Twitter Word2vec model
As part of our ACL W-NUT 2015 shared task paper, we release a Twitter word2vec model trained on 400 million tweets, as described in detail in this paper. The model, including Python code to load and access it, can be downloaded here.
Dynamic Convolutional Neural Networks
Contextual Decomposition for CNNs
As part of my paper, I have implemented a decomposition technique for convolutional neural networks. I allows you to understand which patterns the neural networks is looking for to make a certain decision. In my case, I was interested in comparing character-level patterns of CNNs and BiLSTMs. The implementation can be found here.