In the method described above the anticipation feedback loop operates
only on the semantic level, i.e. whether an elliptic utterance will be
ambiguous or not is determined only under semantic considerations. If a
candidate partial subtree has been found the grammatical construction of
it is done without checking the utterance's grammatical ambiguity. The implicit
assumption made here is that the string determined for an unambiguous
SR 
will also be analyzed definitely as SR . But this
does not avoid the
risk of being misunderstood in general, because it could be the case
that the resulting string is still ambiguous.
Most interesting from a reversibility standpoint is that Jameson and Wahlster
JamesonWahlster:82 also vote for the use of reversible
knowledge sources
, the recognition component must operate on the same kind of structure as those returned by the system's ellipsis generation component.
In HAM-ANS the generation and understanding components share the same internal representation language. However, two different grammars are used during parsing and generation. If AFL were also designed for grammatical processing (actually they do not consider integration of parsing and generation) the same problems as already mentioned for Levelt's model would also occur.