currently the best known to the public. Although from a conceptual standpoint one and the other are indeed based on the concept of Artificial Intelligence, the specific functions and capabilities of one and the other differ in significant ways.
Let's start with ChatGPT: what is it? It is a language model based on the transformer architecture, developed by OpenAI. This model has been trained on a huge range of data collected largely on the Internet, and is programmed to generate coherent and fluent responses - hence, almost human-like in tone - to questions asked in natural language. However, it is essential to understand that it is not specifically designed to perform text translation.
Machine translation, on the other hand, has been known to us for longer: the term refers to systems that use AI models exclusively to translate text from one language to another. To name a few very famous names in this regard we can mention the very popular Google Translate, DeepL, and Microsoft Translator. These systems are, unlike their predecessors, built and trained specifically for the task of translating texts between different languages, and to perform this task in a wide range of languages and styles, from business translation, to scientific translation, to many other possible registers.
ChatGPT and machine translation systems are powered by machine learning algorithms, but as we mentioned they are therefore "trained" with different goals.
ChatGPT is a language model whose training is based on a large corpus of Internet data in order to teach it to understand and generate consistent responses in reaction to as wide a variety of natural language inputs as possible. The stated purpose of those who designed it is to mimic human conversation, and to generate credible and consistent responses to user prompts, ideally going so far as to respond in a manner perfectly confusable with that of a human being. This broad and generalized training has given ChatGPT the ability to juggle a wide range of situations and topics, to be able to answer specific questions as well as to create original stories, or to generate of coherent content from a cue provided in natural language by the user.
Machine translation systems, on the other hand, have always been trained with a much more specific goal, namely precisely the conversion of texts between different languages. The training carried out on these systems is therefore of an entirely different nature: it consists of exposing them to huge amounts of aligned bilingual (or multilingual) texts, so that they are in a position to learn with which expressions they can accurately translate words, phrases, and whole passages from one language to another. Like all very focused training, this has made them exceptionally effective (unthinkably so, at first) in translation, but has in no way prepared them to operate in other areas, such as generating original text or interacting in dialogue contexts with a human being.
ChatGPT does indeed have some translation skills; it can translate simple sentences or even a few longer expressions from one language to another, but the moment it is exposed to more complex, technically challenging texts, or texts that have content closely related to particular cultural nuances that require deep understanding, it begins to make even egregious errors. The versatility we mentioned earlier results in less expertise on such specific tasks, inevitably.
For machine translation systems, we could say that the situation is almost mirrored. These are software designed and optimized to handle a wide range of translation tasks among many different languages, which they are able to perform on a very wide variety of content where both simple texts and complex technical documents are included, and they perform them with great accuracy. However, it is important to remember that even such targeted programming and specific training have not made these systems capable of handling with absolute accuracy linguistic and cultural nuances, language-specific idioms, and even highly specific or technical contexts, all areas in which they still make mistakes.
ChatGPT is designed to be an excellent tool for natural language interaction, and consequently it must be able to produce human reactions as natural responses to questions asked. This has led it to excel in creating original content such as stories or articles, brainstorming ideas, and simulating human conversation.
Machine translation systems, on the other hand, were in no way designed to generate new content or interact while maintaining a human appearance. Their main function is translation, and they are designed and trained exclusively to make them capable of performing it in even complex contexts, with special terminologies and advanced grammatical and semantic structures.
In summary, both ChatGPT and machine translation systems have in common that they are based on Artificial Intelligence, but having different purposes and competencies they also have very little overlap between their capabilities. ChatGPT, therefore, cannot reliably act as a translator except for the simplest sentences in the most commonly used languages: a machine translation job (which, in any case, must then be reviewed by a human translator) requires a dedicated tool.
Photo by Monika Grafik on Pixabay
Sections