Figure 1
A socially appropriate letter for declining an invitation1
In order to clarify what I mean by "socially appropriate" messages and show how an agent could fail to meet that criterion, I will start with an example of a message which is socially appropriate. The context for the example is that one has been invited to speak before an academic department about one's research, but the scheduled time of the speech conflicts with a previous commitment, so one must decline. The note in Figure 1 is an appropriate response to such an invitation because it shows a fitting level of gratitude for the offer, in addition to communicating the basic message that the writer cannot accept the offer.
Figure 1
A socially appropriate letter for declining an invitation1
Without the knowledge that a human should show gratitude for offers he or she receives, an agent acting on behalf of such a person might respond merely, "I decline your offer." Upon receiving such a curt response, the inviter would probably feel insulted that the invitee did not value the invitation. Thus, an agent which interacts with humans on behalf of a human must have some knowledge of appropriate social communication (even if that knowledge is implicit in a procedure or data structure, e.g. a decision tree of response-letter templates indexed by the category of message received).
At first glance, the effects resulting from an induced belief may seem nearly innumerable. Certainly, the relations between these effects can be very complicated. But if one enumerates a small set of "effects of interest" and maps a limited set of connections, one can establish the core of a model which can be extended gradually. Figure 2 illustrates the network of abstract states which I used as a guide in developing the model. In the network, a hearer's belief in the content of a speech act often leads to hearer goals, emotions, beliefs about the personality traits of the speaker, and beliefs about the hearer's own rights and responsibilities. Furthermore, changes to a hearer's emotions about another person, or changes to his beliefs about another's traits, may cause changes in the hearer's attitude toward his relationship with that person. As an example of a chain of effects which conforms to this abstract description, consider the effects of the expression of thanks in Figure 1. The usual effect of a Thank act is a belief of the hearer that the speaker feels gratitude. Such a belief may then lead the hearer to have a goal to say, "You're welcome" in response. Also, the belief may give the impression that the speaker is conscientious (due to his acknowledgment of his social debt). In turn, such an impression may lead the hearer to begin to like the speaker (i.e. to feel a positive emotion toward him). If the feeling is mutual, the two individuals might enter into a cordial relationship.2
Figure 2
A network of abstract act effects
The network depicted in Figure 2 is useful for the analysis of the effects of many speech acts. Figure 3 illustrates several such analyses for acts which culminate in a positive effect on cordial relationships.3 (The model has similar analyses for acts such as Criticize and Threaten which have negative or corrective effects on cordial relationships.)
Figure 3 also illustrates a pair of act preconditions (or "appropriateness conditions"). Preconditions are as important to the success of social acts as to the success of physical acts. For example, inappropriate praise may appear awkward or even sarcastic to a hearer, which would thwart any motive the speaker may have to ingratiate himself.
As in many physical domains, social act effects and act preconditions often mesh to form standard sequences of interaction, such as a speaker's praising of a hearer followed by the hearer's thanking of the speaker for his praise. These standard sequences or scripts often create expectations to which one must be sensitive as a social participant. For example, a speaker who praises a hearer but does not receive any thanks in return is likely to believe that the hearer is ungrateful. (Failing to show gratitude for an invitation causes a similar effect, due to the failure of a similar script expectation.) Avoiding such perceptions seems to be a motivation for many social actions.
An interesting feature of the model is its representation of self-reinforcing cycles of interaction. In Figure 3, cordial relationships are shown to oblige cordial acts, which links the "top" of the model to the "bottom". Similarly, unfriendly acts often lead to adversarial relationships, which may induce the participants to act on opportunities to inconvenience each other. The model represents cordial relationships as mutual liking among participants, and it represents adversarial relationships as mutual dislike. Thus, an agent can escape a cycle by indicating to others in a relationship that feelings of like or dislike are not mutual. Indicating that feelings are not mutual should be as simple as performing an act from the cycle one wishes to move into. If the other persons in the relationship reciprocate, a new type of relationship becomes active and one enters a new cycle of interaction. This description appears to reflect human relationship interactions somewhat faithfully.
The note in Figure 1 was generated by LetterGen. In response to an input goal to decline politely, the agent suggests seven acts: Thank, Decline, Apologize, Make-excuse, Advise, Reassure, and Request. An organizational template is used to place these acts in the e-mail message in the order given above.
Unlike traditional planners, LetterGen does not generate all of its suggestions in a means-end manner. Only the Decline act is generated in that way. The Thank, Apologize, and Make-excuse acts are suggested in response to the agent's fear that the Decline act endangers the user's goal to have the addressee like him or her. (This danger would come from the failure of the addressee's expectation that the user be grateful and polite.) The Reassure act is planned in a similar way, as a reaction to a "fear" that the addressee may be skeptical about the user's advice. The Advise act is an expression of the user's personality trait of helpfulness. This expression of helpfulness is triggered when the agent notices that the addressee needs someone to replace the user in giving a speech. LetterGen has four general plan types:
The last three plan types are triggered opportunistically.
If one attempts to use LetterGen's plan types to understand why a particular speech act was included in a letter, one finds that often more than one plan type could have been involved. For example, the Thank act might have been included in the example of Figure 1 in order to lessen the social debt the invitee owes to the inviter, or to avoid insulting the inviter through curtness, or to reinforce the relationship, or simply out of polite habit. LetterGen's model allows for all of these different plans, but in practice only one plan is used to generate an act suggestion.
