Fine-grained Entity Set Refinement with User Feedback

State of the art semi-supervised entity set expansion algorithms produce noisy results, which need to be refined manually. Sets expanded for intended fine-grained concepts are especially noisy because these concepts are not well represented by the limited number of seeds. Such sets are usually incorrectly expanded to contain elements of a more general concept. We show that fine-grained control is necessary for refining such sets and propose an algorithm which uses both positive and negative user feedback for iterative refinement. Experimental results show that it improves the quality of fine-grained sets significantly.