Word Vector Model Descriptions In Depth :

​ : Many state of the art NLP algorithms currently use deep learning techniques. A 2014 paper by Goodfellow et al demonstrates that, at least in image classification, error rates from deep learning can be improved by adversarial training. Generating adversarial examples relies on having high­dimensional input spaces and many output labels. We present two methods of generating adversarial examples, and also introduce a new loss function for training word vectors in a CBOW model. Using QVEC as a metric for our word vectors, these techniques both improve on the vanilla CBOW model, and these methods can be used in conjunction with other generalization techniques such as dropout and early stopping.