Fine-Grained Tasks for Crowdsourced Entity Resolution

Entity resolution aims to identify records that point to the same entity in the real world. In recent years, crowdsourcing approaches have provided new ideas for entity resolution, using human intelligence to bring entity resolution to a higher level that can meet the short-term needs of a variety of users, unlike deep learning models that require large amounts of labeled data to train the model and are therefore subject to more research and development. Earlier approaches relied excessively on propagation relationships, and manual tasks were typically about whether record pairs matched. Balancing quality and cost is an important research problem. To address the above issues, this paper proposes a crowdsourced entity resolution approach with fine-grained tasks, proposing two fine-grained manual tasks that focus on the connections within the data, combine human intelligence, and are easier to answer. The entity resolution is considered a splitting problem of the dataset. And the problem of a large number of records sharing the same attribute in the data is solved by using scoring tasks. The experimental results on real datasets show that the FGTER algorithm proposed in this paper can achieve a higher F-1 score with cost control.