In order to implement a domain-independent dialogue-processing module, we need to be able to generate referring expressions that help us to discriminate different representations. As an example, consider two hotels carrying the same name but being located in different addresses. From a representational point of view, we are looking at a set of feature structures some of which contain common information. In order to generate a clarification question, prompting the user to select, say, one of the two hotels, the system should be able to separate similarities and differences in the representations. This is a necessary precursor for generating clarification questions in a domain-independent way.
Sets of feature structures can be represented in an underspecified representation
factoring out similarities and differences in the different feature structures.
For example, two feature structures of the form
and
respectively can more
compactly be represented as
, 
being the greatest lower bound of 
and 
in the type
hierarchy. In addition, the types and features are annotated with indices of feature
structures in order to avoid overgeneralization, being similar in spirit to named
disjunctions. Figure 3 depicts a generic form for a compact representation of
feature structures in which the feature is defined.
Figure 4 shows the informational content
of two feature structures F and G and their common information H from an
information-set perspective.
Figure 3: An underspecified feature structure. The types 
, 
, 
are represented in trees that preserve the subsumption relation from the type hierarchy. Types and
features are annotated with indices referring to the feature structures that contain them in order
to be able to extract the feature structures correctly from the compact representation.
By the same token, we can determine compatible information between feature structures.
As a result of this operation, we obtain a representation separating compatible from incompatible information.
This is helpful for example in determining which constraints specified by the user can not be fulfilled
and to establish close solutions. The structures separating incompatible information
are similar in structure to, yet differ in semantics with, the one shown in
figure 3. In figure 4, I depicts the compatible information of the feature structures F and G.
Figure 4: The information represented in two feature structures F and G
In both cases, the nodes trees consist of decision trees whose elements are annotated with the indices of the original feature structures. For disambiguation, the dialogue strategy may select one or more of the decision trees according to some strategy specific criterion. The selection criteria might be to disambiguate the feature path whose value has a decision tree of maximal or minimal entropy, according to the way the question is generated (for a more detailed presentation on the generation of clarification questions, see [Denecke1997]). Due to the co-indexed types and features in the underspecified representation the disambiguation of one feature path typically reduces the ambiguity in other feature paths as well. The compact representation helps us to select discriminating information when generating clarification questions.
It should be noted that although the construction of the decision trees relies on domain-specific knowledge (e.g. a museum is more specific than an object in the above example) the implementation of the underspecification algorithm does not since the selection of the decision tree can be formulated in terms of entropy and specificity and constitutes thus a necessary prerequisite for domain-independent specification of dialogue strategies.
Not only may underspecified feature structures be used to represent differences and similarities of objects being ambiguously referred to, but they also serve to represent ambiguous references to goals or actions. The same clarification strategies may then be used to disambiguate between multiple objects, intentions or actions that are referred to by one description.