Research And Development Of Crowdsourcing Platform For Professional Attributes Labeling

The big data era coming, artificial intelligence and machine learning get rapid development, and the demand of dataset is becoming more and more urgent, which leads to the requirement of rapid label to mass data. Researchers proposed the compromising crowdsourced labeling method, between the traditional expert labeling and computer labeling. Crowdsourcing is outsourcing the tasks originally finished by experts to the public. It has the advantage of low cost, high efficiency and high quality and has been widely used in recent years. With the development of crowdsourcing, crowdsourcing platform began to appear and develop rapidly. In the past few years, a variety of crowdsourcing platform appeared to finish various tasks abroad, while at home, crowdsourcing platform put more attention on software crowdsourcing, which results in the lack of crowdsourcing platform for professional attributes labeling.In this paper, I research and develop crowdsourcing labeling platform "RenCongZhong" for professional visual attributes. After requesters provide pictures and attributes, the crowdsourcing platform will generate tasks automatically and release tasks for workers to label. The platform design proper function models for three different attributes tasks. After tagging is completed, platform integrates final crowdsourcing label result automatically or semi-automatically.We cooperated with China Academy of Art and based on "RenCongZhong", collected aesthetic attributes dataset. After three months’ labeling, I get the labeling result.In this paper, I propose a semi-supervised crowdsourcing learning algorithm, which is based on the imbalance between tags. According to the imbalance between label classes, weighted parameters are introduced. By a small count of ground truth labels, I train the weighted parameters which adapt to the task label class, and add weights to the original expectation maximization algorithm to improve the final integrated label’s accuracy.