A new method for computational cultural cartography: From neural word embeddings to transformers and Bayesian mixture models

Abstract

Recently, computational social scientists have proposed exciting new methods for 'mapping meaning space' and analysing the structure and evolution of complex cultural constructs from large text datasets. These emerging approaches to 'cultural cartography' are based on a foundation of neural network word embeddings that represent the meaning of words, in relation to one another, as vectors in a shared high-dimensional latent space. These new methods have the potential to revolutionize sociological analyses of culture, but in their current form, they are dually limited. First, by relying on traditional word embeddings they are limited to learning a single vector representation for each word, collapsing together the diverse semantic contexts that words are used in and which give them their heterogeneous meanings. Second, the vector operations that researchers use to construct larger 'cultural dimensions' from traditional embeddings can result in a complex vector soup that can propagate many small and difficult-to-detect errors throughout the cultural analysis, compromising validity. In this article, we discuss the strengths and limitations of computational 'cultural cartography' based on traditional word embeddings and propose an alternative approach that overcomes these limitations by pairing contextual representations learned by newly invented transformer models with Bayesian mixture models. We demonstrate our method of computational cultural cartography with an exploratory analysis of the structure and evolution of 120 years of scholarly discourse on democracy and autocracy.