ICT FML Workshop May 7-9 2010
- David DeVault
- Dirk Heylan
- Stefan Kopp
- Brent Lance
- Jina Lee
- Stacy Marsella (co-chair)
- Paul Piwek
- Alicia Sagae
- Prasan Samtani
- David Traum (co-chair)
- Hannes Vilhjalmsson
- Michael Rushforth
- Jon Gratch
- Ari Shapiro
- Ross Mead
- Marcus Thiebaux
Examples of FML in use
- Relationship of FML and BML: division of labour - what goes
where, e.g. role of linguistic syntax
- Agent vs System point of view: should FML/BML be for a single
autonomous agent or for groups of agents working as part of a
unified presentation system?
- How to express unintentional funcion - e.g. unsuccesful deception
- Breaking/opening up the SAIBA pipeline
- Clearly confine what FML is mainly for (<- important to take a practical stance here)
- spec of everything that needs to be realized by behaviors in order to achieve some higher-level goal?
- spec of everything that might possibly causal of overt behavior (e.g. cognitive operations)?
- context variables: part of FML?
- goal/intention hierarchy - define where FML has to pick up from (e.g. include full-blown Gricean communicative intentions?)
- required to make everything explicit (e.g. implicit speech acts)?
- Structural issues (<- this is in my view one of the most important first steps we should tackle now)
- what is the basic layout of fundamental structures: units, operations, dimensions, other parameters?
- what dimensions are to be involved? how can they be combined (e.g. emotion and speech acts)? how can each of them be organized within sub-spaces (e.g. taxonomy of speech acts)?
- compositionality and (in-)dependencies
- preferences and conflict resolution rules
- Possible starting points
- pro's and con's of taking things like DIT++, which provide a lot of insight but were not really meant for the same thing (e.g. annotation system vs. generation specification language)
- working towards FML bottom-up from behavior planning issues (i.e. what needs a Behavior Planner?) or top-down from content and action planning issues?
- adopting info-state update approach for FML? do we need an underlying concept of a dialogue model in relation to which FML is to work?
- How is incrementality dealt with?
- How to balance generality and representation of domain-specific information?
- How is contextual information dealt with?
- Do we need to make a clear distinction between reactive and planned behaviours?
- Do we need to take into account multilinguality and also accessibility issues?
- How does FML fit in with language generation architectures (such as Reiter & Dale and RAGS) and language production architectures (e.g., Levelt and De Ruiter)?
- To what extent does FML allow for representation of non-communicative behaviours.
Possible FML Elements
- Dialogue Acts
- Information Structure
- explicitness/importance of information (implications for whether
one should be subtle or overt, implicated or explicit, complementary
or redundant in expressing the function)
- interpersonal relationship management (status, dominance/submission, closeness/distance)
- culture-specific factors & personality differences (is this a function itself or a method of realizing function in behavior, or some of each)
- behavior-level function (e.g. to nod head, regardless of what
- Environmental effects (i.e. monitoring status of objects/characters, capture of attention)
Conversational regulation (i.e. nonverbal backchannel feedback)
Paul: Rather than introduce specific elements, I have added
some bullet points on issues related to specific
elements/attributes. (References at the bottom
of the page)
- Should we distinguish between felt and expressed emotion?
- Should we distinguish between explicitly communicated and implicated information?
- Should we have a representation of private and shared information?
- For referring expression generation, some kind of representation of context may be required. Do we also include salience (see Krahmer & Theune 2002 and Piwek 2009) and importance of objects in this?
- Do we allow for multiple dialogue act tags for a single segment?
- Dialogue participants: Do we need to distinguish participants, side participants, bystanders and eavesdroppers.
- Should pragmatic parameters be included. For example those of Hovy
(1988) about the setting:
- time (much, some, little)
- tone (formal, informal, festive)
- conditions (good, noisy)
- Should parameters influencing politeness be included, e.g, those from Brown and Levinson (1987) which were operationalized in Walker et al. (1996):
- Social distance
- Power that the speaker has over the hearer
- Do we need to represent bias, e.g., for agents that are not cooperative?