Classifying evolutionary forces in language change using neural networks.

Folgert Karsdorp,Lauren Fonteyn,Enrique Manjavacas,Mike Kestemont

doi:10.1017/ehs.2020.52

Abstract

A fundamental problem in research into language and cultural change is the difficulty of distinguishing processes of stochastic drift (also known as neutral evolution) from processes that are subject to selection pressures. In this article, we describe a new technique based on deep neural networks, in which we reformulate the detection of evolutionary forces in cultural change as a binary classification task. Using residual networks for time series trained on artificially generated samples of cultural change, we demonstrate that this technique is able to efficiently, accurately and consistently learn which aspects of the time series are distinctive for drift and selection, respectively. We compare the model with a recently proposed statistical test, the Frequency Increment Test, and show that the neural time series classification system provides a possible solution to some of the key problems associated with this test.

Highlights

To study the mechanisms underlying cultural change, detailed information is needed about the complex mix of, for example, cognitive, social, and memory-based biases of individuals that bring about a certain change
We show that the neural networks are affected less by distortions of the time series compared with the Frequency Increment Test’ (FIT), making the method more applicable to the noisy, sparse and incomplete data that we often find in historical collections of cultural data
A machine learning approach As an alternative to the FIT, we propose to conceptualize the task of detecting evolutionary forces in language and cultural change as a binary time series classification (TSC) task

Summary

Introduction

To study the mechanisms underlying cultural change, detailed information is needed about the complex mix of, for example, cognitive, social, and memory-based biases of individuals that bring about a certain change. By investigating divergences between simulated and real-world frequency distributions (Bentley et al, 2004, 2007; Hahn & Bentley, 2003; Herzog et al, 2004; Ruck et al, 2017) or turnover rates (Acerbi & Bentley, 2014; Youngblood, 2019), or applying likelihood-free inference techniques (Carrignon et al, 2019; Crema et al, 2014, 2016; Kandler & Powell, 2015; Kandler & Shennan, 2013, 2015; Lachlan et al, 2018), arguments for the presence of individual-level biases underlying cultural change have been made, such as conformity bias in bird song (Lachlan et al, 2018) and music sampling traditions (Youngblood, 2019), or anti-conformity bias in archaeological pottery data (Crema et al, 2016). Knowledge of such biases operating at the individual level is crucial to better understand how they ‘can affect the populational profile of a collection of ideas, skills, beliefs, attitudes, and so forth’ (Lewens, 2015, p. 57, and see Boyd & Richerson, 1985; Cavalli-Sforza & Feldman, 1981; Richerson & Boyd, 2005)

Objectives

Methods

Results

Conclusion