CHANGING TOPICS OF DIALOGUE FOR NATURAL
MIXED-INITIATIVE INTERACTION OF CONVERSATIONAL
AGENT BASED ON HUMAN COGNITION AND MEMORY
Sungsoo Lim, Keunhyun Oh and Sung-Bae Cho
Department of Computer Science,Yonsei University, Seoul, Korea
Keywords: Mixed-initiative interaction, Global workspace, Semantic network, Spreading activation theory.
Abstract: Mixed-initiative interaction (MII) plays an important role in conversation agent. In the former MII research,
MII process only static conversation and cannot change the conversation topic dynamically by the system
because the agent depends only on the working memory and predefined methodology. In this paper, we
propose the mixed-initiative interaction based on human cognitive architecture and memory structure. Based
on the global workspace theory, one of the cognitive architecture models, proposed method can change the
topic of conversation dynamically according to the long term memory which contains past conversation. We
represent the long term memory using semantic network which is a popular representation for storing
knowledge in the field of cognitive science, and retrieve the semantic network according to the spreading
activation theory which has been proven to be efficient for inferring in semantic networks. Through some
dialogue examples, we show the usability of the proposed method.
1 INTRODUCTION
Conversational agent can be classified into user-
initiative, system-initiative, and mixed-initiative
agent with the subject who plays a leading role when
solving problems. In the user-initiative
conversational agent, the user takes a leading role
when continuing the conversation, requesting
necessary information and services to the agent with
the web searching engine and question-answer
systems. On the other hand, in the system-initiative
conversational agent, the agent calls on users to
provide information by answering the predefined
questions. Although the various conversational
agents have been suggested with the user-initiative
or system-initiative, these techniques still have
significant limitations for efficient problem solving.
To overcome their limitations, the mixed-initiative
interaction has been discussed extensively.
Mixed-initiative conversational agent is defined
as the process the user and system; which both have
the initiative; and solve problems effectively by
continuing identification of each other’s intention
through mutual interaction when needed (Hong et al.,
2007, Tecuci et al., 2007). Macintosh, Ellis, and
Allen (2005) showed that mixed-initiative
interaction can provide better satisfaction to the user,
comparing system based interface and mixed-
initiative interface through ATTAIN (Advanced
Traffic and Travel Information System) (Macintosh
et al., 2005).
The research to implement mixed-initiative
interaction has been studied widely. Hong used
hierarchical Bayesian network to embody mixed-
initiative interaction (Hong et al., 2007), and Bohus
and Rundnicky utilized dialogue stack (Bohus and
Rudnicky, 2009). However, these methods, which
utilize predefined methodology depending on
working memory solely, can process only static
conversation and cannot change topics naturally.
In this paper, we focus on the question – how
could the conversational agent naturally change the
topics of dialogue? We assume that the changed
topics in human-human dialogue are related with
their own experiences and the semantics which are
presented in the current dialogues.
We apply the global workspace theory (GWT)
on the process of changing topics. GWT is a simple
cognitive architecture that has been developed to
account qualitatively for a large set of matched pairs
of conscious and unconscious processes. On the
view of memory, we define the consiousness part as
a working memory where the attended topics on the
107
Lim S., Oh K. and Cho S. (2010).
CHANGING TOPICS OF DIALOGUE FOR NATURAL MIXED-INITIATIVE INTERACTION OF CONVERSATIONAL AGENT BASED ON HUMAN
COGNITION AND MEMORY.
In Proceedings of the 2nd International Conference on Agents and Artificial Intelligence - Artificial Intelligence, pages 107-112
DOI: 10.5220/0002733501070112
Copyright
c
 SciTePress

working memory are candidates of next topics, and
the unconsiousness part as a long term memory. By
the broadcasting process of GWT, the most related
unconsious process is called and becomes a consious
process which means one of the candidates of next
topics in this paper. We model the unconsiousness
part (or long term memory) using semantic network
and the boradcating process using the spreading
activate theory.
2 RELATED WORKS
2.1 Global Workspace Theory
Global workspace theory models the problem
solving cognitive process of the human being. As
figure 1 shows, there is independent and different
knowledge in the unconscious area. This theory
defines the independent knowledge as the processor.
Simple work such as listening to the radio is possible
in the unconscious but complex work is not possible
in the unconscious only. Hence, he calls the
necessary processors through the consciousness and
solves the faced problem by combining processors
(Moura, 2006).
Figure 1: Global workspace theory.
