This is not just distributional information analysis in the sense that ‘tokens’ are grounded in other ‘tokens’. They’ve grounded these calls in naturalistic situational context. This is hard won data.
If I understand, the finding here is that bonobo calls are “non-trivially” compositional, e.g., the semantic embeddings of pairs of vocalizations point in different directions surrounding the base vocalization. But it seems there is no “trivial” compositionality in the sense that constructions like [good __] might point in a similar direction. I would expect this latter result. This seems like a conspicuous absence? Is this really compositionality? Not sure what to make of it.
Some interesting context: bonobos and other (non-human) great apes are believed to have more intentional and flexible control over their gestural repertoire than their vocal repertoire and that these gestural repertoires are larger. Human language likely evolved from gesture (or so some believe). So, if their vocalizations are in fact compositional, it may be a separate evolutionary prong.