Like an Apostroph chatbot?
R: Yes, focussing on translation and proofreading, specifically relating to the services we offer here at Apostroph. While more general models like ChatGPT could actually be useful to us in our world of language, they still can’t be relied upon to deliver the accurate linguistic results we need. It's not a case of one model fits all. So we’re working on a custom solution of our own.
How are things looking in terms of data security?
R: This is such an important topic. The data security infrastructure here in Switzerland is forever evolving. We keep on adding to our capacity for internal storage solutions here.
S: We view it as absolutely essential that any data we process and any systems we develop remain right here in Switzerland at all times. We take extreme care to keep all the data we process safe and secure – and that includes data generated using AI.
What’s going on with the likes of machine translation and translation memories?
S: Machine translation is an interesting area to follow at the moment because lots of people are experimenting and some are even trying to bring about a paradigm shift, with large language models (like GPT) being used for automated translation instead of machine translation models that are specifically trained. We are keeping a close eye on these developments and looking into the results and implications of an approach along these lines.
R: One of the biggest problems with multilingual models is that there is conflict between different languages within a model when it comes to parameters. This means that the training data for different languages is out of balance, which can result in poorer results being delivered for less common languages. A new innovation published by Apple is exclusive model parameters that apply to a single language. It just goes to show that there is still plenty of scope for innovation in machine translation. I know that I’m excited to see how good MT can get.
Looking into your crystal balls, what do you think the future holds for IT in the world of languages?
S: I hope that models for text comprehension and generation continue to improve. They are becoming increasingly intelligent and learning to solve complex problems using examples. A single Shakespeare poem will be enough of a prompt for more poems to be written in the style of Shakespeare.
I hope we see more of the new efficient ways of training large language models so that a development process no longer takes a single person 100 years. This will open things up beyond the realms of tech giants.
I also think that our devices are only going to become more powerful, presumably allowing many language models to be run locally rather than just in US-based clouds. It’s possible that we will soon start to see more end-to-end translations, like audio-to-text, audio-to-audio and image-to-text, without complex pipelines consisting of multiple models. We’d be excited to see the rise of simpler, more elegant end-to-end pipelines trained using high-quality data.
R: The role of our linguists will shift towards text managers. They will need to manage a number of technology systems to create linguistic processes. Translations and texts generated by machines will need less and less linguistic input. Linguistic expertise and cultural sensitivity will go hand in hand with an understanding of technology.
