Defending Compositionality in Emergent Languages
This addresses a foundational problem in AI and linguistics by clarifying the role of compositionality in emergent languages, though it is incremental as it refutes recent claims rather than introducing a new paradigm.
The paper tackles the debate on whether compositionality is essential for generalization in artificial neural networks, showing that compositionality is indeed crucial for successful generalization when evaluated on a proper dataset in a two-agent communication game.
Compositionality has traditionally been understood as a major factor in productivity of language and, more broadly, human cognition. Yet, recently, some research started to question its status, showing that artificial neural networks are good at generalization even without noticeable compositional behavior. We argue that some of these conclusions are too strong and/or incomplete. In the context of a two-agent communication game, we show that compositionality indeed seems essential for successful generalization when the evaluation is done on a proper dataset.