Halliday & Matthiessen (2014: 54):
If there is no description to draw on, this means that we will gradually have to develop one based on the analysis of a representative sample of texts (a corpus). In other words, describing a language is a process of generalising from the analysis of textual data. The outcome of this process is a description of the system of the language, and we keep testing such descriptions by deploying them in continued text analysis and by applying them to different tasks such as language education or natural language processing.