Practice of Efficient Data Collection via Crowdsourcing at Large-Scale (1912.04444v1)

Abstract: Modern machine learning algorithms need large datasets to be trained. Crowdsourcing has become a popular approach to label large datasets in a shorter time as well as at a lower cost comparing to that needed for a limited number of experts. However, as crowdsourcing performers are non-professional and vary in levels of expertise, such labels are much noisier than those obtained from experts. For this reason, in order to collect good quality data within a limited budget special techniques such as incremental relabelling, aggregation and pricing need to be used. We make an introduction to data labeling via public crowdsourcing marketplaces and present key components of efficient label collection. We show how to choose one of real label collection tasks, experiment with selecting settings for the labelling process, and launch label collection project at Yandex.Toloka, one of the largest crowdsourcing marketplace. The projects will be run on real crowds. We also present main algorithms for aggregation, incremental relabelling, and pricing in crowdsourcing. In particular, we, first, discuss how to connect these three components to build an efficient label collection process; and, second, share rich industrial experiences of applying these algorithms and constructing large-scale label collection pipelines (emphasizing best practices and common pitfalls).