Crowdsourcing systems empower individuals and companies to outsource labor-intensive tasks that cannot currently be solved by automated methods and are expensive to tackle by domain experts. Crowdsourcing platforms are traditionally used to provide training labels for supervised machine learning algorithms. Crowdsourced tasks are distributed among internet workers who typically have a range of skills and knowledge, differing previous exposure to the task at hand, and biases that may influence their work. This inhomogeneity of the workforce makes the design of accurate and efficient crowdsourcing systems challenging. This dissertation presents solutions to improve existing crowdsourcing systems in terms of accuracy and efficiency. It explores crowdsourcing tasks in two application areas, political discourse and annotation of biomedical and everyday images. The first part of the dissertation investigates how workers' behavioral factors and their unfamiliarity with data can be leveraged by crowdsourcing systems to control quality. Through studies that involve familiar and unfamiliar image content, the thesis demonstrates the benefit of explicitly accounting for a worker's familiarity with the data when designing annotation systems powered by the crowd. The thesis next presents Crowd-O-Meter, a system that automatically predicts the vulnerability of crowd workers to believe \enquote{fake news} in text and video. The second part of the dissertation explores the reversed relationship between machine learning and crowdsourcing by incorporating machine learning techniques for quality control of crowdsourced end products. In particular, it investigates if machine learning can be used to improve the quality of crowdsourced results and also consider budget constraints. The thesis proposes an image analysis system called ICORD that utilizes behavioral cues of the crowd worker, augmented by automated evaluation of image features, to infer the quality of a worker-drawn outline of a cell in a microscope image dynamically. ICORD determines the need to seek additional annotations from other workers in a budget-efficient manner. Next, the thesis proposes a budget-efficient machine learning system that uses fewer workers to analyze easy-to-label data and more workers for data that require extra scrutiny. The system learns a mapping from data features to number of allocated crowd workers for two case studies, sentiment analysis of twitter messages and segmentation of biomedical images. Finally, the thesis uncovers the potential for design of hybrid crowd-algorithm methods by describing an interactive system for cell tracking in time-lapse microscopy videos, based on a prediction model that determines when automated cell tracking algorithms fail and human interaction is needed to ensure accurate tracking.
| Identifer | oai:union.ndltd.org:bu.edu/oai:open.bu.edu:2144/27482 |
| Date | 22 February 2018 |
| Creators | Sameki, Mehrnoosh |
| Source Sets | Boston University |
| Language | en_US |
| Detected Language | English |
| Type | Thesis/Dissertation |
Page generated in 0.002 seconds