Project ideas

Here are a few ideas for more focused course projects (beyond just replicating the key results in a paper):

1. Implement the RSA model of the Nie et al. (2020) architecture and explore its predictions for synthetic data; i.e., explore the underlying RSA model outside the context of neural NLG. This would be a project for a single person at most.

2. Re-implement the model of Nie et al. (2020) and test it on the A3DS data set. This data set has exhaustive captions, unlike the CUB-captions data set, and so should be more informative about the working of the model under ideal-data conditions.

3. Implement the Emitter-Suppressor model as a non-neural RSA model. Systematically compare it to an \(S_{2}\) RSA model. Additionally: plan (or even execute) an experiment to test whether human data is better predicted by one or the other model.

4. Re-implement the incremental RSA model of Cohn-Gordon et al. (2020) and check if performance boost is mostly or only due to difference between beam-search and greedy decoding.