This thesis presents a collection of research articles that make contributions in the area of semantic classification and evaluation. Semantic classification describes the automatic processing of data, such as text, by machines, with the goal of simulating “understanding” the intended semantics, and as a result of this making a decision, for instance about the topic being discussed, or how some text should be translated into another language, or whether some piece of information constitutes fake news. This area has seen tremendous development in recent years, especially with the wide spread use of artificial neural network architectures, practically leading to almost human-like performance. This thesis presents a series of contributions in the design of artificial neural network architectures that: 1) can capture with high accuracy the most salient parts of text, in terms of syntax, semantics and grammar; 2) can capture semantic compositionality accurately; and 3) that can accurately detect fake news using different types of supporting evidence. This thesis also presents a series of contributions in how text processing is evaluated. Specifically, this thesis presents: 1) a family of novel evaluation measures that can evaluate rankings with respect to several aspects, such as relevance, and credibility and usefulness; 2) the biggest to this day evaluation dataset for fake news classification; and 3) a method for improving the evaluation capacity of incomplete evaluation datasets. Collectively, the above contributions advance the state of the art in how machines process and understand text.