Stefan Kostadinov, Georgi Petkov and Maurice Grinberg
Central and Eastern European Center for Cognitive Science, New Bulgarian University
Montevideo str. 21, Sofia 1618, Bulgaria
Keywords: ECA, cognitive agent, software assistant, context-sensitivity, priming effects.
Abstract: A working model of an ECA with cognitive capabilities based on the DUAL cognitive architecture is
described. The cognitive model used inherits the advantages of a high context-sensitivity, general and
episodic memory, and reasoning by analogy of the DUAL/AMBR model. However, several crucial new
mechanisms are proposed which allow for the continuous functioning of the agent and the completion of
several question-answer cycles with meaningful priming and context effects. This paper presents these
mechanisms and discusses the results of simulations of a user-agent interaction session.
RASCALLI (Responsive Artificial Situated
Cognitive Agents Living and Learning on the
Internet) is a FP6 EC project (see
http://www.ofai.at/rascalli for more information)
aimed at the development of a platform whose
purpose is to help users to search for information in
Internet and in large data-bases and ontologies by
communicating with an Embodied Conversational
Agent (ECA). This ECA should be able to
understand and answer questions, look for and find
information in Internet, but also memorize the
interactions with the user and the environment, and
learn from its experience. Thus it will start to know
its user and his/her preferences and adapt its
activities in order to achive better completion of the
given tasks.
In order to naturally conceptualize and model
Rascalli’s virtual life in a virtual environment a
‘human’ metaphor has been adopted.
The mind – specialized to Rascalli’ specific
knowledge structure and tasks – communication
with its owner, type of knowledge (e.g. music),
events etc. The mind operates only on represented
knowledge and has only a mediated connection to
the body and the environment. Thus it contains a
partial, selected representation of the environment at
abstract conceptual level and experiential memories
related to specific episodes: interaction of Rascalli
with user(s), other Rascalli and the environment.
The interaction with the environment and the body is
mediated by the sensory-motor layer.
The Sensory-Motor Layer consist of two main
parts – the Perception Layer that selects the
information provided by the Sensors (e.g. the
translation of specific question from the user) and
translates this information in the symbolic form
required by the mind and the Action Layer that
translates action commands from the symbolic form
used by the mind into specific command to the body.
The body (e.g. specific tools for translating a
question or send a query to DB) consists of various
sensors and effectors which allow Rascalli to acquire
information from the environment and to perform
actions in it.
Rascalli acts in an environment, which is
defined as everything outside the Rascalli like the
user(s), other Rascalli, knowledge bases (KB),
external tools that would be able to function without
the Rascalli, etc.
The work, presented in this paper mainly focuses
on the mind. Where necessary for the explanation of
the integration of the mind in the general Rascalli
platform some communication tools with the
Sensory-Motor layer will be mentioned and their
function expained.
The core of the Rascalli, their mind, is based on
the cognitive architecture DUAL and the analogy-
making cognitive model AMBR (Kokinov, 1994,
Kokinov & Petrov, 2001). The mind includes a Long
Kostadinov S., Petkov G. and Grinberg M. (2008).
In Proceedings of the Fourth International Conference on Web Information Systems and Technologies, pages 325-328
DOI: 10.5220/0001529503250328
Term Memory (LTM) where general and episodic
knowledge is stored and a Working Memory (WM)
which is the active part of the LTM, perceptual input
and goals. LTM contains concepts (including
relations) as well as instances of concepts, organized
in coalitions to represent tools, episodes of
interaction with the owner, already acquired
knowledge, etc.
The DUAL architecture consists of a large number
of relatively simple interconnected hybrid
(connectionist and symbolic) micro agents. The
main advantages of the architecture are its context-
sensitivity, based on spreading of activation,
dynamic and emergent symbolic computations.
The main purpose of DUAL/AMBR
development has been the modeling of human
analogy making (Kokinov & Petrov, 2001). Various
simulations have been performed with the AMBR
model and compared successfully to empirical data
of analogy-making related to structural constraints,
context effects, and blending of memory episodes.
The following section discusses the added
mechanisms that allow the mind to perform cycles
of perception-action-communication.
As stated in the first section, the mind is part of
Rascalli platform together with the body. This
section is about the mechanisms that allow the mind
of Rascalli to function inside the general platform
dealing with a question (perception);
extracting information from its own memories
or from an source in the Environment (action
and formation of a solution space);
selecting the right solution (judgment and
decision making);
presenting it to the user (communication);
store the interaction as an episode in LTM
(evaluation and learning);
being ready for the next question without losing
the context of the previous one (continuous
functioning in a given context).
