Abstract

Human Computation is an emerging paradigm that organizes human efforts to carry out a computation or task that computers cannot do well, like detecting objects in videos. In this paradigm, humans become a type of processing unit and crowdsourcing serves as a tool for distributing computations to humans. In recent years, crowdsourcing has been popularized by commercial crowdsourcing platforms like Amazon’s Mechanical Turk and has reached an industrial scale, involving millions of Internet users who offer the human processing power. However, unlike CPUs, humans are, well, humans and are influenced by a great many things, including their own motives and aspects of the computation task.

In this lecture we use the evaluation of search engines as a context for studying various aspects of crowdsourcing. In order to measure the performance of a search engine human assessments of search results are needed. Since online search engines involve millions of queries a day and at least 10 times as many retrieved documents, such a task seems like a perfect match for human computing. However, there are many issues to consider, as we will show through a number of experiments conducted using Amazon’s Mechanical Turk.