The conflict detection and resolution in knowledge merging for image annotation

Semantic annotation of images is an important step to support semantic information extraction and retrieval. However, in a multi-annotator environment, various types of conflicts such as converting, merging, and inference conflicts could arise during the annotation. We devised conflict detection patterns based on different data, ontology at different inference levels and proposed the corresponding automatic conflict resolution strategies. We also constructed a simple annotator model to decide whether to trust a given piece of annotation from a given annotator. Finally, we conducted experiments to compare the performance of the automatic conflict resolution approaches during the annotation of images in the celebrity domain by 62 annotators. The experiments showed that the proposed method improved 3/4 annotationaccuracy with respect to a naïve annotation system. The Conflict Detection and Resolution in Knowledge Merging for Image Annotation Cheng-Yu Lee and Von-Wun Soo Department of Computer Science, National Tsing Hua University HsinChu Taiwan 300 Department of Computer Science and information engineering, National University of Kaohsiung, Kaohsiung, Taiwan, 811 E-mail: {leoli;soo}@cs.nthu.edu.tw Abstract Semantic annotation of images is an important step to support semantic information extraction and retrieval. However, in a multi-annotator environment, various types of conflicts such as converting, merging, and inference conflicts could arise during the annotation. We devised conflict detection patterns based on different data, ontology at different inference levels and proposed the corresponding automatic conflict resolution strategies. We also constructed a simple annotator model to decide whether to trust a given piece of annotation from a given annotator. Finally, we conducted experiments to compare the performance of the automatic conflict resolution approaches during the annotation of images in the celebrity domain by 62 annotators. The experiments showed that the proposed method improved 3/4 annotation-accuracy with respect to a naïve annotation system.Semantic annotation of images is an important step to support semantic information extraction and retrieval. However, in a multi-annotator environment, various types of conflicts such as converting, merging, and inference conflicts could arise during the annotation. We devised conflict detection patterns based on different data, ontology at different inference levels and proposed the corresponding automatic conflict resolution strategies. We also constructed a simple annotator model to decide whether to trust a given piece of annotation from a given annotator. Finally, we conducted experiments to compare the performance of the automatic conflict resolution approaches during the annotation of images in the celebrity domain by 62 annotators. The experiments showed that the proposed method improved 3/4 annotation-accuracy with respect to a naïve annotation system.