Timezone: »

Improved Text Classification via Test-Time Augmentation
Helen Lu · Divya Shanmugam · Harini Suresh · John Guttag

Test-time augmentation---the aggregation of predictions across transformed examples of test inputs---is an established technique to improve the performance of image classification models. Importantly, TTA can be used to improve model performance post-hoc, without additional training. Although test-time augmentation (TTA) can be applied to any data modality, it has seen limited adoption in NLP due in part to the difficulty of identifying label-preserving transformations. In this paper, we present augmentation policies that yield significant accuracy improvements with language models. A key finding is that augmentation policy design–for instance, the number of samples generated from a single, non-deterministic augmentation–has a considerable impact on the benefit of TTA. Experiments across a binary classification task and dataset show that test-time augmentation can deliver consistent improvements over current state-of-the-art approaches.

Author Information

Helen Lu (Massachusetts Institute of Technology)
Divya Shanmugam (MIT)
Harini Suresh (MIT)

Harini is a student at MIT in CSAIL (Computer Science and Artificial Intelligence Laboratory), pursuing a Master of Engineering degree in Computer Science. She received her Bachelor's degree in Computer Science from MIT as well. She is interested in what machine learning can do to improve health and medicine, and particularly in the application of deep neural networks to clinical time series data. Her current research in the Clinical Decision Making group at MIT (under Pete Szolovits) aims to use deep autoencoders to uncover latent patient phenotypes, and use these in a recurrent LSTM neural network to predict mortality and intervention onset/duration. Her past research in the Computational Biophysics Group at MIT also utilized clinical data to make mortality predictions, and explored the question of how selecting and engineering various features affected prediction performance.

John Guttag (Massachusetts Institute of Technology)

More from the Same Authors