Defining the meaning of terms in a controlled vocabulary is difficult. 
This is best demonstrated in the BBC's Blackadder III (Ink and incapability).

Ontologies should be developed and maintained in the context of
literature which underpins the data to be annotated with the ontology.
To link ontologies to text, textmining is useful. Although it is hard to
extract terms from text, there have been successess such as 
80 percent and higher success rate in disambiguation tasks and
gene identification tasks. There are also many systems around such
as GoPubMed.org, Textpresso, EbiMed, to name a few. Textmining can
help to identify candidate terms for ontologies and even definitions,
although the process is only semi-automatic due to false positives.

Thus, textmining can help to build and then maintain ontologies by
linking terms to text.

Meaning can be defined with logical formulae. However, the problem arises of
representation and reasoning. Science is geared towards positive statements, but negative
ones can be very helpful too. Negative statements can be made with different meaning: Two
proteins may have been shown by an experiment not to interact or two proteins may be
assumed not to interact since the relevant databases do not state otherwise. The two types
of negation can be distinguished. However, there are many different ways how to define the
meaning.  It is often favourable to state in an ontology that something is normally the
case, but that there may be exceptions. There is a long tradition in artificial
intelligence defining logics which can handle this form of reasoning. No matter what the
expressiveness of a logic is, users may not use all its capabilities. Snomed e.g.
uses a simplified description logic which only provides for existential restrictions.
Logical statements are useful for reasoning and representation, but they must be hidden
form end users. It would be desireable to have a system in which authors submit abstracts
for both humans and machines. The latter are formulated in a controlled natural language,
which ensures that the author can read them.