This project contains code in the form of an R package for modelling bilingual-led divergence. It implements a cognitive model of bilingual lexical selection, and then builds community-level agent-based modelling on top of this implementation.
The software takes the form of an R package, called bldR.
This software is companion to the following publication, hereafter called EM:
- Ellison, TM and L Miceli (2017) Language Monitoring in Bilinguals as a Mechanism for Rapid Lexical Divergence. Language 93(2):255-287. (Once the embargo is lifted, this article will be made publicly available. Before then, copies can be requested from the authors.)
The package contains all new data reported in EM, and all software used in the simulations, data-processing, statistical testing and graph-drawing.
The code consists of three parts:
- the cognitive simulation, implemented in the R6 class TensorAgent,
- the population simulation, implemented in the R6 class PopulationSimulation, and
- the specific simulations run for EM, including the ggplot2 code to create its graphs, as implemented in the R6 class L2017.
Also included in the package are three data-sets:
- EM2017_Stimuli contains the stimuli used in the experiment reported in EM,
- EM2017_Responses contains the raw results of the experiment,
- EM2017_ParameterFit contains the probability the model assigns to the experimental data for 10201 parameter settings.
You can easily install this package from github directly. Here are the steps (see this page for more details):
-
You need a suitable development environment on your computer: enough so you can run the package devtools. The readme for this package (just follow the link and scroll down) tells you how to setup this environment on various platforms (Linux, OSX, Windows). It's quite easy.
-
Install the devtools package, and the other libraries used by \code{bldR}. For example, start up R and run:
install.packages("devtools")
install.packages("ggplot2")
install.packages("tensorA")- Now load the devtools library:
library(devtools)- Then install this R package:
install_github("tyrannomark/bldR")Here are some examples of code using the classes defined in the package. Other examples are given in the R package documentation.
Here is a very simple model of a bilingual lexical production in a context where English (EN) can use either picture or photo (spelled foto to indicate to the software the identity with its Dutch doppel) but prefers photo, and where in the corresponding context, Dutch (DU) would only use foto.
We assume that speakers are in bilingual mode - i.e. somewhat prone to intrusions from their non-intended language if they don't monitor - at a level of 0.5. To minimise these intrusions, however, they engage in monitoring at a reasonably strenuous level, namely 0.7.
require( bldR );
ta <- TensorAgent$new()
ta$clearLexicon();
ta$addExample("PHOTO","EN","picture", ct=40);
ta$addExample("PHOTO","EN","foto", ct=60);
ta$addExample("PHOTO","DU","foto", ct=100);
ta$constructDataTensor();
ta$setLanguageMode(0.5);
ta$setMonitoringLevel(0.7);
ta$make_p_f_st__bm();
print(ta$as.data.frame());
Note that the production probability of foto is reduced for bilinguals when speaking English, and probability of using picture is increased.
In this example, let's imagine a microcommunity with 20 monolingual English speakers, 30 monolingual Dutch speakers, and 40 speakers who are bilingual. The bilinguals have the same level of bilingual mode and monitoring as in the above example. Let's look at what happens to the use of alternative forms in the community over time. Since only English has initial variation, this is the only case in which we will see change.
Each agent in each generation produces a sample of 50 forms, based on their inputs and biases. These samples as a whole are used as the input (divided by language of course) for the next generation.
Bilinguals are assumed to be taking input from, and producing output in, each of their languages equally.
require( bldR );
p <- PopulationSimulation$new();
p$setPopulationStructure(A=20,B=30,AB=40);
p$setLanguageMode( 0.5 );
p$setMonitoringLevel( 0.7 );
p$setSamplesPerAgent( 100 );
p$setNumberOfGenerations( 30 );
p$setLexicon(A_d=0.4,A_nA=0.6,A_nB=0.0,B_d=1.0,B_nA=0.0,B_nB=0.0)
history <- p$simulate(exact = FALSE);
historyAd <- history[history$Language == "A" & history$Form == "d",]
doppelUse <- aggregate(p ~ g,data=historyAd,FUN=mean)
plot(doppelUse)
The resulting graph is not very pretty, but you can see the fall-off among English speakers (in this simulated microcommunity) in probability of their use of the doppel form.
The following code draws graph 4a from EM. The functions for analysing data and constructing all the other graphs from the paper can also be called from the class L2017. It will also save a copy of the graph in your current working directory.
require( bldR );
l <- L2017$new();
l$G$setIsColour(TRUE);
l$draw_4b();
A snapshot of this package will be kept at the repository refered to in EM. While the code used in the article will not change (that is the three files TensorAgent.R, PopulationSimulation.R and L2017.R) unless bugs are found, the package codebase as a whole will change.
Later versions will include:
- simulations of language differentiation where information about the relative frequency of forms is not available.
- enriched cognitive models incorporating other biases that can change production frequency away from input frequency.
- data, analysis and presentation code from papers detailing future experiments.
The code and documentation presented here were written by T. Mark Ellison. Conceptual development as reported in the paper EM, and the design and collection of experimental data, were jointly undertaken with Luisa Miceli.
GNU Public License version 3