Interpretable Textual Neuron Representations for NLP

Data

We use the dataset from Chrupała, G., Gelderloos, L., & Alishahi, A. (2017). Representations of language in a model of visually grounded speech signal. ACL., which can be downloaded here.

Code

train.py

Trains a keras implementation of the GRU IMAGINET architecture from Kádár, A., Chrupała, G., & Alishahi, A. (2017). Representation of linguistic form and function in recurrent neural networks. Computational Linguistics..

optimize.py [search|embfree|logits|softmax|gumbel] [I|T]

Finds optimal n-gram for the image or text modality using the specified method.