Designing for quality in real-world mobile crowdsourcing systems

Abstract

Crowdsourcing has emerged as a popular means to collect and analyse data on a scale for problems that require human intelligence to resolve. Its prompt response and low cost have made it attractive to businesses and academic institutions. In response, various online crowdsourcing platforms, such as Amazon MTurk, Figure Eight and Prolific have successfully emerged to facilitate the entire crowdsourcing process. However, the quality of results has been a major concern in crowdsourcing literature. Previous work has identified various key factors that contribute to issues of quality and need to be addressed in order to produce high quality results. Crowd tasks design, in particular, is a major key factor that impacts the efficiency and effectiveness of crowd workers as well as the entire crowdsourcing process. This research investigates crowdsourcing task designs to collect and analyse two distinct types of data, and examines the value of creating high-quality crowdwork activities on new crowdsource enabled systems for end-users. The main contribution of this research includes 1) a set of guidelines for designing crowdsourcing tasks that support quality collection, analysis and translation of speech and eye tracking data in real-world scenarios; and 2) Crowdsourcing applications that capture real-world data and coordinate the entire crowdsourcing process to analyse and feed quality results back. Furthermore, this research proposes a new quality control method based on workers trust and self-verification. To achieve this, the research follows the case study approach with a focus on two real-world data collection and analysis case studies. The first case study, Speeching, explores real-world speech data collection, analysis, and feedback for people with speech disorder, particularly with Parkinson’s. The second case study, CrowdEyes, examines the development and use of a hybrid system combined of crowdsourcing and low-cost DIY mobile eye trackers for real-world visual data collection, analysis, and feedback. Both case studies have established the capability of crowdsourcing to obtain high quality responses comparable to that of an expert. The Speeching app, and the provision of feedback in particular were well perceived by the participants. This opens up new opportunities in digital health and wellbeing. Besides, the proposed crowd-powered eye tracker is fully functional under real-world settings. The results showed how this approach outperforms all current state-of-the-art algorithms under all conditions, which opens up the technology for wide variety of eye tracking applications in real-world settings.