There also seem to be certain arguments against the reversibility of understanding and generation in natural language processing, e.g., differences between active and passive language use (although this argument -- as far as I know -- has only been discussed for lexical material), decision-making during generation vs. hypothesis maintainance during understanding, or the differences in input structures, which cause a different traversal of the problem space.
The above discussion can be seen as an argument against this view.
Clearly, it is the case that humans only use a subset of the lexical
material they are able to understand. But this need not necessarily lead to the
conclusion that understanding and generation are substantially different
processes. Otherwise the above mentioned adaptability would be very hard to
model. In a similar way [Wilks1991] also argues that phenomena such as
active vs. passive lexical usage do not contradict a symmetrical position.
Moreover he claims that ``if speaking a language is to utter new and creative
metaphors all the time as many researchers assert, then we can also
presume that a language generator must have access to the inverse of
that very same process, 
''
Some researchers (like [McDonald1987, Mann1987]) assume that the
space of problems and complexity is too different such that understanding and
generation could be described by the same underlying processes.
Although McDonald McDonald:89 could adopt the view that both
processes could employ the same knowledge represented in the same
way he states that ``the two processes that draw on the knowledge cannot be the
same because of the radical differences in information flow: Decision-making
[during generation] is a
radically different kind of process than hypothesis maintenance [during
understanding]
Understanding processes must cope with ambiguity and under-specification,
problems that do not arise in generation".
However, as we will make clear in
section 5.1, generation has also to cope with ambiguity and
under-specification to be able to produce adequate utterances. Of course, one
could argue that a produced utterance's ambiguity is not of relevance because
the generator can assume that the hearer will be able to disambiguate
it. Under this assumption
generation would be less decision driven on the grammatical level
than understanding, because it
would be the task of the hearer to choose between alternative readings.
In this view, understanding can also be interpreted as a complex decision
process -- the decision on understanding an utterance as best as possible.
Moreover, consider the case where a speaker wanted to talk about new ideas. He has not only to decide how to verbalize these ideas with the material already available, but he has also to hypothesize about the possible interpretations by the hearer, i.e. he has to find a way to express the new ideas such that they will be understandable (as it is usually the case in an expert-novice relationship). However, if the speaker wants to be sure that his utterance will be understood in the intended way, he has to assume that the hearer of the utterance will in principle be able to choose the inferences that he intends to be drawn. If not, it would be more or less randomly that the speaker produces exactly those structures of which analysis would lead to the intended meaning or in other words, that the hearer can reverse the speaker's computation. For example, if the speaker wanted to highlight a particular part of his ideas he can do this by choosing a specific ordering of the elements of the utterance. However, this makes sense only, if he assumes that the hearer is sensitive to linear order in exactly the same way. But then both processes should better be viewed as being symmetrical at least on the linguistic level.
In summary, reversibility has important advantages. If it is possible to develop reversible NLS which have at least the same power and functionality as non-reversible systems, such systems should be preferred. In this thesis however, we also show that the use of reversible grammars leads to better natural language systems.