In spoken English, language users produce linguistic structures which elude a merely syntactic description since, for example, intonational means may conjoin syntactic structures which are otherwise unrelated. This paper discusses a theoretical framework which establishes the level of parasyntax as an appropriate level of linguistic description. At the level of parasyntax, talk units represent linguistic units which are prosodically and syntactically defined. This theoretical framework is put to the test on authentic corpus data. By looking at large amounts of data in context, it is possible to identify an important principal function fulfilled by talk units: information structure in terms of information hierarchy and information packaging. The talk-unit model can also be fruitfully applied to other fields of research, e.g. pausology. In a wider setting, it seems possible that corpusbased methods may lead to new theoretical concepts of spoken language which are able to account for phenomena that have not been covered by existing descriptive frameworks.