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 re

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 ;)