Text generation for small data regimes

In Natural Language Processing (NLP), applications trained on downstream tasks for text classification usually require enormous amounts of data to perform well. Neural Network (NN) models are among the applications that can always be trained to produce better results. Yet, a huge factor in improving results is the ability to scale over large datasets. Given that Deep NNs are known to be data hungry, having more training samples can always be beneficial. For a classification model to perform well, it could require thousands or even millions of textual training examples. Transfer learning enables us to leverage knowledge gained from general data collections to perform well on target tasks. In NLP, training language models on large data collections has been shown to achieve great results when tuned to different task-specific datasets Wang et al. (2019, 2018a). However, even with transfer learning, adequate training data remains a condition for training machine learning models. Nonetheless, we show that small textual datasets can be augmented to a degree that is enough to achieve improved classification performance. In this thesis, we make multiple contributions to data augmentation. Firstly, we transform the data generation task into an optimization problem which maximizes the usefulness of the generated output, using Monte Carlo Tree Search (MCTS) as the optimization strategy and incorporating entropy as one of the optimization criteria. Secondly, we propose a language generation approach for targeted data generation with the participation of the training classifier. With a user in the loop, we find that manual annotation of a small proportion of the generated data is enough to boost classification performance. Thirdly, under a self-learning scheme, we replace the user by an automated approach in which the classifier is trained on its own pseudo-labels. Finally, we extend the data generation approach to the knowledge distillation domain, by generating samples that a teacher model can confidently label, but not its student.