3.1 DUAL/AMBR Mechanisms
As mentioned above, DUAL/AMBR is built of a
relatively large number of interconnected DUAL
micro agents. An utterance is represented in a
structured form (as a coalition of micro agents) and
in order to be ‘perceived’ by Rascalli it must be
attached to the INPUT and GOAL nodes. The micro
agents representing the question become target
micro agents (which comes from the terminology
used in analogy research). The INPUT and GOAL
nodes are the only source of activation of the
architecture, so they activate the question coalition
and via them the concept level micro agents to
which they are linked. The concept micro agents
activate their instances through the inverse links.
Thus, activation spreads throughout LTM and the
micro agents which become active enough enter
WM and start participating in the analogy mapping
and transfer mechanisms. These mappings range
from direct correspondence to distant analogical
objects which allow the transfer of knowledge from
episode in domains different from the utterance
A mechanism based on anticipation, first
introduced in a robot implementation has been
adapted for Rascalli (Petkov et al., 2006).
3.2 Specific Knowledge Transfer
The DUAL/AMBR mapping mechanisms along with
the added anticipatory mechanism (Petkov et al.,
2006), are too unspecific and cannot lead to
knowledge transfer. Thus new mechanisms had to be
developed, on the first place – knowledge extraction
The utterances must be represented in a form
which contains information about the provided
details and, if present - specific answer expected
(e.g. a name of a music album or a child, see the
examples in Section 4). This form is provided
through NLP analysis by the input processing tool
that handles the utterances from the user.
Thus the utterances presented to the mind can
have two tags - ‘:of-interest’ for the elements of
information given and ‘:question’ to define what is
specifically asked for if the latter can be extracted
from the question.
The specific knowledge transfer mechanism
comes into play after one of the arguments of a
certain relation is mapped. Then, the other
arguments are directly transferred after verifying
whether the first argument has tag ‘:of-interest’. At
WEBIST 2008 - International Conference on Web Information Systems and Technologies
the same time, the extracted information can replace
some empty placeholders that have the tag
‘:question’. This new mechanism works locally and
in parallel with all other mechanisms. The relevance
requirement, however, still holds because knowledge
retrieval is constrained in two ways: first, transferred
micro agents should be sufficiently active (i.e.
relevant); and second, the tag ‘:of-interest’ should be
present in the utterance elements for a transfer of
specific information.
3.3 Action Transfer Mechanism
The final mechanism needed to close the perception-
action-communication cycle is the selection and
sending of an action command. It is triggered by the
anticipated cause-relations that are linked to the
GOAL node(s) (Petkov et al., 2006). The cause-
agents, as indicated by their name, represent causal
relations. If a cause-agent is linked to a goal agent
(e.g. ‘find-album’), it receives the ‘close-to-goal’
message. If a ‘close-to-goal’ cause-agent participates
in a winner-hypothesis, it checks its antecedents for
action micro agents (micro agent describing an
action). If all the above conditions are met, the
action mechanism executes the action.
To put it simple, when a whole structure from
INPUT to GOAL, supported by enough winner-
hypotheses establishes, the respective actions would
be triggered for execution. The action is sent to the
Sensory-Motor layer, that further processes it and
sends it to the appropriate tool.
3.4 WM Cleanup and Learning
The capability of Rascalli to give reasonable,
context-sensitive, and flexible answers to simple
questions relies on previous knowledge in LTM.
Without the possibility to acquire new knowledge
and to modify the existing one the system would be
rigid and limited.
Thus, various mechanisms for working memory
cleanup and episode storage have been developed.
They can be summarized with the following
algorithm: (1) Define the moment when the goal is
achieved. After that: (2) erase all current
correspondence hypotheses. (3) Delete all markers in
all concepts. (4) Terminate all suspended symbolic
operations. (5) Create a new episode with all the
elements from the current one including the answer
and the user evaluation. (6) Adjust/create new
inverse links from concepts to instances.
Equipped with these routines for WM cleanup
and episode storage, the system is able to work
continuously, without interruption between the
cycles; it enriches its memory with new information
after each session, and it is able to support and use
the context of a continuous conversation.
All these abilities of Rascalli are demonstrated
with the simulation, presented in the next section.
3.5 Mind and Body
As described above the body of the Rascalli
platform provides an interface to various tools for
communication, exploration and information
acquisition. The tools and the mind communicate via
a sensory-motor layer that translates the agents from
the mind into RDF (see http://www.w3.org/RDF/ for
details) messages to the tools and vice-versa. The
tools themselves carry out various tasks – translating
natural language into RDF graphs, translate RDF
graphs into natural language and voiced by Rascalli,
search in DB, consult Google, etc.
