Teleosemantics for neural word embeddings

This paper applies a consumer‐based teleosemantic framework to give a detailed analysis of a particular algorithm for generating word embeddings. In the process, it addresses several of the challenges facing teleosemantic approaches to artificial neural networks. It distinguishes the content of embeddings from the contents of their features and argues that word embeddings should be seen as pushmi‐pullyu representations whose descriptive contents are facts about the contexts in which words appear. Furthermore, it argues that embeddings meet Millikan's conditions for iconic representation as transformations between embeddings correspond to transformations of the contents of their representations.