Whalesong patterns follow a universal law of human language, new research finds | Hacker News Reader

Whalesong patterns follow a universal law of human language, new research finds

theconversation.com

140 pointssohkamyung22 comments

This does not prove that whale songs are natural language, as Zipf's law is also true for random texts: https://www.nslij-genetics.org/wp-content/uploads/2022/12/ie...

Nor does it prove cultural learning, as various physical phenomema also follow Zipf's law: https://www.cs.cornell.edu/courses/cs6241/2019sp/readings/Ne...

Not just "various physical phenomena", but "physical phenomena that have stabilised and persisted structure over time".

I feel like you buried the lede in your post. People studying these physical systems are generally excited about the shared features, in my understanding

You're correct, but op is also right to point out that the headline is misleading by grouping whale song and language together—it implies that this is something unique to human language, whereas the actual category to which these now both belong is much much bigger than language.

An alternate headline that makes the point clear: "Whalesong patterns follow a universal law of human city populations". Equally true, but with much less dramatic implications.

"Whalesong patterns follow a universal law of human city populations".

Teeing up an Atlantis joke you know will get downvoted to hell on HN is just cruel man.

Note that cultural transmission of humpback whale song is well-known and studied. Here's an example article on movement of songs across Pacific ocean populations in the span of ~2 years: https://royalsocietypublishing.org/doi/epdf/10.1098/rsos.220...

(ed - in other words, cultural transmission of humpback whale song isn't part of the claims of the paper. The result is the Zipf distribution, and they hypothesize that the Zipf-law is emerging because of the cultural component.)

On the converse, human nature languages do not necessarily follow Zipf's Law. Many asian languages including Chinese and Japanese are exceptions.

I think Chomsky showed that, thanks to universality, English is a perfectly general language encapsulating all important properties of human language, so those don't count. (/s)

That Science paper

https://www.science.org/doi/10.1126/science.adq7055

ought to be retracted, it is using the same ad-hoc methods that we used to wrongly fit power laws to statistical distributions when I was a grad student 30 years ago. I wrote a trash paper too and I'm still angry about not feeling psychologically safe to bring the issue up.

A postdoc who was there wrote a paper many years later about the right way to do it

https://arxiv.org/abs/0706.1062

I don't think anybody read it. Note that the Science paper uses the "Pearson product moment correlation" which is "can only be used to measure the relationship between two variables which are both normally distributed" -- and there's nothing normally distributed here.

Ofc nobody reads papers saying your methods are trash; if they read it they simply ignore it.

Fwiw Pearson correlation does not assume normally distributed variables, what people usually mean is homoscedascity (normality of residuals) is an assumption for parametric hypothesis testing.

Cosma has a PowerPoint that may be more accessible to a general audience:

https://www.stat.cmu.edu/~cshalizi/2010-10-18-Meetup.pdf

Exactly, the thing that is "universal" about "universal power laws" is that the evidence is weak for them, not because people did a bad experiment, but because they pulled a data analysis method out of their ass.

My one regret of my scientific career was this paper

https://arxiv.org/abs/cond-mat/9512055

where my coauthor asked me "you want to be a theoretician, do you have an explanation?" but I didn't. (I did have some ideas that would have involved knowing more about the structure of the balls, doing more experiments, to give any real answer)

Actually I had reservations about the analysis method (got different answers trying two reasonable things that I saw other people to do), it would have been a good paper had I actually addressed the tension but the coworker didn't want to go there.

Another interesting thing about my paper was that segmentation was a problem as it was in that Science paper, I didn't do anywhere near the level of sensitivity analysis, but I could have gotten different answers if I'd defined integration windows around the pops differently. Probably the very loud pops also had echos, oh my god that problem was so bad when I tried to do the crumpling in an ordinary room and worse when I made my own small 'soundproof' box, it was better in the anechoic chamber I took the real measurements in.

The mystery of explaining the universality is greatly diminished when you can throw out a large number of the papers though...

Is your objection that maybe it only looks like a power law distribution for the more common "words", but if you go further out the fit isn't very good? Does it really mean anything if some other curve fits the data better?

