From small to large: what is a large language model? A large language model (LLM) is an artificial intelligence (AI) system specially designed to understand and produce human language. Although large language models are a relatively new development in the world of AI and machine learning, the concept of language models dates back to the 1950s and 60s. During this period, researchers experimented with different statistical models to simulate natural language. Around the turn of the century, new machine learning techniques, such as the rise of neural networks, helped to build increasingly sophisticated language models. Crucial for large language models, however, is the recent advent of deep learning and the immense amounts of textual data that the internet has to offer today.

All-in-One

When hanging a painting, you probably need a hammer and nail. If you want to build a complete house, you will need to expand your toolbox. Traditional language models are like a hammer and a nail. Based on rules and statistics, they can perform simple language tasks, such as predicting the next word in a sentence, but they have a limited understanding of context and semantics. Large language models, on the other hand, are a complete toolbox. Thanks to their large size and training data, these models can be used for a much wider range of language tasks. They generate new text, from articles to computer code. Large language models can produce translations and summaries and answer questions. Moreover, they can even perform creative tasks, think of generating art or writing poetry.

How does a large language model work? Simply put, the system is trained on massive amounts of textual data and learns to recognize patterns and understand relationships between text elements. But how does the model manage this complex process? Like an onion, large language models have different layers that each contribute to the complex structure.

The embedding layer converts each word in the input text into a vector – a series of numbers – that the model can understand and process. The distance between these vectors indicates the semantic relationship between the words. In other words, the distance between the words cat and bicycle vacation will be greater than that between coffee and tea. Then the feedforward layer applies certain weights and biases to these vectors, which makes the model more accurate and efficient. The recurrent layer ensures that the information that has been processed earlier in the text is stored in the model’s memory. This allows the model to understand contextual information and interpret the meaning of words in their sentence context. Large language models also have an attention mechanism with which they can focus on the most relevant parts of the input text. Want to know more about this mechanism? Then take a look at our previous Simply Story!

Large language models and translations: is it a match? We keep repeating it, but an important feature of large language models is that they are trained on huge amounts of text. Because these texts come from the internet, the training data contains a variety of languages – though some much more than others. The language model trains itself to learn the underlying structures and patterns of language, not of one language, but of several. Large language models are therefore capable of generating translations.

Pros and Cons

One advantage of the translations from these language models is that they take into account context. Texts are not translated piece by piece or sentence by sentence but as a whole. This is very handy for words whose meaning depends on the entire context. The English word ‘bat’, for example, can be an animal or a baseball bat. In the wrong content, or with a language model that is not sufficiently trained in your context, many outcomes are possible.

Moreover, it is possible to give language models like ChatGPT additional information so that they can take into account the style of the translations. For example, indicate that your text is a poem, news article, or blog post. If you have a young or an older target group in mind, mention this in your assignment to the language model. Large language models also bring benefits for translations in the e-commerce world. You can immediately ask the translation assignment whether the model also processes SEO keywords in the translation.

So, is everything hunky-dory if you want to use language models for your translations? No, unfortunately not. To use the English saying, large language models are (still) a jack of all trades, but master of none. In other words, it can and does so much that it does not yet have the necessary specialization required for delivering quality translations.

A study in which the founders of ChatGPT, among others, participated predicts that about 80% of the workforce in the United States will feel the effect of large language models. For some professions, this means that a small percentage of daily tasks are taken over, while for others, it could be the majority of the work.

The translation industry is expected to fall into the group where a larger part of tasks is automated in the future. Does this mean that translation agencies will be a thing of the past in a few years? Certainly not! Even when using large language models, experts are still needed who can provide the correct input (ask the right question) and who can check and improve the quality of the translations. These first and last steps in the translation process will only become more important in the future.

Do you want to know what large language models can mean for your translations, but you can’t see the forest for the ChatGPT trees? Then contact Simply Translate. Our software experts are fully aware of all new developments in this rapidly growing field. Moreover, the combination of tech know-how and first-class language connoisseurs ensures that Simply Translate is the place for all your questions.

Do you want to know more about our llm-translation, nmt-translation, or the possibility to train your own datasets? Then quickly contact us or come over for a cup of coffee!