Multimodal Output Management - Man-Machine Dialogue: Design and Challenges - page 145

and is the one we will study in this chapter. These are the successive
processing steps for text generation: content planning, i.e. choosing the way
to dispose of the different propositions constituting the semantic content;
sentence aggregation, i.e. allocating propositions to sentences and
determining the discourse relations; lexicalization, that is, i.e. the choice of
words; referring expression generation, for now only in a linguistic
framework, which requires the system to choose between direct reference and
anaphora; linguistic realization, with the application of syntactic and
morphological rules to obtain a well-built sentence [REI 00]. Within the
framework of oral dialogue, we can add a prosody determination phase, and a
text-to-speech synthesis phase, which can include an oral rendition of
emotions and dialogue act management, notably with the generation of an act
that materializes the change in speaking turn. If the dialogue involves an
ECA, the how can also be broken down into several processes: choosing a
type of physically perceptible behavior, taking the what into account, then
rendering of this behavior; see Chapter 9 of [GAR 11]. The management of a
speaking head specifically requires a face animation phase (lips, eyes, hands,
body, in general) including the visual rendition of emotions. All these
processes involve various techniques, from the use of patterns, whether they
are syntactic, prosodic, gestures, animations, with or without variables that
can be set, to the management of linguistic and discourse phenomena as is the
case for natural language generation when it uses information structure
principles.
The links between the system's general behavior and all these processes is
sometimes hard to draw. The rendition of emotions is a favored means to
transmit a few indications, for example on the positive or negative direction of
the message. Making this direction vary depending on the answers, the
incomprehensions, the ambiguities or simply the system's inability to answer
a request increases the realistic aspect of the interaction: a systematically
positive interaction can irritate the user in case of repeated incomprehensions
and, obviously, a mainly negative direction does not help with the cooperating
aspect of the dialogue. Beyond the simple positive or negative direction (some
systems can get upset when they detect malicious behavior on the user's part),
the current emotion models involve various dimensions, each one potentially
able to materialize itself on several modalities: valency (positive or negative),
activation (weak to strong), the level of control (fear is not, for example,
linked to a feeling of situation control, whereas anger is), and the level of
unexpected; see Chapter 3 of [GAR 11].