Multimodal Output Management - Man-Machine Dialogue: Design and Challenges - page 158

that allow it to single out the targeted object from the other objects,
prioritizing certain properties for certain types of objects, following in these
near universal preferences. This algorithm has been extended and adapted
many times to take into account any type of property including spatial
relations between several objects, to manage a consistency criterion between
the properties used to refer in a grouped manner to several references or even
to take into account contextual markers, especially when the properties
chosen are vague (long journey) and gradable [KOP 08, KRA 12].
The linguistic challenges for future research studies are a better
management of salience, redundancy, sets of referents, referents other than
objects, a better use of vague properties, relation managements between more
than two objects and a better control over the involuntary generation of
ambiguities. In a dialogueic framework, it is also a challenge to integrate into
the process the ability of joint construction of a reference with the hearer
[KRA 12]. There are more psychological challenges that can be found, as for
automatic generation in general, in collaboration with psycholinguists,
through, for example, experiments aiming to characterize the notion of
salience, and in taking into account human factors, that is in applying to
reference the principles that were developed above. From a technical point of
view, the current algorithms combine knowledge representation because of
adapted logic, research into graphs, constraint satisfaction and context
modeling. They are increasingly complex, and as is often the case in NLP and
MMD, currently explore combinations of symbolic approaches and
approaches based on corpus data.
9.3.4. Valorizing part of the information and text-to-speech synthesis
At the end of the output processing chain is the valorizing of part of the
information (whether it is visual, in natural language or in any other way) and
the materialization processes, especially ECA and text-to-speech synthesis.
To highlight the part of the information, the process uses the same principle as
that of the allocation of information over communication channels: it starts by
taking the constraints into account, then by managing a set of rules modeling
the preferences. The constraints are the same as previously stated; for
example, the sound level beyond which the system cannot go in spite of the
desire to highlight the part of the message, which can be translated by a
prosodic accentuation. The rules rely on the message's content, the dialogue