Assignment strategies for ground truths in the crowdsourcing of labeling tasks

- Algorithms for exploiting ground truths in crowdsourcing are developed.
- Quite general workers are assumed in the development.
- The algorithms can be of benefit to general EM algorithm-based approaches.
- Evaluation demonstrates that our algorithms can work well in various situations.

It is expected that ground truths can result in many good labels in the crowdsourcing of labeling tasks. However, the use of ground truths has so far not been adequately addressed. In this paper, we develop algorithms that determine the number of ground truths that are necessary. We determine this number by iteratively calculating the expected quality of labels for tasks with various sets of ground truths, and then comparing the quality with the limit of the estimated label quality expected to be obtained by crowdsourcing. We assume that each worker has a different unknown labeling ability and performs a different number of tasks. Under this assumption, we develop assignment strategies for ground truths based on the estimated confidence intervals of the workers. Our algorithms can utilize different approaches based on the expectation maximization to estimate good-quality consensus labels. An experimental evaluation demonstrates that our algorithms work well in various situations.