Active learning and crowdsourcing: a survey of annotation optimization methods

Abstract

High quality labeled corpora play a key role to elaborate machine learning systems. Generally, creating of such corpora requires human efforts. So, annotation process is expensive and time-consuming. Two approaches that optimize the annotation are active learning and crowdsourcing. Methods of active learning are aimed at finding the most informative examples for the classifier. At each iteration from the unplaced set, one algorithm is chosen by an algorithm, it is provided to the oracle (expert) for the markup and the classifier is newly trained on the updated set of training examples. Crowdsourcing is widely used in solving problems that can not be automated and require human effort. To get the most out of using crowdplatforms one needs to to solve three problems. The first of these is quality, that is, algorithms are needed that will best determine the real labels from the available ones. Of course, it is necessary to remember the cost of markup - to solve the problem by increasing the number of annotators for one example is not always reasonable - this is the second problem. And, thirdly, sometimes the immediate factor is the rapid receipt of the marked corpus, then it is necessary to minimize the time delays when the participants perform the task. This paper aims to survey existing methods based on this approaches and techniques to combine them. Also, the paper describes the systems that help to reduce the cost of annotation.

Keywords

• активное обучение
• краудсорсинг
• аннотация корпусов
• крауд-вычисления