This repository contains the code for the following paper:

"Text classification by untrained sentence embeddings",
D. Di Sarli, C. Gallicchio, A. Micheli,
Submitted to Intelligenza Artificiale.

The folder structure is organized on two levels. At the top level we have the datasets (QC, SMS, SNLI), and inside each folder we find the models that are applied to that dataset.

To launch an experiment, for example a leaky ESN on the QC task, run this command from the root of the project:

$ python3 QC/leaky_esn  # ...args...

To reproduce the results in the paper, just run the following commands:

$ python3 QC/leaky_esn --searches 0 --final-trials 10
$ python3 QC/leaky_esn_ensemble --searches 0 --final-trials 10
$ python3 QC/leaky_esn_attn --searches 0 --final-trials 10
$ python3 QC/mygru --searches 0 --final-trials 10

$ python3 SMS/leaky_esn --searches 0 --final-trials 10
$ python3 SMS/leaky_esn_ensemble --searches 0 --final-trials 10
$ python3 SMS/leaky_esn_attn --searches 0 --final-trials 10
$ python3 SMS/mygru --searches 0 --final-trials 10

$ python3 SNLI/leaky_esn --searches 0 --final-trials 10
$ python3 SNLI/leaky_esn_ensemble --searches 0 --final-trials 10
$ python3 SNLI/leaky_esn_attn --searches 0 --final-trials 10
$ python3 SNLI/mygru --searches 0 --final-trials 10

It is advisable to run the code on a GPU with at least 16GB of memory. Alternatively, the code can easily be modified to accumulate the gradients of smaller minibatches in the training loop.