NL Generation for Virtual Humans in a Complex Social Environment

NL Generation for Virtual Humans in a Complex Social Environment
NL Generation for Virtual Humans in a Complex Social Environment
David Traum
Michael Fleischman and Eduard Hovy
University of Southern California
Institute for Creative Technologies
[email protected]
University of Southern California
Information Sciences Institute
fleisch,hovy @isi.edu
1
Introduction
Natural language generation is a broad field, given the wide
variety of different applications for text generation. Perhaps
one of the most challenging of these applications is natural
language generation for spoken dialogue systems. In spoken
dialogue systems, real-time throughput is required, which
constrains the processing to less than a second if the system
is to seem natural, especially given other processing of input
and output. Thus text generation approaches which involve
selecting from among many possible alternatives or involve
complex calculations to determine preferences (Langkilde &
Knight 1998) is not appropriate. Generation in dialogue is
also somewhere in between single-shot sentence generation
and generation of extended discourse. On the one hand, single short utterances must be generated because one can not
predict a priori exactly how the other dialogue participant(s)
will react, and subsequent generation may depend more on
the input that is newly provided than any previously available information. On the other hand, dialogues generally
have a coherent structure, depending on the goals and overall structure of the task that is being discussed as well as
the immediately previous utterance (Grosz & Sidner 1986).
Thus text-planning notions are still relevant, even if one can
not count on being able to produce paragraph-level or longer
utterances as pre-planned due to the interactive nature of dialogue.
Another large issue for generation systems is the nature of
the input to the system. While the output is generally well
defined as coherent texts (of whatever size) in the appropriate language, there is no generally agreed-upon format for
the input to a generation system. Even if one were to define
a standard language for this input, this would not necessarily alleviate the problem, since the conversion from existing
representations to this standard language might be just as
difficult as the process of generation to natural language itself. For generation within a dialogue system, there is also
the question of where the “dialogue component” ends and
generation begins. In dialogue systems a large part of the
issue is not just how to convey a given meaning in natural
language, but what meaning should be conveyed, as well as
when this utterance should be spoken. This is to be conCopyright c 2003, American Association for Artificial Intelligence (www.aaai.org). All rights reserved.
trasted with generation for machine translation, in which the
content is already specified, and the utterances are generated one by one into the target language as the inputs are
provided in the source language (e.g., (Dorr 1989)). Likewise, for story generation or instruction manual generation,
the generator may decide how much to express explicitly
and how to order the presentation of content, but the content itself is largely fixed by the input, and interactive issues,
such as whether the next planned text could be produced or
how long to wait before producing the next text are largely
absent. Generation for dialogue systems must also be concerned with dialogue issues such as turn-taking, grounding,
initiative, and collaborative notions such as obligations and
joint goals. There is also an issue of “point of view” of the
speaker, which may be absent from text generation systems.
In this paper, we describe a generation system for virtual humans (Rickel et al. 2002), in a story-based multicharacter Virtual-reality training system (Swartout et al.
2001). Generation for these characters puts additional constraints beyond those of most dialogue systems. The language produced must be appropriate for the character’s role
in the interactive experience, expressing emotions as well
as beliefs and goals. The characters must also be able to
speak to multiple addressees, tailoring language for each.
The agents also need to be able to express content using both
speech and visual modalities.
In the next section, we describe the virtual world and virtual humans that use the dialogue and generation systems we
have developed. In Section 3 we describe the architecture
of the agent system, including how dialogue and generation
processing fits in. In Section 4, we describe the dialogue
representations that are used as motivation and inputs for the
generation system. In Section 5 we provide more detail on
aspects of the generation process, from motivation to speak
through the agent speaking english text. Section 6 has some
final remarks.
2 MRE
The test bed for our dialogue model is the Mission Rehearsal
Exercise project at the University of Southern California’s
Institute for Creative Technologies. The project is exploring
the integration of high-end virtual reality with Hollywood
storytelling techniques to create engaging, memorable training experiences. The setting for the project is a virtual real-
ity theatre, including a visual scene projected onto an 8 foot
tall screen that wraps around the viewer in a 150 degree arc
(12 foot radius). Immersive audio software provides multiple tracks of spatialized sounds, played through ten speakers
located around the user and two subwoofers. Within this setting, a virtual environment has been constructed representing a small village in Bosnia, complete with buildings, vehicles, and virtual characters. This environment provides an
opportunity for Army personnel to gain experience in handling peacekeeping situations.
