Millions of people rely on online opinions to make their decisions. To better help people glean insights from massive amounts of opinions, we present the design, implementation, and evaluation of OpinionBlocks, a novel interactive visual text analytic system. Our system offers two unique features. First, it automatically creates a fine-grained, aspect-based visual summary of opinions, which provides users with insights at multiple levels. Second, it solicits and supports user interactions to rectify text-analytic errors, which helps improve the overall system quality. Through two crowd-sourced studies on Amazon Mechanical Turk involving 101 users, OpinionBlocks demonstrates its effectiveness in helping users perform real-world opinion analysis tasks. Moreover, our studies show that the crowd is willing to correct analytic errors, and the corrections help improve user task completion time significantly.