Over the course of human history, prominent hypotheses for the source of human behavioral uniqueness have included freedom from metaphysical necessity, rational intention, and strong social altruism. However, when applied to questions concerning the evolution and current use of human language, these hypotheses generate more problems than they solve. These include everything from the statistical improbability of "hopeful monster" mutations to the unstable strategy of altruism without reciprocity. Adopting a performative view of human language removes these problematic dependencies, and explains aspects of language like phatic speech, fossils, and register, that require special pleading by one or more of the popular hypotheses.

This thesis provides two pieces of computational evidence for a performative view of human language. First, it shows that the possibility space of human language is much smaller than expected for a rule based but otherwise semantically unrestricted communication system. Second, it shows that the semantic information contained in an entire speech act is poorly predicted by the semantic information of its individual parts, as might be expected for a compositional system designed for efficient communication.

To conduct this second analysis, a novel method was developed based on the distributional theory of semantics for measuring the information content of human speech. Briefly, it is a measure of the distance between the distribution of terms in a language and the conditional probability of their appearance within a context of interest. This same method shows promise for testing other hypotheses about the nature and origin of language that involve an informational component.