The first prototype implementation of a training scenario
within this environment was completed in September 2000
(Swartout et al. 2001). To guide the development, a Hollywood writer, in consultation with Army training experts,
created a script providing an overall story line and representative interactions between a human user (Army lieutenant)
and the virtual characters. In the scenario, the lieutenant
finds himself in the passenger seat of a simulated Army vehicle speeding towards the Bosnian village to help a platoon
in trouble. Suddenly, he rounds a corner to find that one of
his platoon’s vehicles has crashed into a civilian vehicle, injuring a local boy. The boy’s mother and an Army medic are
hunched over him, and a sergeant approaches the lieutenant
to brief him on the situation. Urgent radio calls from the
other platoon, as well as occasional explosions and weapons
fire from that direction, suggest that the lieutenant send his
troops to help them. Emotional pleas from the boy’s mother,
as well as a grim assessment by the medic that the boy needs
a medevac immediately, suggest that the lieutenant instead
use his troops to secure a landing zone for the medevac helicopter.
While the script is fine for a canned demo, we need
to go beyond it in a number of ways, allowing for variation both depending on divergent behavior of the Lieutenant trainee, as well as perhaps a director or simulation
Observer/Controller. Agents must communicate in ways
faithful to their emotional and intellectual assessment of the
situation yet still strive to maintain the immersiveness of a
well-told story.
We currently use full NL generation for two of the characters, the sergeant, and the medic, while using a variety of
templates and fixed prompts for generating the lines of the
other characters. Figure 1 gives an example recorded dialogue fragment from this domain, illustrating a number of
characters involved in multiple conversations across multiple modalities (LT, Sgt and Medic discussing the situation
face to face, LT on radio to base and eagle1-6, Sergeant
shouting orders to squad leaders). The Lt utterances were
spoken by one of our researchers. The Sgt and Medic utterances were generated and synthesized spontaneously by
the agents playing those roles. The other characters (Base,
3rd squad leader, and platoon Eagle 2-6) were controlled
by simple algorithms and used pre-recorded prompts that
were triggered either by keyword utterances or signals from
the simulator illustrating a number of characters involved in
multiple conversations across multiple modalities (LT, Sgt
and Medic discussing the situation, LT on radio to base and
eagle1-6, Sergeant with squad leaders). The utterances are
labelled with conversation.turn.utterance (if a
turn consists of only a single utterance, the “,1” is omitted).
Not shown in the fragment are many non-verbal behaviors
which are coordinated with and part of the communication
between agents. For instance, the agents look at a speaker
who is part of their conversation, while otherwise they might
attend to other relevant tasks. They avert gaze to keep the
turn when planning speech before speaking. Speech is also
accompanied by head and arm gestures. After turns 3.1 and
5.1, troops move into position as ordered, signalling understanding and acceptance of the orders.
3
Overview of Agent Architecture
The animated agents controlling the Sergeant and Medic
characters are implemented in the SOAR programming language (Laird, Newell, & Rosenbloom 1987), and built on
top of the STEVE agent architecture (Rickel & Johnson
1999). SOAR provides a declaratively accessible information state, as advocated in the Trindi project (Larsson &
Traum 2000). Most processing is done by production rules
that examine aspects of the information state and produce
changes to it.1 In general, all rules fire (once) whenever their
left side makes a new match against aspects of the information state, and rules will fire in parallel. There are elaboration cycles of finding matching rules and applying their
results. Part of the information state contains the current operators, which act as a kind of focussing mechanism, so that
an agent serially attends to different functions. Rules can
have their left hand side refer to aspects of the operator so
that they apply only when this operator is active. There is
also a higher level decision cycle, in which new operators
are chosen. The agent can focus on sub-goals by introducing new operators as subordinate to the main goal rather than
replacing the main operator.
