Overview of 'Planning Speech Acts'

OVERVIEW OF "PI..liNNING SPEECli ACI'S"

c. Raymooo Perrault aril Philip R. CohEn Dept. of Canputer SciEnce University of Toronto

We are developing aoo testing a computer model of conversation by constructing a program that engages its user in purposeful conversation to help him perform sane task. Conversation is viewed as a sequence of actions performed by the participants, intentionally affecting each other's model of the world, primarily their beliefs and goals. These actions can only be performed if certain conditions hold.

We show here how sane of these ~ acts (Searle [1969] ) can be described as operators in problem-solving systems such as S'IRIPS (Fikes aril Nilsson[l97l]), aril how they can be incorporated into plans with non-linguistic acts such as scheduling trips or opening doors. Language analysis and generation can thus be related to problems of plan generation, execution, aril recognition. Our work so far has concentrated on the speech act generation aspect.

Consider for example a REQUEST fran a speaker SP to a hearer H that H should perform a certain action ACT. SP would like H to make Acr one of his goals by having him recognize that SP wants him to. Certain conditions must be satisfied before the REQUEST can be usefully performed:

1. SP believes that H can do Acr
2. SP believes H believes H can do Acr,

3) SP wants to perform the REQUEST .

Coooition 2) excludes requests when H may be lacking necessary information although he may be familiar with all the necessary procedures . Want preconditions like 3) will appear on all actions performed by human agents. If one agent SP is planning for another agent H to perform Acr, SP cannot assume that Acr is a goal of H; this want precondition of Acr is precisely what a REQUEST is meant to satisfy. (In fact, we assume that H's believing that SP wants him to do Acr is the effect of the REQUEST, rather than just H's wanting to do Acr, aril we postulate an intermediate act to bridge this gap.) As well as RECUEST, our system includes a speech act INFORM whose effect is that the hearer believes the speaker believes sane proposition P is true. Again, a mediating act is used to proouce the hearer's believing P.

System Overview.

Our current system uses an object-centered representation developed by Levesque [1977], implemented using partitioned semantic networks (Heoorix [1975] ), a semantic network interpreter that executes programs in the net, aril an interactive network definition package. The system is ceded in SPITBOL on an IBM 370/165 II Lmder TSO.

Memory Organization.

We use Heoorix' spaces both as manipulable objects and as contexts for procedure execution. A space is identified as representing the "belief" or "want" of an agent S by placing it as the value of a case in a proposition stating "S believes" or "S wants". The space representing the system's beliefs (SB) includes spaces containing the system's beliefs about the system's wants (SffiW) aril the system's beliefs about the user's beliefs (SBUB) that in turn a:mtains the user's wants (SBUBUW). 'Ihese spaces can share information, and the interpreter can add new levels of nesting as necessary, aril test arbitrarily deeply nested propositions. Plans for the various agents are created in the want spaces.

An Example

Let the world consist of a room with a swinging door and a lock. System is outside and user is inside. Assume acts to move through an open door, push an unlocked door, mlock the door with a key, aril fetch a key. Also assume that the system believes that both it and the user know the state of the world. Our current system can plan the following utterances to achieve its being inside:

1. I request you to unlock the door.

2. I request you to push the door.

3. I want you to unlock the door.

4. I want you to push the door.

Utterances 1. aril 2. are RECUESTs to perform actions, aoo 3. aril 4. are INFORMs of WANTs. If the system did not believe the user knew the location of the key, it would issue an appropriate INFORM. Further details are available in Perrault aril Cohen [ 19 77] ·

Concluding Remarks.

Other natural language systems possess the means for generating utterances (e.g. A'IN's) but not the motivation. Ours, on the other hand, plans what it should say, but currently does not decide how to say it. We are investigating the interactions this approach has with plan recognition and strategies for being helpful, aril trying to extEnd it to the planning of references.