As I have often observed before, when it comes to language, people seem to think they can just make stuff up. Even patently absurd claims about language elicit no fact-checking. Editors limply allow any kind of nonsense to flow into print. I’ve pointed this out in journalism about artificial intelligence and robots, and especially about language in gorillas and orcas and dolphins and dogs and various other species. But in all my years of reading such absurd claims, I’ve never seen anything like the recent article by Charles Hymas, published on September 16 in the National Post, a fairly conservative Canadian newspaper. (The Telegraph also ran the story in Britain, behind a paywall — they will charge you money for this balderdash!)
Hymas reports that “two of the original architects of social media,” namely Aza Raskin (a feature designer and programmer who invented the endless scroll) and Britt Selvitelle (an early engineer at Twitter), are on the brink of “using artificial intelligence to create a translation system that will understand what animals are saying and what they think of us.” Raskin and Selvitelle claim to have “analyzed 70 human languages to establish that all have a universal ‘shape,’ such that a computer can translate one into the other without any prior understanding or knowledge” (emphasis added).
They plan to “add to their 50,000 hours of humpback-whale recordings by setting up a huge array of microphones in the Congo to log elephants’ communications” so that they can “compare the architecture of animals’ communications with human language, from which they aim to create a modern-day AI Rosetta Stone to decode and translate what animals are saying.”
Raskin thinks this “would have profound implications for our judiciary system and the way we pursue our role as stewards of this tiny, pale blue marble that we call home.”
I fear he has lost his tiny, pale blue marbles. There are two known ways to attempt the translation of a sentence S from language L1 to language L2:
- Learn some L2 (and, of course, make sure you know L1 thoroughly), and using an L1-to-L2 dictionary and a description of the grammar of L2, figure out how to express S in L2. (This calls for humans. Simulating it on a computer is too difficult for present capabilities.)
- Ignore meaning and grammar completely; just build a giant multi-billion-word database of paired pre-existing L1-to-L2 translations, aligned sentence by sentence, and use heavy statistical computation to figure out for each substring in S the L2 substring that is most likely to be aligned with it. Then put all the information together and, using the computed substring probabilities, find the string of words most likely to correspond to the whole of S. (This calls for powerful computers. It relies on vast quantities of storage for parallel bilingual texts, and massive high-speed statistical computation.)
Raskin and Selvitelle apparently dispense with both. They don’t need to “know either language” (that rules out method A) or “have any examples of how to translate between the two” (which rules out method B).
Translation between arbitrary human languages without either knowing them or training on text? Really? Douglas Adams’s fantasized babel fish was more sensible than this.
And the method can be extended directly to nonhuman animal species without any prior understanding of their cognitive powers or communicative practices? Are we being serious? Hymas thinks so.
He adds that “Linguistics experts have previously predicted that human-animal translation could be cracked within a decade as research teams, from Sweden’s Royal Institute of Technology to Northern Arizona University, began deploying the AI technology to understand animal communication.” The reference to Northern Arizona University must pertain to Con Slobodchikoff’s work on prairie-dog behavior; it is interesting, but suggesting that it involves anything like human-language use goes way too far (I explained my skepticism in this post).
The Swedish reference probably alludes to recent claims about a start-up company called Gavagai that plans to try using AI to understand dolphins, which (from all that I can tell) is as quixotic as the research plan Hymas describes. (And excuse me, but dumping a corpus of doplphin squeals into a machine-learning program to see what happens does not turn people into “linguistics experts.”)
I remain optimistic about scientific work on language one day achieving things that will actually matter for practical purposes, so I find it discouraging to see crazy-talk stories concerning animal language being repeated in major news sources. It erases the line between doing sensible research and merely tickling the fancy of animal lovers at the dimwitted end of the general public’s intelligence cline. But at present, cynical nonsense-mongers catering to the gullible too often have the floor.