The STEVE agents, as augmented for the MRE project,
have multiple components, each involving one or more operators, sometimes also including parts that work in other
operators as well. These include
a belief model including knowledge of the participants
and other relevant people, objects and locations of the domains, including social relationships and various relevant
properties.
a task model, consisting of knowledge of the events that
can happen in the world as well as plans for sequencing
these tasks into plans to achieve specific goal states.
a perception module that receives messages from the
world-simulator and other agents and updates the agent’s
internal state, mediated by notions of foveal attention.
an emotional model, including basic emotional states, appraisal of emotional state, and coping mechanisms to influence behavior (including mental actions such as adopting goals) on the basis of emotion (Gratch & Marsella
2001; Marsella & Gratch 2002).
a body control module that adjusts body position, gaze,
and gesture as appropriate for engaging and monitoring
tasks as well as involvement in face to face conversation.
1
SOAR also has access to functions in the tcl programming language, when needed.
a dialogue module, maintaining the dialogue layers described in the previous section.
1.1
1.2.1
1.2.2
LT
SGT
SGT
1.3
1.4
1.5
1.6
1.7
1.8
1.9
1.10
1.11.1
1.11.2
1.12.1
1.12.2
1.13
2.1
2.2
2.3
2.4.1
2.4.2
2.4.3
LT
SGT
LT
SGT
LT
SGT
LT
SGT
MEDIC
MEDIC
LT
LT
SGT
LT
BASE
LT
BASE
BASE
BASE
2.4.4
2.4.5
2.5
1.14
1.15
1.16
1.17
3.1.1
3.1.2
3.1.3
3.1.4
3.1.5
3.1.6
4.1.1
4.1.2
4.2
5.1.1
5.1.2
6.1.1
6.1.2
6.2.1
6.2.2
6.3.1
6.3.2
6.3.3
6.3.4
6.4.1
6.4.2
1.18
1.19.1
1.19.2
1.19.3
BASE
BASE
LT
LT
SGT
LT
SGT
SGT
SGT
SGT
SGT
SGT
SGT
SGT
SGT
3SLDR
3SLDR
3SLDR
1-6
1-6
LT
LT
1-6
1-6
1-6
1-6
LT
LT
LT
SGT
SGT
SGT
1.20
LT
what happened here?
there was an accident sir
this woman and her son came from the side street
and our driver didnt see them
who’s hurt?
the boy and one of our drivers
how bad is he hurt?
the boy or the driver
the driver
the driver has minor injuries sir
how is the boy?
Tucci?
sir he is losing consciousness
we need to get a MedEvac in here ASAP
understood
Sergeant where is the medevac
the MedEvac is at the base sir
eagle base this is eagle two six over
Eagle two six this is eagle base over
requesting medevac for injured civilian over
Standby.
Eagle two six this is eagle base
medevac launching from operating base alicia
time now
eta your location zero three
over
roger eagle base two six out
sergeant secure a landing zone
sir first we should secure thee assembly area
secure the assembly area
understood sir
Squad leaders listen up
give me three sixty degree security here
First Squad take twelve to four
Second Squad take four to eight
Third Squad take eight to twelve
Fourth Squad secure thee accident site
Johnson
send a fire team to the square to secure an LZ
Yes sergeant
Sergeant Duran
get your team up to the square and secure an lz
Eagle two six this is one six
whats your ETA over
one six this is two six
ETA 45 minutes over
two six it’s urgent you get here right now
situation’s getting critical
We’re taking fire
over
Roger 1-6
2-6 out
sergeant send two squads forward
sir that’s a bad idea
we shouldn’t split our forces
instead we should send one squad to reconn forward
send fourth squad to recon forward
Figure 1: An MRE dialogue interaction fragment
a generation module, to produce natural language utterances for the agents to say.
an action selection module that decides which operators
should be invoked at which times.
In addition to these core agent functions implemented
within SOAR, the agent also relies on external speech recognition and semantic parsing modules, which send messages
to the agent (which are interpreted by the perception module), and speech synthesis and body rendering, that take the
output body control and speech directives from the agent
and produce the behaviors for human participants to see and
hear.
Dialogue related behavior is performed in three SOAR
operators. Understand-speech is invoked whenever perception detects new utterances, including results from speech
recognition and parsing. The understand-speech operator includes recognition rules, which use both the input as well as
the information state to decide on the set of dialogue acts
that have been performed in the utterance. The same operator is used regardless of whether the input speech came
from a human participant (via speech recognition and parsing), the agent itself (via the output-speech operator – see
below), or from other agents (via agent messages that may
also include some partially interpreted dialogue acts). The
update-dialogue-state operator applies updates to the information state that are associated with the recognized acts.
