How to Talk to Whales

A fascinating research organisation is using AI to decode animal communications

Image by Thomas Lipke on Unsplash

Dear Earth Species Project,

I am writing this post for my 25 followers so that 25 more people may know that it is theoretically possible to understand Whale.

I was going to wait to write this until I was a doctor on the matter. I had already come up with half of the title for my Ph.D. thesis. It was the fun, pop-culture-inspired half, the half academics like to drizzle here and there to demonstrate that their work is consumable and relevant, or just more pleasant to read than the everyday slog through papers with sentences like “Heidegger’s conception of dwelling is both building and care that allows for continued poetic dwelling.” What does that mean? I don’t know. I am finally at an age where I feel OK admitting that I don’t know because maybe someone will tell me and relieve me of the burden of pretending that I’ve read anything by Heidegger.

As for the title of my thesis, it went like this: “Wow. I wish I could speak whale: The role of Natural Language Processing in mapping conceptual space. Or discourse analysis. Or interspecies communication.” I haven’t quite worked out the second half yet, but the important thing is I have the first part down so that when it’s finished and I post my thesis online, my 25 followers will know that it is consumable and relevant. Until that is, I start talking about Heidegger.

It may, however, be a while before I finish my Ph.D. Or start it, for that matter. So I’m just going to go ahead and write about your work now because, in my opinion, it is among the most world-changing endeavors human beings are currently attempting.

Here’s the part where I explain what you do

You, Earth Species Project, are a team of computer scientists, biologists, and dark matter physicists. Collectively of course you know exactly what your aims are and how exciting new research in the field of Natural Language Processing makes them even more achievable. However, for those of us without a degree in computer science, my explanation of what you are trying to do would go something like this:

First of all, we turn a human language into a shape — or model.

"An example of a geometric representation, an ‘embedding,’ of the top 10,000 most spoken words in English. While each word is actually represented by a point in hundreds of dimensions, here it is projected down to three for visualization." The Earth Species Project at https://github.com/earthspecies/project/blob/main/roadmaps/ai.md

See this lovely revolving graphic? That’s a representation of a model of the top 10,000 most common words used in English. The model has hundreds of dimensions and hence cannot be printed into physical space and exists only as a mathematical construct. Each of the tiny dots in the revolving graphic represents an English word, which is each represented in the model as a number — essentially their coordinates in the landscape of the 10,000 most common English words. In 2017, two very exciting papers were published that demonstrated it was possible to translate between languages by comparing their respective shapes — just take the model of English, slip the model of Urdu (for example) on top of it, and rotate until they more or less match up. Now, the dot representing the English word “king” will be in the same place as the dot for the Urdu word for “king”, and the English word “Geezer” will be in the same place as the closest Urdu equivalent.

Now imagine that you can do this for Whale.

Technical insert:

If you do any googling of phrases such as Natural Language Processing you will probably encounter the explanation that it involves the transformation of linguistic components into vectors. A word vector is the list of numbers that represent each word within a mathematical model describing a corpus of text; Wikipedia, the Complete and Unabridged works of William Shakespeare, IMBD movie reviews, and so on. The process of producing these vectors is called ‘word embedding’ and there are many different machine-learning algorithms available with which to do this. See here for an illustrated and very comprehensible explanation of the word embedding algorithm word2vec, and here for an illustrated and more maths-heavy explanation of GloVe and BERT; two neural network models that create word vectors in a manner that accounts for the context of the word and fancy things like polysemy, which AI algorithms have previously been rather bad at.

This method — of translating between languages by comparing their n-dimensional models — is the technique used by modern machine translation tools — like Google translate.

We could seriously, actually understand Whale

Just think of the possibilities! A linguistic model is a conceptual model, a geometric representation of a worldview. Creating a model of Whale would allow us, humans, to enter a completely alien mind-space and access the foundational concepts of an intelligent non-human species. Cetaceans including humpback whales, grey dolphins, and Orcas all have distinctive sound combinations used to call specific individuals, in other words — they have names. Do they also have language for family relationships? Or gender? Does their language operate with opposing pairs like mother-father and male-female — or do whales use multiple conceptual classes and avoid binary thinking altogether? Author Ursula Le Guin created a fictional androgynous society where avoidance of binary thinking meant avoidance of societal war. Is speaking Whale the secret to world peace? Or do Whales indeed have violent, multi-generational wars that manifest not as whale-on-whale violence (which we probably would have noticed) but sustained tonal assault?

The things we could learn if we learned to speak Whale.

I get that it’s not completely straightforward

There is the issue of Unwelt — a delicious word meaning the world as it is experienced through a specific set of senses. The umwelt of a mantis shrimp is polarized; their famous eyes equipped with 16 color receptors (compared to three, in humans) naturally detect circularly polarized light — and they use this ability to detect secret shrimp-to-shrimp communications. Umwelts are encoded into language; studies of color perception between different human cultures indicate that how people see an object or landscape or fellow living thing is shaped by the words they have available to describe it. The possibility of translating between the world experiences of different creatures is intoxicatingly exciting for the insight it offers into that very world experience — but it is also challenging because some perceptions of reality may be just too different to ever make sense to each other. I suggest it would be difficult to chat with a mantis shrimp.

As you say; however, “even with cetaceans there are reasons to be hopeful; We share being social mammals living in the same physical world, we are live-born, breath to live, have families, relationships, personalities, and complex evolving cultures.”

Your technical roadmap, at the Earth Species Project, includes determining the extent to which inter-species realities are comparable. To do this, you are endeavoring to identify where meaning is encoded (what part of the Whale sounds constitute the words), and to separate out individual voices from the chattering of the crowd. Algorithmically generated language models are being used at all stages of this process. Your work is public and has attracted the attention of Machine Learning developers and linguistics across the world. I look forward to delving deeper into your technical content — and I will even overcome the imposter panic I feel every time I use GitHub (what are you doing here? You’re not a computer scientist! You didn’t even finish maths at high school!) to do so.

I may report back, to my 25 followers, where research from the Earth Species Project is currently at.

Now for the conclusion

I will finish this post with an additional point that I think is important to keep in mind. Popular fantasies emphasize talking to animals. Cinderella sang to birds and they made her a dress. A nice lady on the internet called ‘Val’ will charge you $27 for an ebook so that you too can send messages to your cat.

What about the animals who have messages for us? Or the animals who have absolutely no interest in speaking to humans at all and prefer to carry on with their ancient, enduring civilizations that have avoided rending the planet inhospitable for large numbers of carbon-based lifeforms.

If we learn to speak Whale then maybe we can finally keep quiet and listen.

So full steam ahead, dear Earth Species Project. 25 more people should now be thoroughly impressed and more aware of what you do.