A MODEL OF SPEECH ACT PLANNER ADAPTED TO MULTIAGENT UNIVERSES 
Daniel Rousseau, Guy Lapalme 
Department of Computer Science and Operational Research 
University of Montreal, Montreal, Canada 
E-mail: roussead@iro.umontreal.ca, lapalme@iro.umontreal.ca 
Bernard Moulin 
Department of Computer Science 
Laval University, Quebec, Canada 
E-mail: moulin@vm.ulaval.ca 
A multiagent universe is characterized by many agents who cooperate (i.e. share a 
common goal) or compete (i.e. have conflictual goals). With cooperation, an agent may take 
advantage of other agent's abilities and increase her chance to reach her goals. But 
competition between agents decreases the opportunities to reach her own goals, because one 
agent's action might interfere with the ones of the other agent. So, it is very important to 
cooperate when it is possible and to negotiate when there are conflictual goals. 
Conversation is a form of cooperation by means of interrelated speech acts, also called 
illocutionary or linguistic acts, between two or more agents. Each speech act is not executed 
immediately, but is planned by an agent in order to reach some goal in the conversation 
context. In fact, a speech act is for an agent the best mean to transmit her mental states to 
other agents, to try to change their mental states and to affect their behavior in the same 
direction that she wants. A minimal level of cooperation is necessary to achieve a coherent 
and satisfying conversation between all participants. Before executing a speech act, an agent 
must take into consideration many factors such as the context of utterance, including previous 
speech acts of the current conversation and the participants' mental states, and shared rules of 
communication such as Gnce's (1975) conversational maxims. During a conversation, it is 
also possible for an agent to execute non-linguistic speech acts. 
In this context, we propose a model of speech act planner to be used in a cooperative 
responses generation system. The model can explain an agent's general behavior during a 
conversation involving two agents or more. It deals with the reasoning process between the 
perception of a situation and with the execution of a linguistic or non-linguistic action. It 
reasons about mental states belonging to herself or to other agents such as goals, intentions, 
beliefs, low-level actions and plans containing linguistic and non-linguistic actions. It 
integrates ideas from the following domains: intention in speech acts, structure of 
conversation and planning. 
Some people (Searle and Vanderveken 1985, Cohen and Levesque 1990) have assumed 
that the recognition of the speaker's intention and other mental states in producing an utterance 
is extremely important to understand its meaning, but have limited their work on one single 
speech act. Because a conversation is a temporal sequence of connected illocutionary acts 
110 
(Moulin, Rousseau and Vanderveken 1991) where each speech act plays a precise role in the 
context of other speech acts, some researchers have studied the structure of a conversation and 
they all agreed that there are several interrelated components in it and many subconversations 
of different types (Grosz and Sidner 1986, Litman and Allen 1987). The planning approach 
has been used by some scientists to produce speech acts in the context of a dialogue (Allen 
1983, Appelt 1985, Litman and Allen 1987, Lambert and Carberry 1992). 
Starting from the approaches mentioned above, the model of speech act planner we 
propose takes into consideration the following problems: multiagent planning, reasoning on 
other agents' mental states and on one's own mental states, recognition of intentions behind 
direct and indirect speech acts, use of plans integrating linguistic and non-linguistic actions, 
coherence of the conversation between two or more agents, handling of subeonversations and 
modeling of the conversational context. 
We assume that a conversation is a task shared by many agents that is the result of 
"reactive distributed planning" A conversational participant cannot control all the content of a 
conversation, but adapts its behavior by reacting to what has just been said. An agent that 
cannot reach a goal without communicating with another agent initiates a conversation by 
executing a planned speech act. In reaction to a speech act, at any time during a conversation, 
a participant must process it in many steps before executing an action, usually a speech act. It 
must interprete it by recognizing the intention of the speaker and update its beliefs. It must 
revise its goals, particularly when it must take turn according to the conversational state. In 
this case, it has a new communication goal, like answenng a question or clarifying the last 
speech act, for which it constructs a plan to reach it. This plan generally specify one or more 
speech acts to execute. The conversation is ended when there is no more reason to continue it, 
i. e. when the goal for which it was initiated has been reached or cannot be reached anymore. 
We will illustrate the planning process during a conversation in the context of a 
multiagent universe with the following example. Four cooperative robots, that can 
communicate with another robot and execute a particular task, are in a working shop. One 
robot can lift and move an object, including another robot, unless it weights too much. In that 
case, two robots can join forces to lift and move it. When there is a breakdown, a robot has to 
go to the repair station. 
Suppose that a robot, called Dartagnan, cannot move anymore. So, Dartagnan must go 
to the repair station after taking an appointment, but it cannot move. It has to ask another 
robot for help to reach the repair station. A sollicited robot that cannot lift and move 
Dartagnan will have to ask for help before declining assistance, unless it is already too busy. 
Taking an appointment and asking help take the form of speech acts in the context of a 
conversation. For each conversation, there is a goal that an initiator conversational agent tales 
to satisfy. For example, suppose that Dartagnan already has an appointment with the repair 
station. A conversation between Dartagnan and another robot, Portos, may look like this: 
Dartagnan: Portos, please carry me to the repair station by 10 o'clock. 
Portos: Sorry, I cannot help you. 
Dartagnan: OK, I will ask another robot to help me. 