All of the generation functions from deciding what to say
to producing the speech happen in the output-speech operator. Proposal rules for this operator function as goals to
speak, given various configurations of the information state.
Selection of one of these proposal rules (meaning there is
nothing more urgent to do or say2 ) constitutes selection of a
goal for generation. Once in the operator, there are several
phases including also a couple of sub-operators. First is the
content selection phase, in which the agent reasons about
how best to achieve the output goal. Examples are which assertion to make to answer a pending question, or how to respond to a negotiation proposal. Once the content has been
selected, next there is a sentence planning phase, deciding
the best way to convey this content. This is followed by
a realization phase, in which words and phrase structures
are produced. Next, a ranking phases considers the possibly multiple ways of realizing the sentence, and selecting
a best match. This final sentence is then augmented with
communicative gestures including lip synch, gaze, and hand
gestures, converted to XML, and sent to the synthesizer and
rendering modules to produce the speech. Meanwhile, messages are sent to other agents, letting them know what the
agent is saying. The output-speech operator continues until
callbacks are received from the synthesizer, letting the agent
know either that the speech was completed or has been interrupted (perhaps by someone else’s speech). The last part of
2
In SOAR, one can write preference rules for operator selection,
which are used by the action selection module to decide on the
current operator.
the operator prepares the content for the understand-speech
operator, so that other dialogue acts, beyond those that were
explicitly planned can be recognized. For example, an operator might be concerned with computing and generating
an answer to a previously asked question. While the main
goal is to provide the answer, this utterance will also involve
speech acts relating to grounding the question, taking the
turn, and making an assertion.
4 Dialogue Representation
While there is no generally accepted notion of what a dialogue model should contain, there is perhaps growing consensus about how such information should be represented.
Following the Trindi project (Larsson & Traum 2000), many
choose to represent an information state that can serve as a
common reference for interpretation, generation, and dialogue updates.
Depending on the type of dialogue and theory of dialogue
processing, many different views of the specifics of information state and dialogue moves are possible. A complex
environment such as the MRE situation presented in the previous section obviously requires a fairly elaborate information state to achieve fairly general performance within such
a domain. We try to manage this complexity by partitioning
the information state and dialogue moves into a set of layers, each dealing with a coherent aspect of dialogue that is
somewhat distinct from other aspects.
Each layer is defined by information state components, a
set of relevant dialogue acts, and then several classes of rules
relating the two and enabling dialogue performance:
recognition rules that decide when acts have been performed, given observations of language and nonlinguistic behavior in combination with the current information state
update rules that modify the information state components
with information from the recognized dialogue acts
selection rules that decide which dialogue acts the system
should perform
realization rules that indicate how to perform the dialogue
acts by some combination of linguistic expression (e.g.,
natural language generation), non-verbal communication,
and other behavior.
contact
attention
conversation
– participants
– turn
– initiative
– grounding
– topic
– rhetorical
social commitments (obligations)
negotiation
Figure 2: Multi-party, Multi-conversation Dialogue Layers
The layers used in the current system are summarized in
Figure 2. The contact layer (Allwood, Nivre, & Ahlsen
1992; Clark 1996; Dillenbourg, Traum, & Schneider 1996)
concerns whether and how other individuals can be accessible for communication. Modalities include visual, voice
(shout, normal, whisper), and radio. The attention layer
concerns the object or process that agents attend to (Novick
1988). Contact is a prerequisite for attention. The Conversation layer models the separate dialogue episodes that go
on during an interaction. A conversation is a reified process entity, consisting of a number of sub-layers. Each of
these layers may have a different information content for
each different conversation happening at the same time. The
participants may be active speakers, addressees, or overhearers (Clark 1996). The turn indicates the (active) participant with the right to communicate (using the primary
channel) (Novick 1988; Traum & Hinkelman 1992). The
initiative indicates the participant who is controlling the
direction of the conversation (Walker & Whittaker 1990).
