Understand · Barnard coursework · Computational modeling
Language as Social Inference: Computational Models of Pragmatics and Perspective-Taking
TL;DR
A comparative synthesis of how computational models explain the social reasoning behind language use.
Skills & stack
- 01Computational model evaluation
- 02Bayesian reasoning
- 03Cross-disciplinary synthesis
- 04Technical communication
Background
The problem
Language models must account for more than literal meaning. Successful communication also depends on reasoning about another person's beliefs, intentions, attention, and perspective.
What I did
The paper compares symbolic, probabilistic, and neural approaches to identify what each framework explains—and where its assumptions fall short in interactive communication.
Comparative synthesis
How can models represent language as social reasoning?
01 / Core question
The paper asks whether language simply reflects social cognition or also supports it. Its central argument is that communication depends on real-time inferences about what other people know, intend, and attend to.
02 / Models compared
-
Symbolic models
Make rules explicit and interpretable, but struggle when a listener's knowledge is uncertain.
-
Bayesian inverse planning and Theory of Mind
Infer beliefs and goals by reasoning backward from observed actions.
-
Rational Speech Act models
Formalize recursive speaker-listener reasoning to explain context-dependent meaning.
-
Response-time models
Use choices and decision times to infer another person's preferences and processing style.
-
Capacity-limited Bayesian RL and LLM-based agents
Offer complementary accounts of cognitive constraints and adaptive social behavior.
03 / What the comparison reveals
The paper concludes that no single framework captures the full complexity of social interaction. Dynamic feedback, hidden motives, cognitive limits, and context remain difficult to formalize.
04 / Transferable practice
This work demonstrates how I approach unfamiliar technical domains: define the underlying question, compare competing assumptions, connect formal models to human behavior, identify limitations, and communicate the result clearly across disciplines.
The project is a literature-based synthesis. It does not claim that I implemented or fitted the models discussed.