Here’s a curious paradox related to American Sign Language, the system of hand-based gestures used by around 2 million deaf people in the US and elsewhere to communicate.
Almost 40 years ago, researchers discovered that although it takes longer to make signs than to say the equivalent words, on average sentences can be completed in about the same time. How can that be possible?
Today, Andrew Chong and buddies at Princeton University in New Jersey give us the answer. They say that the information content of the 45 handshapes that make up sign language is higher than the information content of phonemes, the building blocks of the spoken word. In other words, there is greater redundancy in spoken English than signed English.
In a way, that’s a trivial explanation, a mere restatement of the problem. What’s impressive about the Princeton contribution is the way they have arrived at this conclusion.
The team has determined the entropy of American Sign Language experimentally, by measuring the frequency of handshapes on video logs for deaf people uploaded to youtube.com, deafvideo.tv and deafread.com as well as from video recordings of signed conversations taken on campus.
It turns out that the information content of handshapes is on average just 0.5 bits per handshape less than the theoretical maximum. By contrast, the information content per phoneme in spoken English is some 3 bits lower than the maximum.
This raises an interesting question. The spoken word has all this redundancy for a reason: it allows us to be understood over a noisy channel. Lessen the redundancy and your capacity to deal with noise is correspondingly reduced.
Why would sign language need less redundancy? “Entropy might be higher for handshapes than English phonemes because the visual channel is less noisy than the auditory channel…so error correction is less necessary,” say Chong and co.
They go on to speculate that signers cope with errors in an entirely different way to speakers. “Difficulties in visual recognition of handshapes could be solved by holding or slowing the transition between those handshapes for longer amounts of time, while difficulties in auditory recognition of spoken phonemes cannot always be easily solved by speaking phonemes for longer amounts of time,” they say.
And why is all this useful? Chong and friends say that if sign language is ever to be encoded and transmitted electronically, a better understanding of its information content will be essential for developing encoders and decoders that do the job. A worthy pursuit by any standards.
Ref: arxiv.org/abs/0912.1768: Frequency of Occurrence and Information Entropy of American Sign Language
This new data poisoning tool lets artists fight back against generative AI
The tool, called Nightshade, messes up training data in ways that could cause serious damage to image-generating AI models.
Rogue superintelligence and merging with machines: Inside the mind of OpenAI’s chief scientist
An exclusive conversation with Ilya Sutskever on his fears for the future of AI and why they’ve made him change the focus of his life’s work.
The Biggest Questions: What is death?
New neuroscience is challenging our understanding of the dying process—bringing opportunities for the living.
Driving companywide efficiencies with AI
Advanced AI and ML capabilities revolutionize how administrative and operations tasks are done.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.