An easy way to understand about GWT is in
terms of a “theatre metaphor”. In the “theatre of
consciousness” a “spotlight of selective attention”
shines a bright spot on stage. The bright spot reveals
the contents of consciousness, actors moving in and
out, making speeches or interacting with each other.
In this paper, the bright spot could be interpreted as
current topic of dialogue, the dark part on the stage
as the candidate topics of next dialogue, and the
outside of the stage represents the unconsciousness
part (long term memory).
2.2 Structure of Human Memory and
Semantic Network
Figure 2 shows the structure of human memory.
Sensory memory receives the information or stimuli
from the outside environment, and the working
memory solves problems with the received
information. The working memory cannot contain
the received memory in the long term since it store
the information only when the sensory memory is in
the cognition process of the present information.
Hence, the working memory calls the necessary
information from the long term memory when the
additional memory needed (Atkinson and Shiffrin,
1968).
Figure 2: Human memory structure.
The long term memory of human can be
classified into non-declarative memory which cannot
be described with a certain language and declarative
memory which can be. In the conversation agent, we
deal the declarative memory and construct the
declarative memory in the long term memory. The
declarative memory is divided into semantic
memory and episodic memory. Semantic memory is
the independent data which contain only the
relationship, and episodic memory is the part which
stores data related with a certain event (Squire and
Zola-Morgan, 1991).
In this research, for the domain conversation, we
transfer the needed keywords and the past
conversation record to semantic memory and
episodic memory respectively, and express
descriptive memory as semantic network. In the
field of cognitive science, semantic network is a
popular representation for storing knowledge in long
term memory (Anderson, 1976). Semantic network
is the directional graph which consists of nodes
connected with edges. Each node represents the
concept, and the edge represents the relationship
between concepts the nodes mean. Semantic
network is mainly used as a form for knowledge
symbol, and it is simple, natural, clear, and
significant (Sowa, 1992). Semantic network is
utilized to measure the relationship between the
created keywords during the present conversation
and the past conversation, and generate system-
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
108

initiative interaction when the past conversation with
the significant level of relationship is discovered.
2.3 Spreading Activation Theory
Spreading activation serves as a fundamental
retrieval mechanism across a wide variety of
cognitive tasks (Anderson, 1983). It could be applied
as a method for searching semantic networks. In the
spreading activation theory, the activation value of
each and every node spreads to its neighbouring
nodes. As the first step of the search process, a set of
source nodes (e.g. concepts in a semantic network)
is labelled with weights or activation values and then
iteratively propagating or spreading that activation
out to other nodes linked to the source nodes.
3 CHANGING TOPICS IN
CONVERSATIONAL AGENT
3.1 System Overview
In this paper, we proposed a method for changing
topic of dialogue in conversational agent. Figure 3
represents the overview of proposed method
adapting memory structure of human and global
workspace theory.
Figure 3: Overview of proposed method.
In the working memory, the information needed
to process the present conversation and the
candidates of next topics of dialogues are stored.
Long term memory is composed of semantic
memory and episodic memory. Semantic memory
expresses the relationship between important
keywords in the conversation, and episodic memory
stores the past conversation which is not completed.
The proposed method is to represent such two types
of memory into be a semantic network.
Conversation agent process the current dialogue
using the information in working memory. When the
current dialogue is ended, it selects some candidates
of next topics in the long term memory by
broadcasting using spreading activation. If there are
some topics which have more activation value than
threshold, they are called and become candidates of
next topics. Finally, the conversational agent selects
the most proper topics from the candidate topics.
3.2 Conversational Agent
In this paper, we adjust and adapt the conversation
agent using CAML (Conversational Agent Markup
Language) to our experiment (Lim and Cho, 2007).
CAML has designed in order to reduce efforts on
system construction when applying the conversation
interface to a certain domain. It helps set up the
conversation interface easily by building several
necessary functions of domain services and
designing the conversation scripts without
modifying the source codes of conversational agent.
The agent using CAML works as following.
1) Analyze the user’s answers and choose the
scripts to handle them
2) Confirm the necessary factors to offer service
which is provided by the chosen scripts. (If
there is no information of the factors, system
asks the user about them.)
3) Provide services
This agent use both stack and queue of the
conversation topics to manage the stream of
conversation, and provide different actions to the
identical input by the faced situation. System
initiative conversation with working memory is
already constructed; we focus on the system
initiative conversation using long term memory.