The grounding component of a conversation tracks how information is added to the common ground of the participants (Traum 1994). The conversation structure also includes a topic that governs relevance, and rhetorical connections between individual content units. Once material is
grounded, even as it still relates to the topic and rhetorical
structure of an ongoing conversation, it is also added to the
social fabric linking agents, which is not part of any individual conversation. This includes social commitments —
both obligations to act or restrictions on action, as well as
commitments to factual information (Traum & Allen 1994;
Matheson, Poesio, & Traum 2000). There is also a negotiation layer, modeling how agents come to agree on these
commitments (Baker 1994; Sidner 1994). More details on
these layers, with a focus on how the acts can be realized
using verbal and non-verbal means, can be found in (Traum
& Rickel 2002).
The interface between generation and dialogue is still a
difficult issue, given a lack of general agreement both on
what constitutes the division of labor between those two areas, as well as no general agreement on internal representations of dialogue. Concerning the former point, in some
systems, NL generation is seen as a sort of server in which
meaning specifications are fed in, and NL strings are sent
back, for the dialogue module to decide when to say (or stop
saying). In other systems, e.g.(Allen, Ferguson, & Stent
2001; Blaylock, Allen, & Ferguson 2002), an interaction
manager controls both generation, production, and feedback
monitoring but not other dialogue functions. In starting the
MRE project, we were unsure on exactly what the interface should be between these two modules, but we knew
we wanted to be able to easily make more information available when it is needed and can be used. The simplest way to
achieve this is to have the generation component be part of
the agent itself, implemented using SOAR production rules,
and having full access to the information state provided by
the dialogue modules as well as task reasoning and emotion. It also allows for easy interleaving of processes, such
as synchronizing gaze behavior and turn-taking with utterance planning.
5 Generation Phases
In this section, we look at each phase in the generation in a
little more detail.
5.1 Operator Proposal and Selection
There are a number of output speech operator proposal rules.
One basic one is to ground utterances by other speakers for
which the agent is an addressee, giving evidence of understanding or lack of understanding. Another type of proposal
rule concerns the obligation to address a request (including
an information request regarding a question). Other rules
involve trying to get the attention of another agent, making
requests and orders, clarifying underspecified or ambiguous
input, and performing repairs. There are also preference
rules that arbitrate between multiple possible outputs. Preferences are given to addressing a request or question rather
than merely acknowledging it, or for talking about a higherlevel action rather than a sub-action. Preference is also given
to reactive acts over initiative-taking acts.
5.2 Content Planning
Given a goal, there are often many ways to respond. An
obligation to address a question can be met by answering
the question, but also by refusing to answer, deferring the
answer, or redirecting the question to another agent to answer. Even with a decision to answer, there are always multiple possible answers, including both true and false answers.
Within the set of answers the agent believes to be true, there
are also more or less informative answers available, depending on the assumed mental state of the interlocutor, but also
based on perceptual evidence of accessible information and
likely inferences or general interest. We also use the agent’s
emotion model to focus on which content to present given a
choice of valid answers.
Likewise, for a proposal (issued as either an order or a
request), the agent must decide how to respond, whether to
accept, defer, reject, counterpropose, or perform other negotiation moves. For clarifications, one must decide which
information to ask for.
5.3 Sentence Planning
Creation of sentence plans from content is currently a hybrid process. There is a fully general but simple sentence
planner, which can produce simple sentence plans for any
task or state that is in the agent’s belief or task model. For
more precise and non-standard realization, some sentence
plans are selected rather than generated from scratch, given
certain configurations of the content as well as other aspects
of the information state. Finally, there is a short-cut procedure which also bypasses realization and moves directly to
pre-selected prompts.
Figure 3 shows an example content specification that is
input to the sentence planner. These inputs contain minimal information about the object, state or event to be described, along with references to the actors and objects involved, and values representing the speaker’s emotional attitude toward each object and event. A detailed account of
the emotional aspects of the generation system can be found
in (Fleischman & Hovy 2002). A set of SOAR production
rules expands this information into an enriched case frame
structure (Figure 4) that contains more detailed information
about the events and objects in the input.