The Sensory-Motor Layer essentially translates
RDF graphs into DUAL micro agent structures and
vice-versa. The Action Layer additionally decides
which tool to use based on the RDF command. This
process is completely automated, as the mind’s
internal representation format and the RDF ontology
have a similar structure (e.g. semantic graph).
The current implementation of the mind deals
with three basic tools – for input processing,
database search and output of messages to the user.
This is the minimal set of tools required for Rascalli
to understand a request from the user, undertake
some action(s) to satisfy this request and finally
report the answer back.
The scenario demonstrating the system capabilities
consists of a dialog of five utterances in the music
domain – artists and details about their personal
lives like religion, children, etc.
The first utterance is: “Tell me something about
Britney Spears”. The input processing tool processes
the words and sends the message representation to
the input of the mind. Britney Spears is of interest to
the mind, so it tries to transfer information and link
it to the Britney Spears. The mind has in its LTM
information about Britney Spears so it is activated
by the question and is transferred by the anticipation
transfer mechanism described in Section 3 and the
parts in this information compete among them.
Eventually, the information about the album
Blackout wins the competition, as it is considered
most relevant and it is sent as answer to the user.
The second utterance is a question: “Who are
the children of Madonna?” It can be noticed that this
time the utterance is specific about what is needed –
the names of Madonna’s children – so the node
representing it has the tag ‘:question’. The rest of the
message has the tag ‘:of-interest’.
The system tries to replace ‘child’ (with tag
‘:question’, subsection 3.3) with information from
LTM. We assume that this information is available
to the mind so it is represented and attached to the
corresponding concepts of LTM.
The third utterance from the simulation is the
same as the first one: “Tell me something about
Britney Spears.” One option for the mind is to
answer as in the first question by giving the name of
an album. But its internal state is determined be the
second question related to the children of Madonna.
There is no information about the children of
Britney Spears in LTM so the mind primed by the
second question decides to search for it in DB where
such type of information is available. The command
sent to the data source search tool contains the
Britney Spears, the ‘has-child’ predicate and the
‘child’ as something to be filled in. The former two
are marked with ‘:of-interest’ tag and the latter with
the ‘:question’ tag. This information allows the data
source search tool to transform this message into a
search in the musical DB with key words ‘child’ and
‘Britney Spears’, the answer is completed with the
new information and sent to the user via the output
The fourth utterance is again a specific
question: “What is the religion of Madonna?” The
mind has such information, so it transfers it to the
target and thus provides the answer to the user.
The fifth utterance completes the priming
demonstration of the scenario. It is again the same as
the first and the third question: “Tell me something
about Britney Spears”. Again the mind has this
information in the LTM and directly provides the
answer – Britney is Christian.
In this paper we presented a full working model of
the mind of a future ECA based on the cognitive
architecture DUAL augmented with a number of
new mechanisms. The agent is able to carry on a
simple conversation consisting of a series of
questions and displays context sensitivity in its
answers – an essential trait for a more natural and
flexible conversation with a user.
The performance observed is a combination of
DUAL/AMBR mechanisms and a set of newly
developed ones based on the main principles of this
cognitive architecture.
The simulation demonstrates that the major
mechanisms needed for realistic situations are
available in Rascalli’s mind. Rascalli can encode the
incoming information, can reason using cognitive
mechanisms, can act according to the tasks, and can
learn and adapt itself.
The newly developed agent will be integrated in
the general Rascalli platform developed in the
Rascalli project and efforts are currently in progress
to refine the automatic question encoding for at least
a limited set of simple questions.
This work was supported by the Rascalli FP6 project
with EC.
Kokinov, B. & Petrov, A. (2001) Integration of Memory
and Reasoning in Analogy-Making: The AMBR
Model. In: Gentner, D., Holyoak, K., Kokinov, B.
(eds.) The Analogical Mind: Perspectives from
Cognitive Science, Cambridge, MA: MIT Press.
Petkov, G., Naydenov, Ch., Grinberg, M., Kokinov, B.
(2006). Building Robots with Analogy-Based
Anticipation. In: C. Freksa, M. Kohlhase, and K.
Schill (eds.) Proceedings of the 29th German
Conference on Artificial Intelligence (KI-2006), LNAI
4314, Bremen, Germany, 2007. Springer, pp. 72-86.
Krenn, B. et al. (2007) A Smart Music Companion. In:
AAMAS 2008 Special Track on Virtual Agents,
WEBIST 2008 - International Conference on Web Information Systems and Technologies