Reverse active learning for optimising information extraction training production

When processing a noisy corpus such as clinical texts, the corpus usually contains a large number of misspelt words, abbreviations and acronyms while many ambiguous and irregular language usages can also be found in training data needed for supervised learning. These are two frequent kinds of noise that can affect the overall performance of machine learning process. The first noise is usually filtered by the proof reading process. This paper proposes an algorithm to deal with noisy training data problem, for a method we call reverse active learning to improve performance of supervised machine learning on clinical corpora. The effects of reverse active learning are shown to produce results on the i2b2 clinical corpus that are state-of-the-art of supervised learning method and offer a means of improving all processing strategies in clinical language processing.