Language Agents demo

This is a demo of an agent-based model based on the Rational Speech Act (RSA) model[1]. Each agent has knowledge that represents which signal (Sx) the agent thinks can in principle be used to refer to a referent (Rx). The speaker computes per referent a probability distribution over words: "If I want to communicate Rx, what is the probability of Sx getting that message across?" These Theory of Mind (ToM) distributions are computed for a specific theory of mind level according to the RSA model. The speaker also has a probability distribution over referents, representing her (weighted) desire to communicate Rx: "I both want to communicate R1 and R3, therefore P(R1) = 0.5 and P(R2) = 0.5." The posterior signal distribution is computed by integrating the ToM map and the desired referent distribution.

The listener performs this computation in reverse. She computes per word a probability distribution over referents: "If I head Sx, what is the probability of Rx being the correct referent?" Given a probability distribution over words, representing uncertainty in the observed signal, she computes a probability distribution over referents.

Feel free to try it out for yourself. Play around with the agents' knowledge, desires and observations. See if you can make them understand each other; or try to see when communication breaks down. Have fun. For questions, please contact me.

References

  1. Michael C. Frank and Noah D. Goodman. Predicting pragmatic reasoning in language games. Science, 336(6084):998, 2012