Semi-monitoring crowdsourcing marking data integration method facing imbalance of labels

Abstract

The invention discloses a semi-monitoring crowdsourcing marking data integration method facing imbalance of labels. According to the phenomena that (1) the label marking accuracy is not related to objects and (2) the weights of different marking tasks of the same object are the same, a new weighted parameter evaluation method and a marker capability evaluation methods are provided, and the semi-monitoring crowdsourcing marking data integration method facing imbalance of labels. An iteration manner is used for solving, so that estimation for the weighted parameters and the marker capability is more objective and accurate, and an integrated marking result is more accurate; and the method is suitable for different types of crowdsourcing marking data including but not limited to images, texts and videos.

Description

technical field [0001] The invention belongs to the technical field of data labeling, and in particular relates to a semi-supervised crowdsourcing labeling data integration method oriented to label imbalance. The method comprehensively considers weighting parameters and labeler capabilities. Background technique [0002] With the advent of the big data era, extracting knowledge from big data is the most important research direction in the computer field today, attracting the attention of fields such as artificial intelligence and machine learning. Methods such as machine learning rely on high-quality labeled datasets for algorithm and model training. Therefore, it is of great significance to quickly and efficiently construct high-quality data sets. In the past, the construction of datasets mainly relied on expert labeling, through the way of hiring, after a period of high-intensity work to manually label the data. This method has the characteristics of high quality, high c...

AI Technical Summary

Problems solved by technology

Although the above algorithm improves the integration accuracy of labeled data to a certain extent, in the process, when considering the final result integration, it is assumed that each label type has the same probability of being labeled.

However, in actual labeling, the types of tags are often unbalanced, and there are certain weight differences between them. It is necessary to introduce weighted parameters to rebalance the relationship between tags.

At the same time, the weights in different labeling tasks must be different, which can only be trained in actual tasks and cannot be pre-set

Images