This paper proposes relationship discovery models using opinions mined from the Web instead of only conventional collocations. Web opinion mining extracts subjective information from the Web for specific targets, summarizes the polarity and the degree of the information, and tracks the development over time. Targets which gain similar opinionated tendencies within a period of time may be correlated. This paper detects event bursts from the tracking plots of opinions, and decides the strength of the relationship using the coverage of the plots. Companies are selected as the experimental targets. A total of 1,282,050 economics-related documents are collected from 93 Web sources between August 2003 and May 2005 for experiments. Models that discover relations are then proposed and compared on the basis of their performance. There are three types of models, collocation-based, opinion-based, and integration models, and respectively, four, two and two variants of each type. For evaluation, company pairs which demonstrate similar oscillation of stock prices are considered correlated and are selected as the gold standard. The results show that collocation-based models and opinion-based models are complementary, and the integration models perform the best. The top 25, 50 and 100 answers discovered by the best integration model achieve precision rates of 1, 0.92 and 0.79, respectively.