Date of Completion
Language self-organization evolution lexicon convention
Field of Study
Doctor of Philosophy
In the emergence of natural languages, two processes typically co-occur: new form-meaning mappings are being conventionalized, and these forms reduce, i.e., shorten, in various ways. Despite this oft-noted co-occurrence, and the existence of theories of conventionalization and reduction separately, there are no explicit, mechanistic, and empirically-motivated theories of the connection between the two processes. To fill this gap, the present dissertation aims to better understand this relationship both empirically and theoretically, with experimental semiotics and agent-based modeling.
In the experiment, eight groups of four hearing, non-signing English-speaking participants were brought into the lab. Quads split into rotating dyads which took turns using gestures – usually highly iconic – to communicate (images of) basic objects (e.g., cow, orange, boy) to each other. As in naturalistic language emergence, quads conventionalized, reduced, and improved success at the task in a self-organizing fashion. Across objects, slope of conventionalization negatively correlated with slope of reduction, suggesting that those objects that conventionalize more also reduced more. Critically, we found that participants reduced more after communication failure, inconsistent with a listener-oriented/rational-agent model of the link between conventionalization and reduction, whereby conventionalization causes communication success, which causes agents to try reducing their utterances (i.e., to try to save effort when they think communication is likely to succeed).
To develop a new theory of the link between conventionalization and reduction, we implemented an agent-based model of gesture production, comprehension, and learning. Three key properties subserve the model: (1) probabilistic gesture-object mappings initially set to model iconicity (which participants in the experiment spontaneously use), (2) language production dependent on a notion of informativeness (understanding one’s own utterance), and (3) a single learning mechanism in listeners, whereby listeners align their probabilistic associations with speaker’s utterances. This model simultaneously conventionalized and reduced, and captured other aspects of our experimental setting.
Consistent with previous naturalistic and experimental work, our experiment and model each show parallel conventionalization and reduction. To our knowledge, our model is the first to simultaneously capture these two phenomena, and shows that conventionalization and reduction can simultaneously emerge from independently necessary mechanisms of language production, comprehension, and learning, rather than from agents rationally selecting optimal lexicons, or optimally responding to their interlocutors’ state of understanding. With our work as a starting point, future work could further investigate how conventionalization and reduction relate to the emergence of phonology and grammaticalization.
Richie, Russell, "Conventionalization and Reduction in Natural Language Emergence: An Experimental and Computational Model Investigation" (2017). Doctoral Dissertations. 1541.