Abstract: At present, people have ever-increasing preference for the Internet for expressing their personal experiences and opinions on almost anything at review sites, forums, discussion groups, blogs, etc. Those user-generated content contains very valuable emotional information. How to mine those emotional information automatically and efficiently will hence be a very challenging question, as well as be promising in applications and development of enterprise business intelligence and public opinion survey and so on. Text-leveled sentiment analysis technology is considered as an extension and enhancement of traditional topic detecting and tracking (TDT) technology by adding some new approaches and ideas, which is based on word semantic orientation computing. In this paper, a novel scalable word semantic orientation computing framework is proposed, in which the word semantic orientation computing is transformed into the function optimization. As an instance of the proposed framework, the authors build an undirected graph in the use of word similarity computing technology first, and then partition the word-to-word graph by the idea of ‘minimum-cut’, thereby function optimization is adopted in this word semantic orientation computing framework and resolved by using simulated annealing algorithm. The experimental results prove that the proposed framework is reasonable and the algorithm performs better than those existing counterparts.