Touch and Speech: Multimodal Interaction for Elderly Persons - Biomedical Engineering Systems and Technologies - page 364

3 A Formal Unified Dialogue Modelling Approach

As a typical recursive network based approach, generalized dialogue models were
developed by constructing dialogue structures at the illocutionary level (cf. [20]).
However, it is criticized for its inflexibility of dealing with dynamic information ex-
change. Meanwhile, information state update based theories were deemed the most
successful foundation of agent based dialogue approaches (cf. [21]), which provides a
powerful mechanism to handle dynamic information and gains a context sensitive
dialogue management. Nevertheless, such models are usually very difficult to manage
and extend (cf. [22]).
Thus, a unified dialogue modelling approach was developed. It combines the gene-
ralized dialogue models with information state updated based theories. This approach
is supported by a formal development toolkit, which is used to implement an effec-
tive, flexible, yet formally controllable dialogue management.

3.1 A Unified Dialogue Modelling Approach

Generalized dialogue models can be constructed with the recursive transition net-
works (RTN). They abstract dialogue models by describing illocutionary acts without
reference to direct surface indicators (cf. [23]). Fig. 1 (left) shows a simple genera-
lized dialogue model as a recursive transition network diagram. It is initiated with an
assertion from a person A, responded by B with three actions: accept, agree or reject.

Fig. 1. A generalized dialogue model as a simple recursive transition network (RTN) (left) & a
generalized dialogue model as a simple deterministic RTN with conditional transitions (right)

The generalized dialogue model above is a none-deterministic model. To build a
feasible interaction model, deterministic behavior should be assured for the interac-
tion flow. Thus, conditional transitions are introduced to improve the above dialogue
model (cf. Fig. 1 right). Let checkAssert be a method to check whether an assertion
holds with B's knowledge and a an assertion given by A, if the assertion holds, B can
agree with it; otherwise, B rejects it and initiates further discussion; if the assertion is
not known by B, then B accepts it. Such conditional transitions can only be activated
if the relevant condition is fulfilled. We call it the conditional RTN.
Although the conditional RTN based generalized dialogue model defines a determi-
nistic illocutionary structure, it does not provide the mechanism to integrate discourse
information. Thus, information state based theory was integrated into our unified di-
alogue model by eliminating some typical elements, e.g. AGENDA for planning the
next dialogue moves, because such information is already captured by the generalized
dialogue model; furthermore it complements illocutionary structure with update rules,
which is associated with the information state of current context, and can update the