ˆtime past
ˆspeech-act assert
ˆevent
:reference
:attitude
ˆagent
:reference
:attitude
ˆpatient :reference
:attitude
collision
-1
driver
+4
mother
+1
Figure 3: Example input to sentence planner: content annotated with speaker’s attitudes toward objects and events
(<utterance> ˆtype assertion
ˆcontent <event>)
(<event> ˆtype event ˆtime past
ˆname collision ˆagent <agent>
ˆpatient <patient> ˆattitude -1)
(<agent> ˆtype agent ˆname driver
ˆdefinite true ˆsingular true
ˆattitude +4)
(<patient> ˆtype patient ˆname mother
ˆdefinite true ˆsingular true
ˆattitude +1)
Figure 4: output of sentence planning
The task of expansion involves deciding which frame is to
be chosen to represent each object in the input. For example,
Figure 5 shows several possible frames that could be used to
represent the agent driver. The decision is based on the
emotional expressiveness, or shade, of each semantic option.
A distance is calculated, using an Information Retrieval metric, between the shade of each semantic frame representing
the driver and the emotional attitude of the speaker toward
the driver. The frame with the minimum distance, i.e., the
frame that most accurately expresses the agent’s emotional
attitudes, is chosen for expansion. This is done for each of
the objects associated with the event or state. Once all objects have been assigned a frame, planning is complete, and
realization begins.
5.4
Realization
Realization is a highly lexicalized procedure, so tree construction begins with the selection of main verbs. Each verb
in the lexicon carries with it slots for its constituents (e.g.,
agent, patient), as well as values representing the emotional
shade that the verb casts both on the event it depicts and the
constituents involved in that event.
Once the verb is chosen, its constituents form branches in
a base parse tree. Production rules then recursively expand
the nodes in this tree until no more nodes can be expanded.
As each production rule fires, the relevant portion of the semantic frame is propagated down into the expanded nodes.
Martinez
(<agent> ˆtype agent ˆname martinez
ˆjob driver ˆproper true
ˆsingular true ˆshade +5)
The driver
(<agent> ˆtype agent ˆname martinez
ˆjob driver ˆdefinite true
ˆsingular true ˆshade 0)
A private
(<agent> ˆtype agent ˆname martinez
ˆrank private ˆdefinite false
ˆsingular true ˆshade -2)
Figure 5: Subset of possible case frame expansions for object driver.
Thus, every node in the tree contains a pointer to the specific
aspect of the semantic frame from which it was created.
For example, in Figure 6, the NP node of ”the mother”
contains in it a pointer to the frame <patient> from Figure 4. By keeping semantic content localized in the tree, we
allow the gesture and speech synthesis modules convenient
access to needed semantic information.
For any given state and event, there are a number of theoretically valid realizations available in the lexicon. Instead of attempting to decide which is most appropriate at
any stage, we adopt a strategy similar to that introduced by
(Knight & Hatzivassiloglou 1995), which puts off the decision until realization is complete. We realize all possible
valid trees that correspond to a given semantic input, and
store the fully constructed trees in a forest structure. After
all such trees are constructed we move on to the final stage.
5.5 Ranking
In this stage we examine all the trees in the forest structure
and decide which tree will be selected and sent to the speech
synthesizer. Each tree is given a rank score based upon the
tree’s information content and emotional quality.
The emotional quality of each tree is calculated by computing the distance between the emotional attitudes of the
speaker toward each object, and the emotional shade that the
realization casts on each object. Realizations that cast emotional shades on objects that are more similar to the agent’s
attitudes toward those objects are given higher scores.
The information content of each tree is judged simply by
how much of the semantic frame input is expressed by the
realization. Thus, realizations that do not explicitly mention the agent (through passivization), for example, are given
lower scores.
The score of each tree is calculated by recursively summing the scores of the nodes along the frontiers of the tree,
and then percolating that sum up to the next layer. Summing
and percolating proceeds until the root node is given a score
that is equivalent to the sum of the scores for the individual nodes of that tree. The tree with the highest root node
score is then selected and passed to the speech synthesis and
gesture modules.
