METM23 Chronicles: Luisa Bentivogli

Large language models and language professionals: understanding the promises, risks and impact

METM23 was the first of our annual meetings since the appearance of ChatGPT in November 2022, and the subject of artificial intelligence (AI) was widely discussed among the attendees. Luisa Bentivogli’s keynote talk on large language models (LLMs) therefore provided a very welcome contribution to our knowledge of the subject. In her clear and thorough presentation, she explained how LLMs work and the threats and possibilities that they entail.

Luisa first explained the basis of generative AI systems, which use massive amounts of data, train them to create “foundation models”, and then adapt them to perform tasks and generate new content. LLMs are foundation models that use text and text-like material. The structure of the language and the knowledge contained in the texts is encoded, and the models are then fine-tuned to follow instructions and carry out tasks such as translation, question-answering and text classification. They also learn to perform instructions for which they are not trained. The results are refined by human feedback, and the training can continue with smaller, more specific datasets for a new domain or task.

With regard to translation, Luisa explained that LLMs do not yet match the results of traditional machine translation engines at the sentence level but offer great potential for document-level translation because they are able to use context. They are currently English-centric because they are trained on monolingual (mostly English) input, so performance into English is better than in the reverse direction, and it is better to use English as the input language even if we are seeking results in other languages. LLM translations are prone to errors such as omissions, “hallucinations”, copying and mistranslation, so the results have to be fact-checked before they can be used.

Despite the impressive performance of generative AI, its “emergent abilities” can involve risks. Mistakes can have consequences for entire systems. There is a risk of bias, discrimination and toxic content. AI systems can be used to create disinformation, and their human-like interaction can allow them to persuade and influence users. The legal issues that arise include possible infringement of copyright (if you are concerned, tell the system not to save your conversations) and the fact that it is not yet clear who owns, and is responsible for, the outputs. Finally, a major drawback of these systems is that they use a colossal amount of energy.

With regard to the implications of AI for the translation industry, Luisa stated that human translators are highly exposed. They will perhaps not be replaced, but the number of them needed will probably fall, and/or the tasks they perform will change. Less translation and more editing will be required, as will more translation of texts summarized by generative AI. Humans will always be needed to create, curate, adapt and evaluate in the context of AI, but they will require new specializations and new skills. For her closing slide, Luisa had asked ChatGPT about the future of language professionals. Its response was “…the future of language professionals in the LLM era is a path of transformation, not obsolescence… Stay adaptable, committed to quality, and culturally sensitive as we navigate this evolving landscape.”

This METM23 keynote talk was chronicled by Alan Lounds.

Featured photo by METM23 photographer Leonardo Rizzato.

Leave a Reply