3.3 Semantic Networks and Spreading
Activation
Figure 4 shows the structure of semantic network we
used in this research. The internal nodes in the
network consist of semantic memories, the leaf
nodes episodic memories, and the edges show the
intensity of connections.
Formally, a semantic networks model can be
defined as a structure
),,,( WRCN
δ
=
consisting of
CHANGING TOPICS OF DIALOGUE FOR NATURAL MIXED-INITIATIVE INTERACTION OF
CONVERSATIONAL AGENT BASED ON HUMAN COGNITION AND MEMORY
109

0.76
0.85
0.95
0.90
0.45
0.78
0.81
0.93
0.81
0.68
0.87
0.83
0.85
0.97
0.89
0.92
0.96
Episodic
Memory
Semantic
Memory
Figure 4: Semantic networks.
z A set of concepts C = {C
s
, C
e
} which represents
the semantics of keywords and episodic memory
respectively,
z A set of semantic relation R = {R
s
Æ
s
, R
s
Æ
e
}
which represents the semantic relations between
keyword and keyword, and between keyword
and episodic memory respectively,
z A function
δ : Cⅹ C→ R, which associates a
pair of concepts with a particular semantic
relation,
z A set of weight functions W = {W
c
(c
x
), W
r
(c
x
,
c
y
)}which assign weights to concepts and
relations respectively.
The past conversation stored in episodic memory
C
e
goes up to working memory by broadcasting, and
we customize this process with applying spreading
activation theory. Searching network is done by BFS
(breath first search) algorithm using priority queue,
and calculates the level of relationship between the
corresponding node and the working memory when
searching from the node to the next node. Finally, it
calculates the level of relationship of episodic
memory, the leaf node, and if it has relationship over
a certain level, corresponding information goes up to
working memory.
Figure 5 shows the pseudo code of Semantic
network searching process. For the first stage of
spreading activation, it gets the initial semantics
which is appeared keywords in the current dialogue.
The activation values for these concepts are set 1.
Then the activation values are spread through the
semantic relations. The spreading values are
calculated as follow:
c
y
.value = c
x
.value*W
r
(c
x
,c
y
)*W
c
(c
y
)
procedure SpreadingActivation
input: C
s
// a set of semantics
in working memory
output: E // a set of episodic
memory
begin
Q.clear() // Q = priority queue
for each c
x
in C
s
Q.push(c
x
)
end for
while Q is not Empty
c
x
= Q.pop()
for each linked concept c
y
with c
x
c
y
.value =
c
x
.value*W
r
(c
x
,c
y
)*W
c
(c
y
)
if c
y
.value < threshold1
then continue
if isVisitedConcept(c
y
) is false
then Q.push(c
y
)
if c
y
is episodic memory
then E.insert(c
y
)
end for
end while
end proc
Figure 5: Pseudo code for Spreading Activation.
The weight function W
c
(c
x
) returns 1 when c
x
∈
C
s
, and returns the priority of episodic memory
when c
x
∈ C
e
.
Figure 6 represents a flow chart of the
conversation process in conversation agents. In this
paper, we use the 2 level of the threshold values. If
the level of relationship is over the threshold value 1,
the information has a significant link. If the level is
over the threshold value 2, it means that it has both a
significant link and urgency. If there is episodic
memory over the threshold value 2, the agent stops
the current conversation to deal the past
conversation. And if there is episodic memory with
only threshold value 1, the agent waits the current
conversation to be finished and then it precedes the
past conversation.
4 DIALOGUE EXAMPLES
We use the scheduling program domain to test the
utility of the proposed method. Agent needs 4
factors: the type of schedule, subject, time and
location. It leaves the conversation unhandled if the
factors are unknown because of user’s situation.
Hence, Semantic network for this scheduling
conversation agent have 4 types of internal nodes of
in the network: the type of schedule, subject, time,
location. The leaf nodes have episodes with 4 values.
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
110

Figure 6: Flow chart of the conversation process.
In this paper, we show how the scheduling program
works on 4 situations, when the episodic memory is
constructed.
The First scenario shows the general schedule
registration process. The agent asks the necessary
information and registers the corresponding schedule
when all information is obtained. The second
scenario shows the past schedule reprocessing.
Whenever the schedule information gathered, the
agent confirms that there is any correspondence, and
continue the past conversation if there is the same
data as showed in the scenario. In the third scenario,
there exists related conversation which has
activation value between threshold value 1 and
threshold value 2, so that the agent continues the
related conversation after the present conversation
had finished. If the activation value of related
conversation is bigger than threshold value 2, it
stops the present conversation and process the past
conversation ahead because it means the related
conversation has high priority.
