Domain-adaptive Worker Modeling-based Label Completion for Crowdsourcing
Congxuan Zhang , Liangxiao Jiang , Liangjun Yu , Limin Wang
Worker modeling has been proven to be an effective strategy for label completion, which serves as a preprocessing method to improve the performance of the downstream label integration. However, in real-world crowdsourcing scenarios, each worker often annotates only a few instances, leading to insufficient worker modeling and consequently limiting the performance of label completion. To address this problem, we propose a novel domain-adaptive worker modeling-based label completion (DWMLC) algorithm. Specifically, for each worker, we first estimate worker similarity to find its similar workers (including itself). Then, we use the instances annotated by its similar workers to pretrain a classification model. Next, we use the instances annotated by this worker to fine-tune the pretrained model. Finally, we use the fine-tuned model to predict the missing labels of the instances unannotated by this worker. Extensive experiments demonstrate that DWMLC can significantly outperform its competitors in enhancing the performance of downstream label integration algorithms.
Crowdsourcing learning / label completion / worker modeling
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(2907KB)
