Work & Play: Simulating language contact
Introduction. English shares cognates with many other languages, notably German and French. The Norman conquest of 1066 infused new words into the English lexicon, creating a divergence in register and context. Historians point to social conditions like the lifestyles of the Anglo-Saxon laborers and the French-speaking Norman nobility as the cause. But could the results have been different? Can the insights from Game Theory and Network Analysis be of help?
Contact and Context. The distinctions between animal and meat (as seen in Table 1) follow a salient trend in the English language: vocabulary of French origin tends to the prestigious, whereas words of Germanic stock often fall into working class topics; i.e. the laboring class has to work (Ger. Werk), the upper class can play (from Fr.plaisir = pleasure). It can be surmised that i) words likeswineandsheep once bore an animal/meat polysemy like their German cognates1and were used context-dependently and ii) lost this polysemy by adopting French alternatives, which caused lexical blocking (see e.g. Blutner 2000). We call this processemancipation from context dependence. In this article we simulate language contact and its impact on language use by applying signaling games (Lewis, 1969) to social networks. By combining social network environments with recent advances in game-theoretic linguistics, we want to i) simulate the emancipation from context dependence and ii) investigate the social parameters responsible for the salient systematic division of meaning between words of Germanic and French origin.
Experimental Set-Up. Language use is modelled by the repeated context signaling game (CSG), a Lewisean signaling game, played between two agents (speaker and recipient) and augmented by a set of contexts and probabilities on these contexts based on social statusσ∈[1,9]⊆N, assigned to each agent.
Figure 1 depicts the extensive form of a CSG. A detailed description of this game and its relation to the social status parameterσ will follow in the full paper. For our simulations 300 agents with a randomly chosen social statusσand arranged on a social network played theCSGrepeatedly as both speaker and recipient with their neighbors. Based on previous encounters, agents employ alearning dynamicto guide their decisions on how to communicate with their neighbors. At a simulation’s start the agents’ lexicons contain only one expression (e.g. ”swine”), usable for the animal itself and its meat. We then simulate the conquest of 1066 by replacing 10% of the agents by Norman invaders, agents whose lexicons also contain only one expression, but the french variant (e.g. ”pork”). When an agent as a recipient encounters an unknown expression, he adds it to his lexicon. The simulation ends when every agent’s lexicon contains both expressions and a unique meaning-expression allocation, i.e. a strategy, governs the whole society.
Result 1: Emancipation from Context Dependence The initial network’s agents learn to use expression ”swine” in a context-dependent fashion for the animal itself and its meat as well. The invasion, done by replacing 10 randomly chosen agents with invaders, triggers the word ”pork” spreading, entering every agent’s lexicon. Every agent, as both speaker and as recipient, learns a new context-independent strategy. The new expression makes the context irrelevant. After each trial every agent has learned one of two possible strategies: the expected strategy (S1) for using the word of French origin for the meat and word of Germanic origin for the animal or the other way around (S2) (Figure 2 (left)). Hundreds of simulations revealed that i) in each trial only one strategy spreads and stabilizes society-wide and ii) both strategies emerge equiprobably. But if chance determines whetherS1 orS2emerges, then how can we explain that language use likeS1 is predominant?
Result 2: Influence of Social Status. In the previous simulation runs, randomly chosen agents were replaced by invaders. We now account for the fact that the Normans in many cases displaced the English nobility. We consider the social statusσof the replaced agents as follows: we set a lower limitβ ∈Nsuch that only agents with a social statusσ≥β could be replaced by invaders. Beginning with Experiment 1 (β = 1) we incrementally increasedβ stepwise up to Experiment 9 (β = 9) with 300 trials per experiment.
Figure 2 (right) depicts the percentage of trials resulting with strategy S1 for each experiment. As you can see, the higher the lower limit β for the social status of agents replaced by invaders, the higher the probability for the emergence of the expected strategyS1.
Discussion. We showed that agents in a social network playing the CSG with a single-word lexicon resolve polysemy context-dependently. The simulated Norman invasion provided lexical alternatives, allowing speakers to distinguish between previously context-dependent meanings. But without further assumptions, the new words could have described any meaning equiprobably. By considering social status, the probability shifted to the expected strategy, confirming assertions made by historical linguists. In the full paper, we detail further effects of network structure and context on agents’ communication strategies.
1E.g.Schwein(swine) andSchaf (sheep) both can have the meaning of the animal or its meat depending of the context.
personal abstract freedom liberty knowledge science belief faith brotherly fraternal
animal meat
swine (Ger. Schwein) pork (Fr. porc) cow (Ger. Kuh) beef (Fr. bœuf) sheep (Ger. Schaf) mutton (Fr. mouton) deer (Ger. Tier) venison (Fr. venaison) Table 1: Systematic division of meaning space between words of Germanic and French origin, possibly a result of lifestyle and education differences. The left table shows the division between personal (Germanic origin) and abstract (French origin) concepts. The right table shows the division betweens words for animals and their meat, possibly a result of the French speaking nobility eating the meat of the animals raised by the English speaking workers.
N
N
S
R
1 0
R
1 0
S
R
0 1
R
0 1
N
S
R
1 0
R
1 0
S
R
0 1
R
0 1
PC(dinner) PC(farm)
.9 .1 .1 .9
dinner farm
’meat’ ’animal’ ’meat’ ’animal’
”swine” ”pork” ”swine” ”pork” ”swine” ”pork” ”swine” ”pork”
’m.’ ’a.’ ’m.’ ’a.’ ’m. ’a.’ ’m.’ ’a.’ ’m.’ ’a.’ ’m.’ ’a.’ ’m.’ ’a.’ ’m.’ ’a.’
Figure 1: Extensive form of the context signaling game for two contexts (dinner, farm), two topics (animal, meat) and two expressions (”swine”, ”pork”), played by a speaker (S) and a receiver (R).
Moves not influencable by any player are made by nature (N).
S1: ’meat’ ”pork”
’animal’ ”swine”
S2: ’meat’
”swine”
’animal’
”pork”
perc.S1-trials 20%
40%
60%
80%
100%
1 2 3 4 5 6 7 8 9 β
Figure 2: Left: The two strategiesS1 andS2, whose emergence is equiprobable if the replaced agents are chosen randomly. Right: By increasing the thresholdβof the replaced agents, the probability for a society-wide spread of expected strategyS1 raises.
References.
1. Blutner (2000), “Some Aspects of Optimality in Natural Language Interpretation”,Semantics 17 (3), pages 189–216.
2. Lewis (1969),Convention, Harvard University Press.
