Cognitive and social factors in agent-based models of conversational priming in repetitional responses

Full text

Cognitive and social factors in agent-based models
of conversational priming in repetitional responses
Peter Dekker (Vrije Universiteit Brussel)
Sonja Gipper (Universität zu Köln)
Bart de Boer (Vrije Universiteit Brussel)

Hypothesis
💬 Not much research on the role of conversation in language change
🚊As the prime habitat for language, conversation must provide infrastructures
for linguistic innovations to spread from one speaker to another
HYPOTHESIS:
💡 Conversational priming in repetitional responses leads to faster spread of
innovative forms

Asymmetries in repetitional responses in conversation
(Yurakaré, Isolate, Bolivia)
(1) A: adojla balip
bali-p
go -2PL.SBJ
‘Did you (pl.) walk?’
(2) A: dulawla
dula -w=la
make -3PL.SBJ=COMM
‘Did they build it?’
B: adojla balitu
a-dojjo=la bali -tu
go -1PL.SBJ
‘We walked.’
B: dulaw
dula-w
make-3PL.SBJ
‘They built it.’
data from: van Gijn, Hirtzel, Gipper & Ballivián Torrico 2011: Conversation-NL, YURGVDP08oct06-01
Priming
Repeat may force B to copy form from A.
Priming facilitates reuse, repetition boosts
priming (Gipper 2020)
No priming
Deictic shift, no reuse of person marker

Spread of innovations through priming?
Inspiration for model: Lithuanian dialects Zietela and Lazūnai
3SG/PL is only form that changes in innovative dialect, caused by priming?
eĩti ‘to go’ Zietela
(conservative)
Lazūnai
(innovative)
inf. eĩti eĩti
1SG eimù eimù
2SG eimì eimì
3SG=PL eĩti eĩma
1PL eĩmam eĩmam
2PL eĩmat eĩmat
Vidugiris 2014: 198–200
Rozwadowski 1995: 136, thanks to Eugen Hill
Zietela, Belarus (Wikipedia)
PREDICTION In languages with person marking, if an innovation occurs, third person
markers will diffuse faster than first and second person markers.

Agent-based model
Computer model of repeats in conversations (cf. general models of innovation
spread: Pierrehumbert et al. 2014; Josserand et al. 2021)
● Meanings: 1sg, 2sg, 3sg
● Forms: conservative vs. innovative
● Initialize 2 agent types: conservating (0% innovative form) vs. innovating (90%
innovative form) agents
● Interaction: question and answer, boost distribution based on interlocutor's and own
utterances
Priming:
If person different (1/2sg) → Sample form from own distribution
If person same (3sg) → Use same form as questioner
3SG
3SG
1SG
2SG

Results basic model
● Priming (3SG) converges faster than non-priming (1SG)
● Model converges to population mean (differential equations)

Surprisal
● Favour innovative forms implicitly?
● Forms with higher surprisal (Hale 2001; Levy 2008) show stronger
priming effects (Bernolet & Hartsuiker 2010; Jaeger & Snider 2013)
● Our model: multiply boost with surprisal → larger update when
receiving unexpected form
I = - log2 p(unit|context)
Result:
Converge to 0.5, faster for 3SG (priming)

Network structure
● Favour innovative form implicitly by position of
innovating speakers in network
● Network: lower clustering coefficient for innovating
than for conservating speakers
● Result: innovative form bit higher, but not taking over
Basic Clustering coefficient network

Conclusion & future work
● Persons with priming (3SG) converge
faster than non-primed persons
(1SG/2SG): support for hypothesis?
● Innovative form implicitly privileged
through surprisal and network structure, but
innovative form does not take over
●Explicit privilege necessary for innovative
form to take over in S-curve (see Blythe &
Croft 2012)
Explicit privilege innovative form

Acknowledgments THANK YOU!
The Yurakaré Nation
Jeremías Ballivián Torrico
Eugen Hill
Pascal Coenen
University of Cologne Excellent Research Support Programme, FORUM
University of Cologne Cluster Development Programme
DFG Netzwerk Interaktionale Linguistik (413161127)
Volkswagenstiftung (81821 & 83448)
Research Foundation – Flanders (FWO) PhD Fellowship (11A2821N)