The paper:

Inbal Arnon, et al. Whale song shows language-like statistical structure. Science 387, 649-653 (2025). DOI: 10.1126/science.adq7055

https://www.science.org/doi/10.1126/science.adq7055

Abstract: Humpback whale song is a culturally transmitted behavior. Human language, which is also culturally transmitted, has statistically coherent parts whose frequency distribution follows a power law. These properties facilitate learning and may therefore arise because of their contribution to the faithful transmission of language over multiple cultural generations. If so, we would expect to find them in other culturally transmitted systems. In this study, we applied methods based on infant speech segmentation to 8 years of humpback recordings, uncovering in whale song the same statistical structure that is a hallmark of human language. This commonality, in two evolutionarily distant species, points to the role of learning and cultural transmission in the emergence of properties thought to be unique to human language.

Relevant study of prairie dog language: https://animalstudies.org.au/archives/8211

It amuses me greatly how resistant many people are to the notion that animals who clearly communicate with each other might have something akin to "language". Or for that matter, how plants who exchange myriad signals to each other might be "communicating". Or for that matter how plants who send electrical signals between distant parts of their bodies might have something akin to a "neural network". https://pmc.ncbi.nlm.nih.gov/articles/PMC8331040/

Like most arguments it's an argument about the definition of the word. And many seem to define anything related to intelligence and communication as only applying to humans. And maybe dogs. And maybe to a lesser extent cats. And maybe certain parrots. Maybe dolphins.

I don't think anyone disputes that animal vocalizations carry information. You can demonstrate pretty directly that they get some meaning from human words. But afaik (and please correct me if I am wrong) no species other than humans have been demonstrated to have anything like a grammar.

This is an interesting way to frame the finding:

  We suggest we found these similarities because humans and whales share a learning mechanism: culture

I feel like we often have such an anthropocentric view of intelligence, learning, and culture that it’s fascinating to consider what we’re not seeing from other species just because we aren’t looking for it, especially with a growing list of potentially sapient[0] species.

  Cultural evolution inevitably leads to the emergence of properties that make learning easier.

I wonder if this idea about cultural evolution could apply in some tangential way to model distillation[1] and synthetic training data and active inheritance[2] and the gains in efficiency and efficacy we've seen lately:

While Zipf’s Law[3] might not perfectly apply to every language and also applies to some data that isn’t linguistic, it is an interesting indicator that something we don’t understand might be a language.

There was also similar research[1] into bottlenose dolphins years ago, and some researchers speculate that if we ever detected a signal from an extraterrestrial intelligence[4] it would also follow Zipf’s Law.

[0]: https://en.m.wikipedia.org/wiki/Animal_consciousness#example...

[1]: https://www.datacamp.com/blog/distillation-llm

[2]: https://www.deeplearning.ai/the-batch/active-inheritance-a-s...

[3]: https://en.m.wikipedia.org/wiki/Zipf%27s_law

[4]: https://www.seti.org/animal-communications-information-theor...

In its Dr Doolittle implications the article is comforting to read but not convincing as a scientific discovery; this "law" seems to be of the same class as things like the Pareto (80/20) principle and possibly more abstract and nebulous than attributable to some shared intellectual co-living space with our fellow animals.

Even if whalesong is actually a language, translation to human language could prove challenging if Wittgenstein was correct that meaning is use. Whales won't be using sounds the same way we do. Thus the "If a lion could speak, we would not understand him" quote.

It's a problem we'll face if an extraterrestrial signal is ever detected. We can hope they'll use familiar mathematical patterns to help us translate like in Contact, but we can't be sure about that. They might be too different from us physiologically, culturally and technologically.

Stanislaw Lem’s novel “His Master’s Voice” is about just such a project, where it proves impossible to translate the alien signal.

It's interesting that most animal to human communication seems to be humans training animals to speak human.

Took me a second that this is not about higher order translation but human to non-human.

Do HN post categories have an equivalent exponential frequency distribution?

Reply on news.ycombinator.com