SPOTLIGHT ON… KILIAN SEEBER

Kilian G. Seeber is Professor at the University of Geneva’s Faculty of Translation and Interpreting. He was the Director of the Interpreting Department until 2022 and joined the FTI’s leadership team as Vice Dean in 2018. Kilian’s main research interests are the cognitive aspects of interpreting, with a particular focus on multimodal processing. As director of two research labs, Kilian has been harnessing new technologies both in research (LaborInt) and training (InTTech). He has organized training courses for the European Institutions and the United Nations and has co-developed the Master of Advanced Studies in Interpreter Training of which he is the program director. In this e-bulletin, he tells us more about interpreting in the age of Artificial Intelligence (AI).

Lisez cet entretien en français

Professor Seeber, the media are abuzz with reports projecting the demise of entire professions, thanks to the rapid progress of Artificial Intelligence (AI). Language professions are a case in point: they are often said to be dead. As the Faculty’s Vice Dean, what is your take on these doomsday prophecies?

In keeping with Mark Twain, I am tempted to say that the rumors about the death of language professions are greatly exaggerated. It's true that recent progress in the application of large language models (LLMs) to chatbots has strengthened the illusion of conscious artificial intelligence. However, the truth is that we are still talking about machines, and therefore tools. These tools mimic human speech; they are increasingly autonomous and capable of processing large amounts of data in very little time. Consequently, they have a lot of potential when it comes to facilitating many of the tasks involved in various language professions. That said, such tools are prone to perpetuating mistakes, depending on the dataset used to train them and the prompts used to interact with them. While I don’t see AI tools replacing human intelligence in many language professions, it is neither realistic, nor perhaps desirable, to reverse these developments. Instead, we need to be asking critical questions about where, when, and how we could and should integrate them into the process to make it more efficient and more effective.

Your main expertise is in conference interpreting. It is sometimes said that translators have been more open to technology than interpreters, readily integrating these tools into their workflow. Is that true, and is it fair to say that conference interpreters are being overtaken by events?

It is accurate to say that translators adopted new tools such as translation memories and machine translation into their workflow long before interpreters. That is, however, a bit of a truism. One of the fundamental differences between translation and interpreting, especially simultaneous, lies in their temporal mechanics. It boils down to translators having time to draft, redraft, edit and correct. Time is the one luxury interpreters don’t have. Only very recently has technology produced tools that can be integrated into a workflow that requires interpreters to make decisions in a matter of seconds, with little or no time to make corrections. What's more, translators work with written text, which is much more rule-based than spoken discourse. Not only is the latter not nearly as well-formed as written text, but it is also highly contextual, and even something like an accent, native or otherwise, can be a major hurdle for a tool. Only recently have AI-powered tools for speech recognition and translation become fast and reliable enough to bring added value to the interpreting process.

You mention the time constraints interpreters work under, and the fact that any tool would have to be both fast and reliable to help rather than slow them down. What would such tools look like? And are these tools already available, or are they still more fiction than science?

We still don’t have a fully comprehensive model of the simultaneous interpreting process, but we do know of some factors that tend to make the process more difficult – things that require more processing capacity and things that slow down processing. These include proper nouns, numbers, abbreviations, acronyms and highly technical terms. Because their processing is often mainly signal-driven, as opposed to context-driven, it is conceivable that AI tools could facilitate the interpreter’s job by identifying them and, where appropriate, perhaps even suggesting a translation. This is no longer science fiction and pilot projects are currently underway at several international organizations to test their viability in real-life settings.

You say these systems are already being tested. Are you and your research team working in this area at all?

We are indeed. Rather than looking at the usability side of things, however, we are interested – and this is probably shaped by our tradition and expertise – in the effects integrating such technology into the workflow has on the interpreting process. I mentioned the possibility of « augmenting » the interpreting process by offloading the recognition and perhaps even the translation of certain constituents like proper nouns, numbers, and acronyms to the machine. At this point, we know very little about if and how interpreters can integrate this information into their workflow. In fact, it is not enough to know that such constituents are correctly uttered. We want to see whether this augmentation requires a disproportionate amount of attention that would negatively affect the overall interpretation. In other words, while an AI tool might be able to correctly identify a number, that number remains meaningless if the interpreter is unable to put it in a sentence that makes sense and accurately reflects the original. The potential cost and benefit of integrating these « augmented » components into the simultaneous interpreting process is what we are currently studying.

What about training? Do you see any way AI tools can help train conference interpreters?

In fact, I see more than one. We already use AI-powered tools in our MA in Conference Interpreting, where we recently introduced a seminar series dedicated exclusively to the use of new technologies. As LLM-based chatbots were designed to imitate the way humans communicate, these tools can be used to create training materials for interpreters. They can also be used to streamline the preparation process by extracting terminology from background documents or by summarizing them. Integrating these tools into our training program serves the dual purpose of teaching students how to interact with them effectively (as results vary greatly depending on the prompt used) but also to become critical users who understand both the potential and the limitations of the new tools. As for the use of AI in the booth, we need to understand better how it impacts the interpreting process: once we have reliable research findings we will feed them back into our training approach.