1) Register the schedule
U: I got a dinner appointment.
A: Who will you meet?
U: I’m going to meet Gina.
A: What time will you meet her?
U: At 7 P.M.
A: Where do you meet her?
U: We decided to meet at Yeouido.
A: Okay. The new dinner schedule had registered.
2) Register unprocessed schedule
U: I got a meeting appointment with Jaenam.
A: Is that a dinner appointment at Shinchon you
told me on June 13.
U: Yes, that’s right.
A: Did you arrange the time to meet?
U: Yes. I decided to meet him at 6 P.M.
A: Okay. The dinner schedule had registered.
3) Process the related conversation
U: I and Kangseok will go out to drink.
A: What time will you meet him?
U: Tomorrow. At 8 P.M.
A: Where do you meet him?
U: At Apgujeong.
A: Okay. The new schedule had registered.
Did you arrange the place of tomorrow’s dinner
with Gina?
U: Not yet. I’ll tell you when I decide.
A: Okay.
4) Process the related conversation 2
U: I will study with Gina on this weekend.
A: Did you decide where you and Gina have
dinner tonight?
U: Oh! Let it be Shinchon.
A: Okay. The location had registered. When do
you study?
U: At 2 P.M.
A: Where do you study?
U: At Gongdeok.
A: Okay. The study schedule had registered.
CHANGING TOPICS OF DIALOGUE FOR NATURAL MIXED-INITIATIVE INTERACTION OF
CONVERSATIONAL AGENT BASED ON HUMAN COGNITION AND MEMORY
111

5 CONCLUSIONS
Former conversation agent provides mixed-initiative
conversation according to the predefined
methodology and it only depends on the working
memory so that only static conversation can be
processed. Hence, in this research, we have studied
to give an active function to conversational agent
that can change the topic of dialogues. The proposed
method models the declarative memory of long term
memory with the semantic network, and implements
the broadcasting process in global workspace theory
using spreading activation. By searching relevant
episode memory with current dialogue, the
conversational agent can change the topic of
dialogue naturally.
As showed in the dialogue examples, the
proposed method works according to the
relationship between the present conversation and
long term memory so that the various mixed
initiative conversation can be occurred.
Hereafter, it is necessary to form the semantic
network automatically by using the frequency of
appeared keywords during the conversation and
coherence of keywords. Also, the adaptation of
memory reduction function is required to calculate
the relationship smoothly.
ACKNOWLEDGEMENTS
This research was supported by the Conversing
Research Center Program through the National
Research Foundation of Korea(NRF) funded by the
Ministry of Education, Science and Technology
(2009-0093676).
REFERENCES
Anderson, J.R., 1976. Language, memory and thought,
Erlbaum, Hillsdale, NJ.
Anderson, J.R., 1983. A spreading activation theory of
memory. Journal of Verbal Learning and Verbal
Behavior, New York, Academic Press.
Atkinson, R.C. and Shiffrin, R.M. , 1968. Human
memory: A proposed system and its control processes,
The psychology of learning and motivation, Elsevier.
Bohus, D. and Rudnicky, A. I., 2009. The RavenClaw
dialog management framework: Architecture and
systems, Computer Speech & Language, Elsevier.
Hong, J.-H. et al., 2007. Mixed-initiative human-robot
interaction using hierarchical Bayesian networks. IEEE
Trans. Systems, Man and Cybernetics, IEEE.
Lim, S. and Cho, S.-B., 2007, CAML: Conversational
agent markup language for application-independent
dialogue system, 9th China-India-Japan-Korean Joint
Workshop on Neurobiology and Neuroinformatics.
Macintosh, A. et al., 2005. Evaluation of a mixed-
initiative dialogue multimodal interface, the 24th SGAI
International Conf. on Innovative Techniques and
Applications of Artificial Intelligence, Springer-Verlag.
Moura, I., 2006. A model of agent consciousness and its
implementation, Neurocomputing, Elsevier.
Sowa, J. F., 1992. Semantic networks revised and
extended for the second edition, Encyclopedia of
Artificial Intelligence, John Wiley & Sons.
Squire, L.R. and Zola-Morgan, S., 1991. The medial
temporal lobe memory system, Science, AAAS.
Tecuci, G. et al., 2007. Seven aspects of mixed-initiative
reasoning: An introduction to this special issue on
mixed-initiative assistants, AI Magazine, AAAI.
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
112