A Functionalist Framework for Machine Learning in Animal Communication

Recent breakthroughs in natural language processing inspire optimism that similar methods could decode animal communication systems. But machine learning approaches import assumptions from human language, which could undermine these efforts. In this proposal, we argue that non-human animal communication systems do not have self-contained distributional semantics, are largely non-referential, and function primarily to manipulate the behavior of others rather than exchange information. Not only do these assumptions constrain our ability to investigate signal semantics, but also risk confounding discoveries of signal syntax. To hedge against this possibility, we propose that machine learning efforts should adopt a functionalist framework. This foregrounds ecological and social contexts and the interactional contingencies that give signals their meaning. Our framework provides recommendations about how to account for these variables when building datasets.