5.6
Sequencing Issues
There are a number of issues relating to generation being
part of the deliberate behavior of an agent engaged in taskoriented dialogue. The previous discussion in this section
described the normal process of utterance generation, from
the point at which a goal was proposed until speech was produced. There are, however several cases in which this cycle
does not follow the straightforward path. First, some proposals may later be retracted. For instance, if a goal to address
a request is selected in preference to a goal to acknowledge
the request and fully realized, the goal to acknowledge will
be dropped, since addressing will also count as (indirectly)
acknowledging. Some communicative goals can not be immediately realized, for example communicating content to
characters who are paying attention to the agent. In this case,
one must first adopt and realize a goal to get the attention.
Sometimes a goal must be dropped during the output-speech
operator. For instance if the agent realizes that there is no
need to say anything or doesn’t know how to say what it
wants to. In this case, the agent will produce a verbal disfluency (e.g., “uh”) and continue on with a new realization
goal. Finally, “barge-in” capability is provided by allowing
the agent to back out of an existing output-speech operator
in favor of attending to the speech of others. If the goal still
remains after interpreting the interruption, the agent will readopt it and eventually produce the interrupted utterance. If
on the other hand, the motivations for the goal no longer
hold, the goal will be dropped.
Currently there is a limited facility for multi-utterance discourse plans. For certain content, such as a description of a
charged event or a rejection of an order, the agent plans multiple utterances to give rationale, either assigning causality
or giving explanations and counterproposals. In this case,
single sentence generation is carried out, as normal, but
strong motivations are set up for future utterances, which
will directly follow, unless the agent is interrupted.
6 Summary
. We adopt a hybrid approach to NL generation for virtual characters in a complex interactive environment. For
some simple characters that will have only a limited range
of choices of what to say, we use pre-calculated prompts
and simple templates. For more sophisticated characters,
more complex generation techniques are needed to behave
appropriately given the rich structure of social interaction
and agent emotions that are being tracked. Within the agents
also, multiple methods are used, including prompt generation for very specific situations, selected sentence plans for
intermediate degrees of flexibility while still allowing complex utterance structure, and sentence planning for fully general coverage.
Generation covers simple cases of reactive feedback and
turn management as well as complex representations of sequences of events, negotiation moves and emotional affect.
At this point, we still feel that it is best to keep a fairly tight
coupling between generation and dialogue functions, given
ROOT
S
<event>
NP
<agent>
"the driver"
VP
<event>
"collided with"
S
<event>
NP
<patient>
"the mother"
NP
<agent>
"the driver"
VP
<event>
"smashed into"
S
<event>
NP
<patient>
"the mother"
NP
<patient>
"the mother"
VP
<event>
"was hit"
Figure 6: A subset of the forest output of realization.
the fairly broad range of fairly quickly changing information that can affect generation. The SOAR architecture is
well suited for this, allowing declarative information to be
made available for the use of either module as processing
is continuing. In the future, we are planning a number of
extensions, including information-structure based sentence
planning, more elaborate discourse planning, and statistical
sentence plan selection and ranking.
Acknowledgements
The work described in this paper was supported by the Department of the Army under contract number DAAD 19-99D-0046. Any opinions, findings and conclusions or recommendations expressed in this paper are those of the authors
and do not necessarily reflect the views of the Department of
the Army. We would also like to thank the rest of the MRE
team for providing a stimulating environment within which
to carry out this research, particularly the others contributing
to aspects of the system that are used here.
References
Allen, J. F.; Ferguson, G.; and Stent, A. 2001. An architecture for more realistic conversational systems. In Intelligent User Interfaces, 1–8.
Allwood, J.; Nivre, J.; and Ahlsen, E. 1992. On the semantics and pragmatics of linguistic feedback. Journal of
Semantics 9.
Baker, M. 1994. A model for negotiation in teachinglearning dialogues. Journal of Artificial Intelligence in Education 5(2):199–254.
Blaylock, N.; Allen, J.; and Ferguson, G. 2002. Synchronization in an asynchronous agent-based architecture for
dialogue systems. In Proceedings of the 3rd SIGdial Workshop on Discourse and Dialogue, 1–10. Philadelphia: Association for Computational Linguistics.
Clark, H. H. 1996. Using Language. Cambridge, England:
Cambridge University Press.
Dillenbourg, P.; Traum, D.; and Schneider, D. 1996.
Grounding in multi-modal task-oriented collaboration. In
Proceedings of the European Conference on AI in Education.
