| Intelligent systems hold the promise of enabling intent-level control and exploration of information, to providing robust accessibility accommodations to people with disabilities. However, automated approaches to solving these problems are far from being able to solve many fundamental problems underlying these applications, such as understanding visual information, intention, and natural language.;Leveraging crowds of human workers to provide input as part of a structured, algorithmic process (called Human Computation) can effectively solve many problems that cannot currently be completely solved by automated approaches, or individual people. For example, converting speech to text in real time is difficult for computers because they cannot understand the meaning behind the audio signal they are presented with, which varies greatly in real settings due to acoustic conditions and speaker variations. People also find this task challenging because it requires typing hundreds of words per minute to match natural speaking rates.;This thesis presents solutions to two critical limitations of current crowdsourcing approaches: maintaining context over multiple turns of interaction, and supporting continuous interactions between systems and their users or environments. It also explores how groups of people can go beyond what individuals can, and presents a unified architecture for abstracting the underlying complexity of crowd-powered systems as a single Crowd Agent. |
