Modeling language contact with the Iterated Learning Model

TL;DR

The paper investigates how contact between two mature, expressive, and compositional languages influences retention of core traits during transmission. It employs the Semi-Supervised Iterated Learning Model, simulating language mixing with a weighting parameter $p$ to control exposure from languages $\mathbb{A}$ and $\mathbb{B}$, and tracks expressivity $x$, compositionality $c$, stability $s$, and similarities $a$ and $b$ across generations. Results show that when exposure is balanced ($p=0.5$), the resulting language often diverges from both originals yet rapidly attains high $x$ and $c$ and stabilizes; when one language dominates ($p$ closer to 1), the emergent language tends to resemble that parent. These findings indicate that the same learning dynamics driving compositionality and stability also promote resilience and re-emergence of parental traits under contact, even in large binary meaning-signal spaces up to $2^{20}$. The work provides a scalable, population-light framework for analyzing language contact and offers insights relevant to creole formation and language evolution studies.

Abstract

Contact between languages has the potential to transmit vocabulary and other language features; however, this does not always happen. Here, an iterated learning model is used to examine, in a simple way, the resistance of languages to change during language contact. Iterated learning models are agent-based models of language change, they demonstrate that languages that are expressive and compositional arise spontaneously as a consequence of a language transmission bottleneck. A recently introduced type of iterated learning model, the Semi-Supervised ILM is used to simulate language contact. These simulations do not include many of the complex factors involved in language contact and do not model a population of speakers; nonetheless the model demonstrates that the dynamics which lead languages in the model to spontaneously become expressive and compositional, also cause a language to maintain its core traits even after mixing with another language.

Figures (5)

• Figure 1: Neural network. In the Semi-Supervised ILM both the encoder, $e$ (yellow), and decoder, $d$ (blue), are neural networks; in $e$ the meaning vector forms the input, it is mapped to hidden layer and then to a signal vector; in $d$ a signal vector forms the input, there is again a hidden layer and the output is a meaning vector. The autoencoder, $a$, is formed by concatenating the two neural networks, with the signal layer, green, in this case forming a third hidden layer for the autoencoder.
• Figure 2: Evolution of the mixed language. Here the language is evolved for 20 generations and the results of 50 simulations are plotted for five different language attributes. In every graph the individual simulations are plotted with thin pale lines, the mean with a thicker, darker line. The first three columns plot attributes of the language itself, showing its expressivity ($x$), compositionality ($c$) and the stability ($s$) of the iterated learning. The final two columns show how similar the languages are to the initial $\mathbb{A}$ and $\mathbb{B}$ languages, these similarities are $a$ and $b$. The three rows present results for three different values of $p$, the fraction of the initial language taken from language $\mathbb{A}$, these are $p=0.5$ when the new language is almost always different from either $\mathbb{A}$ or $\mathbb{B}$, $p=0.55$ when the new language is typically similar to $\mathbb{A}$, sometimes to $\mathbb{B}$ and sometimes to neither, and $p=0.75$ when the new language almost always resembles $\mathbb{A}$. In each case a $10\times 10\times 10$ agent is used.
• Figure 3: Modeling language contact. Language contact is simulated 50 times for $p$ values from 0.5 to 1.0 in increments of 0.05. A plots the fraction of simulations which result in a value of $a$, upper line, or $b$, lower line, greater than 0.9. In B$a$ is plotted vertically, although the simulations were only run for $p\in [0.5,1.0]$ there is no distinction between the $\mathbb{A}$ language and the $\mathbb{B}$ language, this means that the $b$ values in $p\in [0.5,1.0]$ can be used to extend the range over which $a$ is plotted to all of $p\in[0.0,1.0]$. Each black dot corresponds to one of the 50 simulations at each $p$ value. Horizontal jitter of $\pm 0.025$ has been added to aid visualization. The red line is the average. Both A and B describe the same set of simulations using a $10\times 10\times 10$ agent.
• Figure 4: The behaviour does not change substantially when the size of hidden layer or signal space is changed. Four different architectures are compared here. Here the average value of $a$ is given for the $10\times 10\times 10$, $10\times 12\times 10$ and $9\times 11\times 12$ and $10\times 15\times 20$ agents. The four curves are similar.
• Figure 5: Mixing larger languages. Here the language is evolved for 20 generations and the results of 50 simulations are plotted for the usual five different language attributes: $x$, $c$, $s$, $a$ and $b$. The two rows give two different values of $p$. In each case the encoder and decoder maps have a 20-30-20 layout.