4 OMT and large language models (LLM)
The use of automated translation tools has become part of daily life, whether we are aware of this or not. Today, social media platforms (such as Facebook or Instagram) and internet browsers (such as Microsoft Edge and Google Chrome), use integrated plug-in applications that support the instant translation of any text on a website. Sometimes these are visible, other times they are not. Websites usually inform users that a text is machine translated, but this may not always be obvious. In some browsers, by highlighting the text and then using the right-click function on the mouse, the option to translate any text is also given.
This course has recommended specific OMTs [Tip: hold Ctrl and click a link to open it in a new tab. (Hide tip)] , as using integrated machine translation plug-ins within browsers does not offer the opportunity to evaluate these tools in full. Indeed, it may not be clear which translation engine you are using, so you will also have problems when trying to reference your sources correctly.
You may also be wondering why we are not suggesting using AI tools, such as Chat-GPT (Generative Pre-Trained Transformer) to practise automated translation. These tools, also available online, are defined as Large Language Models (LLMs), and provide an interactive and conversational experience comparable to chatting with human users. Similarly to Neural Machine Translation (NMT), they use Artificial Neural Networks to predict a sequence of words and formulate a sentence in a specific language. They show capabilities in English and other languages. However, there are two main differences between OMTs and LLMs to consider: how machines are trained and how they interact with users. This is shown in Table 1.
| OMT | LLMs | |
|---|---|---|
| Training | OMT systems are trained on ‘parallel corpora’, which means that the machine is specifically and exclusively trained to translate from language A to language B (also called ‘language pair’). Therefore, they specialise in some language pairs, although not all. There are over 7000 languages spoken in the world, and Google Translate, at the time of writing this course, is only available for 243 of them. | LLMs are predominantly trained in English, but can still translate to some extent. Research to understand how LLMs produce translations is still ongoing, and their use cannot be encouraged for this task or, indeed, in this course, as yet. |
| Interaction | OMT rigidly and statically provide the translation of a source text into a (mostly) equivalent target text. The relationship between the user and the platform is unambiguous and confined to the translation of a specific source text. | LLMs require ‘prompts’ to function. Prompts are instructions that the user dialogically provides to the machine so the machine can respond. Prompting is not a straightforward skill, and small errors may cause the machine to ‘hallucinate’, or to respond to the dialogical prompt in an unexpected way, thus providing incorrect answers. As above, more research is required to evaluate LLMs translation functionalities. |