Machine translation has gone from a laboratory experiment to something huge numbers of people use every day, often without a second thought. But how far can it be trusted, and what does it mean for people who translate for a living? For this month's debate, our presenter, Tom Ellery, talked to Dr Ruth Okafor, a computational linguist, and Daniel Marsh, a professional translator based in Leeds.
Tom: Ruth, let's start with the basics. How do these systems actually work? Ruth: The earliest attempts, back in the 1950s, were rule-based. Linguists tried to write down the grammar and vocabulary of two languages and tell the computer how to convert one into the other. It was painstaking work, and the results were often comical. Later, systems became statistical: instead of following rules, they compared enormous collections of texts that humans had already translated and calculated which translation was most likely. Today's systems use neural networks, which learn patterns from vast amounts of data. The leap in quality has been remarkable. For many common language pairs, the output is now fluent enough that a casual reader might not realise a machine produced it. Tom: Daniel, as a translator, does that worry you?
Daniel: Fluency is precisely what worries me. Older systems produced obvious nonsense, so you knew not to rely on them. Modern ones produce sentences that read beautifully but can be quietly wrong. A plausible mistake is far more dangerous than an absurd one, because nobody spots it. I once checked a machine-translated safety notice for some garden machinery in which a single 'not' had vanished, so the instruction meant the exact opposite of what was intended. Ruth: That's a fair point, and it's why I'd never use raw machine output where the stakes are high – in a hospital, a courtroom or a legal contract.
Tom: What are the systems worst at? Daniel: Anything that depends on context, culture or tone. Humour, wordplay, poetry, the difference between friendly and curt. Languages also divide up the world differently. Some have several words where English has one; some have no grammatical gender while others mark it everywhere, so the machine has to guess. When it guesses, it tends to repeat the bias in its data – making doctors male and nurses female, for example. Ruth: Developers are working hard on that, though I agree it's a real problem. There's another imbalance, too. These systems learn from existing translations, and for widely spoken languages there's a vast amount to learn from. For many smaller languages there's very little, so quality drops sharply. Ironically, the communities that might benefit most are often the ones served worst.
Tom: So is the professional translator on borrowed time? Daniel: I don't think so, but the job is changing. A lot of my work is now post-editing: I take a machine draft and correct it. For routine technical documents, that's genuinely efficient. For a novel or an advertising campaign, it can take longer to fix a draft than to start from scratch, because the machine irons the nuance out of everything. Ruth: I'd add that the technology has opened doors as well. A tourist can read a menu, a family new to the country can understand a letter from their child's school, a scientist can skim research published in another language. That's not trivial.
Tom: Final thoughts? Ruth: Treat it as a remarkably powerful tool, not an oracle. Use it to get the gist, then check anything important with a human. Daniel: I'd agree with that – which, coming from a translator, may surprise you. Machines are very good with words. Understanding what someone actually means is still, for now, a human skill.