This repository contains code and models for the paper Learning the Rational Speech Acts Model (Monroe and Potts, Amsterdam Colloquium 2015).
You'll first need Python 2.7 with NumPy, as well as the TUNA corpus:
./get_tuna
To run the main model presented in the paper (last line of Table 1):
bash ./runs/308/_rerun.sh # furniture
bash ./runs/309/_rerun.sh # people
Check out the above shell scripts for the full command that runs the program with tunable options.
Each of the "learned" models in Table 1 can be rerun with a similar command, replacing the number in the run directory:
Model | Furniture | People |
---|---|---|
S0 basic | 275 | 276 |
S0 gen | 11 | 12 |
S0 basic+gen | 306 | 307 |
S1 basic | 251 | 252 |
S1 gen | 3 | 4 |
S1 basic+gen | 308 | 309 |
Be prepared to wait a while for the models to finish! 2 days is normal for the S1 models (on 2016 hardware); the S0 ones should be faster. (Replicating results is great, but if you'd rather not wait, see stdout.log in each run directory for the output of my own run of the program.)