Just ask a human? - controlling quality in relational similarity and analogy processing using the crowd

Abstract

Advancing semantically meaningful and human-centered interaction paradigms for large information systems is one of the central challenges of current information system research. Here, systems which can capture different notions of `similarity' between entities promise to be particularly interesting. While simple entity similarity has been addresses numerous times, relational similarity between entities and especially the closely related challenge of processing analogies remain hard to approach algorithmically due to the semantic ambiguity often involved in these tasks. In this paper, we will therefore employ human workers via crowd-sourcing to establish a performance baseline. Then, we further improve on this baseline by combining the feedback of multiple workers in a meaningful fashion. Due to the ambiguous nature of analogies and relational similarity, traditional crowd-sourcing quality control techniques are less effective and therefore we develop novel techniques paying respect to the intrinsic consensual nature of the task at hand. These works will further pave the way for building true hybrid systems with human workers and heuristic algorithms combining their individual strength.