←back to thread

189 points docmechanic | 4 comments | | HN request time: 0.473s | source
1. smusamashah ◴[] No.43655793[source]
> They recorded over 300 of these observations, including what the caller was doing at the time, what was happening in the environment and the behaviour of the caller and audience after the vocalisation.

> To reveal the meaning of each call, they used a technique from linguistics to create a cloud of utterance types, placing vocalisations that occurred in similar circumstances closer together. “We kind of established this dictionary,” says Berthlet. “We have one vocalisation and one meaning.”

This is lots of manual effort, could the recent advancement in language models help decode animal languages more easily? I guess it will need lots 24/7 capture of physical movement/action and sound data and train a model (that already understands vocal English too) perhaps.

replies(3): >>43655947 #>>43656408 #>>43657163 #
2. bbor ◴[] No.43655947[source]
I definitely think you’re touching on some exciting possibilities, but adding a language model at this early stage would endanger the goal of this particular research: proving that the compositionality exists in the first place. If there was a foundational language model we’re involved, it might be reading patterns into the calls regardless of whether they’re really there — that is what it’s designed to do, after all!

Re:”lots of work”, I think you’re misunderstanding the quotes a bit. They applied PCA to categorical data to generate semantic positions for each call type—or, in other words, ran a prewritten mathy algo on a big csv. Collecting the CSV data in the first place certainly sounds extremely hard, but that’s more of a practical issue than a scientific one! Bonobos aren’t known for living in easy-to-reach places ;)

3. HappMacDonald ◴[] No.43656408[source]
If you make that cloud exist in a high enough number of dimensions you'll find yourself emulating a machine learning language model. :)
4. keybored ◴[] No.43657163[source]
Making models of the physical world is a lot of work. Can’t they install cameras and record hundreds of thousands of hours of objects getting shot through cannons, birds flying and trees swaying in the wind? Maybe with some more nuclear power plants they could get close to approximating something like Newton’s Laws of Motion (kind of close is good enough, no need to be nerdy about it).