Dorr, B. J. 1989. Lexical Conceptual Structure and Generation in Machine Translation. In Proceedings of the Ninth
Annual Conference of the Cognitive Science Society, 74–
81.
Fleischman, M., and Hovy, E. 2002. Emotional variation in
speech-based natural language generation. In Proceedings
of The Second International Natural Language Generation
Conference (INLG’02).
Gratch, J., and Marsella, S. 2001. Tears and fears: Modeling emotions and emotional behaviors in synthetic agents.
In Proceedings of the Fifth International Conference on
Autonomous Agents.
Grosz, B. J., and Sidner, C. L. 1986. Attention, intention,
and the structure of discourse. Computational Linguistics
12(3):175–204.
Knight, K., and Hatzivassiloglou, V. 1995. Two-Level,
Many-Paths Generation. In Proceedings of the 33rd Annual
Meeting of the Association for Computational Linguistics
(ACL-95), 252–260.
Laird, J. E.; Newell, A.; and Rosenbloom, P. S. 1987.
SOAR: an architecture for general intelligence. Artificial
Intelligence 33(1):1–64.
Langkilde, I., and Knight, K. 1998. Generation that Exploits Corpus-Based Statistical Knowledge. In Proceedings of COLING-ACL ’98, 704–710.
Larsson, S., and Traum, D. 2000. Information state and dialogue management in the TRINDI dialogue move engine
toolkit. Natural Language Engineering 6:323–340. Special
Issue on Spoken Language Dialogue System Engineering.
Marsella, S., and Gratch, J. 2002. A step towards irrationality: using emotion to change belief. In In proceedings of
AAMAS 2002: First International Joint Conference on Autonomous Agents and Multi-Agent Systems.
Matheson, C.; Poesio, M.; and Traum, D. 2000. Modelling
grounding and discourse obligations using update rules. In
Proceedings of the First Conference of the North American
Chapter of the Association for Computational Linguistics.
Novick, D. 1988. Control of Mixed-Initiative Discourse
Through Meta-Locutionary Acts: A Computational Model.
Ph.D. Dissertation, University of Oregon. also available as
U. Oregon Computer and Information Science Tech Report
CIS-TR-88-18.
Rickel, J., and Johnson, W. L. 1999. Animated agents for
procedural training in virtual reality: Perception, cognition,
and motor control. Applied Artificial Intelligence 13:343–
382.
Rickel, J.; Marsella, S.; Gratch, J.; Hill, R.; Traum, D.; and
Swartout, W. 2002. Toward a new generation of virtual humans for interactive experiences. IEEE Intelligent Systems
17.
Sidner, C. L. 1994. An artificial discourse language for
collaborative negotiation. In Proceedings of the Fourteenth
National Conference of the American Association for Artificial Intelligence (AAAI-94), 814–819.
Swartout, W.; Hill, R.; Gratch, J.; Johnson, W.; Kyriakakis,
C.; Labore, K.; Lindheim, R.; Marsella, S.; Miraglia, D.;
Moore, B.; Morie, J.; Rickel, J.; Thiebaux, M.; Tuch, L.;
Whitney, R.; and Douglas, J. 2001. Toward the holodeck:
Integrating graphics, sound, character and story. In Proceedings of 5th International Conference on Autonomous
Agents.
Traum, D. R., and Allen, J. F. 1994. Discourse obligations
in dialogue processing. In Proceedings of the 32 Annual
Meeting of the Association for Computational Linguistics,
1–8.
Traum, D. R., and Hinkelman, E. A. 1992. Conversation
acts in task-oriented spoken dialogue. Computational Intelligence 8(3):575–599. Special Issue on Non-literal language.
Traum, D. R., and Rickel, J. 2002. Embodied agents for
multi-party dialogue in immersive virtual worlds. In Proceedings of the first International Joint conference on Autonomous Agents and Multiagent systems, 766–773.
Traum, D. R. 1994. A Computational Theory of Grounding in Natural Language Conversation. Ph.D. Dissertation,
Department of Computer Science, University of Rochester.
Also available as TR 545, Department of Computer Science, University of Rochester.
Walker, M. A., and Whittaker, S. 1990. Mixed initiative in
dialogue: An investigation into discourse segmentation. In
Proceedings ACL-90, 70–78.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Related manuals

Download PDF

advertisement