111 
This quite simple conversation involves a lot of problems to be solved by the two 
conversational agents. Before initiating the conversation, Dartagnan has to decide which robot 
to talk to. Then, it has to choose what to say to initiate the conversation in order to allow 
Portos to recognize its intentions. After receiving the request, Portos must recognize 
Dartagnan's intention and update its knowledge base. Then, it has to decide if it can help 
Dartagnan or if not. In the example, Portos is too busy. Dartagnan, after receiving the 
response from Portos, must identify Portos' mental states and decide if it accepts this response 
or if it proposes a compromise. Here Dartagnan accepts the response and therefore ends the 
conversation. 
So, before and after each speech act, a lot of reasoning about mental states is necessary 
for all the conversational agents involved in a conversation. The speech act planner we 
propose tries to model all this necessary reasoning to get a conversation such as the above. 
We assume that we must take into account in our planner the general script of a 
conversation to generate coherent ones. A conversation is an evolving object. At a global 
level, each conversation follows the same conversational script that describes the possible 
states of a conversation and the possible transitions between them. A transition is insured by 
the application of a private plan by an agent. According to DeVito (1992), there are five 
possible states for a conversation: opening, feedforward (the general goal of the conversation 
is given), business (the substance of the conversation), feedback (synthesis of the discussion) 
and closing. Note that the feedback state is not always present in an instance of conversation. 
At these five basic states, we add two other ones, interruption and reopening, because we 
consider the cases when a conversation is interrupted by an agent. For instance, an agent may 
interrupt a conversation before answering a request. 
Each state can be considered as a conversational goal and be focalised into lower level 
conversational states during the conversation. These states may be different, depending of the 
conversation. A question is asked or a promise is done are examples of states. We cannot 
find these states in any order to have a coherent conversation. So, we assume that there are 
some rules to respect when we are in some conversational state. These rules precise the 
possible and necessary transitions between conversational states, the agent that is responsible 
of the transitions and the conditions, expressed in terms of mental states, associated with 
them. For example, when a question has just been asked to an agent, it has to answer it if it 
knows the answer and has no reason to hide it. If the question is not clear for it, it can initiate 
a clarification subdialogue by asking for some clarification. But it was not very conventional 
to give an order to the other agent at this time. Therefore, a conversation is like a construction 
made of LEGO TM blocks, where you can put a block of a certain type at a few places only. 
Each conversation has global parameters that respond to the following questions: who? 
(the conversational participants), why? (the goal behind the conversation), when? (the time), 
where? (the place) and how? (the type of the conversation). The answer to the question 'what?' 
is composed of the surface linguistic acts themselves and the corresponding speech acts. All 
these information is part of the conversational context, like all the knowledge and mentals 
states of the conversational participants. In fact, the conversational context contains all the 
information that is important for a speech act planner to participate to a conversation. 
112 
So the model of speech act planner we propose deals with two aspects of a 
conversation: its planning and its structure. Up to now, we have established the global 
framework of the theoretical model and tested it on dialogues like the one between Dartagnan 
and Portos. In the future, we will describe in more detail the different steps of the planning 
process, the necessary data structures and the acceptable transitions between conversational 
states. To handle conversations involving more than two agents, we expect to reduce them to 
many partially ordered dialogues (conversations between two agents only), where one agent 
participating to the conversation is conscious of all the dialogues. Finally, we will test our 
model by integrating it in a prototype that will simulate the behavior of many agents that will 
participate to a conversation in a given situation by producing speech acts from their mental 
states and their knowledge. 

REFERENCES 
ALLEN J. F. (1983), "Recognizing Intentions from Natural Language Utterances", in M. 
Bradie ad R. C. Berwick editors, Computational Models of Discourse, MIT Press, Cambridge, 
Massachusetts, chap. 2 pp. 108-166. 
APPELT D. E. (1985), "Planning English Sentences", Cambridge University Press, 
Cambridge, Great Britain. 
COHEN P. R., LEVESQUE H. J. (1990), "Rational Interaction as the Basis for 
Comuunication", in P. R. Cohen, J. Morgan and M. E. Pollack editors, Intentions in 
Communication, MIT Press, Cambridge, Massachustetts, Chap. 12, pp. 221-255. 
DEVITO J. A. (1992), "The Interpersonal Book", Sixth edition, HarperCollins Publishers, 
New York. 
GRICE H. P. (1975), "Logic and Conversation", in P. Cole and J. L. Morgan editors, "Syntax 
and Semantics, vol. III: Speech Acts", Academic Press, New York, pp. 41-58. 
GROSZ B. J., SIDNER C. L. (1986), "Attention, Intentions and the Structure of Discourse", in 
Computational Linguistics, Vol. 12, No 3, July-September 1986, pp. 175-204. 
LAMBERT L., CARBERRY S. (1992), "Modeling Negotiation Subdialogues", in 
Proceedings of the Conference of the 30th Annual Meeting of the Association for 
Computational Linguistics, Newark, Delaware, pp. 193-200, June 1992. 
LITMAN D. J., ALLEN J. F. (1987), "A Plan Recognition Model for Subdialogues in 
Conversations", in Cognitive Science, Vol. 11, pp. 163-200. 
MOULIN B., ROUSSEAU D., VANDERVEKEN D. (1991), "Speech Acts in a Connected 
Discourse: A Computational Representation Based on Conceptual Graph Theory", in 
Proceedings of the Sixth Annual Workshop on Conceptual Graphs, July 1991, Binghamton, 
New York, pp. 269-282. 
SEARLE J. R., VANDERVEKEN D. (1985), "Foundations of Illocutionary Logic", 
Cambridge University Press, Cambridge, Great Britain. 
