Crowdsourcing Taxonomies

摘要

Taxonomies are great for organizing and searching web content. As such, many popular classes of web applications, utilize them. However, their manual generation and maintenance by experts is a time-costly procedure, resulting in static taxonomies. On the other hand, mining and statistical approaches may produce low quality taxonomies. We thus propose a drastically new approach, based on the proven, increased human involvement and desire to tag/annotate web content. We define the required input from humans in the form of explicit structural, e.g., supertype-subtype relationships between concepts. Hence we harvest, via common annotation practices, the collective wisdom of users with respect to the (categorization of) web content they share and access. We further define the principles upon which crowdsourced taxonomy construction algorithms should be based. The resulting problem is NP-Hard. We thus provide and analyze heuristic algorithms that aggregate human input and resolve conflicts. We evaluate our approach with synthetic and real-world crowdsourcing experiments and on a real-world taxonomy.