We determine the opinion content by looking at a variety of language markers that might indicate the expression of the writer's opinion of events, as opposed to a more dry recitation of facts. Associated Press wire stories typically will score low, whereas opinion pieces, editorials, Op-Eds, and many blog posts will tend to score high.
Sentiment is determined by looking at specific words and language markers that indicate either a postive or negative judgement. The language markers are determined through extensive machine learning, to extract those that are most reliable.