LetterGen is entirely rule-based. Scripts of standard act plans are encoded as chains of rules rather than as special data structures. Forty different speech acts are defined in the agent; many of these acts have more than one sentence template associated with them, to reflect the different types of content communicated by the acts (e.g. declining an invitation to speak versus declining a request for a document because it is out of print). There are approximately three hundred generalized states in the social model. The agent is able to generate a dozen message types, each with many variations due to the user's ability to accept or reject act suggestions.
The greatest limitation of the agent is the genericness of its sentence templates, although this must be weighed against the goal to prevent the agent from becoming too dependent on the user for background information. If the agent requires too much input, it loses its appeal as a work-saving device. (Notice that this genericness is not a problem which could be solved by replacing the templates with a low-level generator.)
Because cordial relationships require mutual feelings of liking, the simplest way of terminating such a relationship is for one to indicate one's dislike of the other person. A less direct method is to induce the other person to dislike oneself. This effect can be induced in a number of ways, such as making oneself appear to have bad traits (e.g. intrusiveness) and giving the impression that one does not respect the other person.
Given the goal of terminating a cordial relationship, LetterGen uses the Goal-pursuit plan type to suggest an Express-anger act, because it would give an impression that the writer is irascible and thereby cause the addressee to dislike him. Similarly, LetterGen suggests Denigrate (i.e. mention a bad trait of the hearer) because it gives an impression of pettiness, and Prohibit because it gives an impression of burdensomeness. Blame is suggested because it gives an impression of one's disrespect for the other person.
The four acts mentioned above are not "focussed" on a particular situation as the acts in the previous example were for the invitation. That is, at the current stage LetterGen does not know what to express anger about, what trait to denigrate, and so forth. Therefore, LetterGen is much more dependent in this case on the user to provide focussed content for the acts in the form of input text. Figure 4 illustrates an e-mail message generated from the four acts suggested plus focussed text (which LetterGen queried the user for in response to the user's approval of each suggestion). It seems possible that focussed situations could be inferred if LetterGen could use its model to parse the user's e-mail.4 But a history of interaction with the user and with his correspondents is likely to be more helpful.
Figure 4
A letter aimed at terminating a cordial relationship
Two related social agent systems are the Affective Reasoner (AR) [Elliott 1992] and the Oz Project [Reilly & Bates 1992]. AR's agents are customers and taxi drivers in a simulated TaxiWorld. The agents construe events and react to them in the emotion-laden manner one would expect in such a domain. The Oz Project is similar to AR in its emphasis on believably emotional agents (who create dramatic situations for the enjoyment of the user through their interactions with each other). In both systems, the agents not only model emotional reactions to their own surroundings but also reason about the likely emotional reactions of other agents.5 The agents represent beliefs, goals, and relationships as these relate to emotions. Yet, the agents interact only with other simulated agents, not with humans and not on behalf of humans. Such an approach has the practical benefit (but potential cost to verisimilitude) of avoiding the great variety of human interaction patterns one finds in speech acts and other forms of pragmatic communication.
Austin, J.L. 1962. How to do things with words. eds., J. Urmson and M. Sbisˆ. Harvard University Press. Back to text
Bruce, B. 1975. Belief systems and language understanding. Technical Report 2973. Bolt, Beranek, and Newman. Back to text
Elliott, C. 1992. The Affective Reasoner: A process model of emotions in a multi-agent system. Technical Report 32, Institute for the Learning Sciences, Northwestern University. Back to text
Etzioni, O., N. Lesh, and R. Segal. 1993. Building softbots for UNIX (preliminary report). Technical Report 93-09-01, University of Washington. Back to text
Moore, J.D. and C.L. Paris. 1994. Planning Text for Advisory Dialogues: Capturing Intentional and Rhetorical Information. Computational Linguistics (19) 4. 651-694. Back to text
Ortony, A., G. Clore, and A. Collins. 1988. The Cognitive Structure of Emotions. New York: Cambridge University Press. Back to text
Perrault, C.R. and J. Allen. 1980. A plan-based analysis of indirect speech acts. American Journal of Computational Linguistics (6) 3-4. 167-182. Back to text
Reilly, W. and J. Bates. 1992. Building Emotional Agents. Technical Report 143, Computer Science Department, Carnegie Mellon University. Back to text
Schank, R. and R. Abelson. 1977. Scripts, Plans, Goals, and Understanding: An inquiry into Human Knowledge Structures. Hillsdale, NJ: Lawrence Erlbaum Associates. Back to text
Tambe, M., W. Johnson, R. Jones, F. Koss, J. Laird, P. Rosenbloom, and K. Schwamb. 1995. Intelligent Agents for Interactive Simulation Environments. AI Magazine 16 (1). 15-39. Back to text
Vanderveken, D. 1991. Meaning and speech acts. Volume 1. New York: Cambridge University Press. Back to text
Wierzbicka, A. 1987. English speech act verbs. Sydney: Academic Press. Back to text
1Upper-case text indicates text provided by the user. Back to text
2For an act which has an effect on a hearer's rights, consider Permit or Prohibit. Back to text
3Effects on goals and rights have been omitted from Figure 3 in order to reduce graphical complexity. Chains of such effects are easily imagined for acts such as Request and Permit. Back to text
4A rule-based framework was chosen over other possibilities with the aim of making the model useful for both generation and interpretation. In future work, LetterGen will use its model to interpret e-mail as a way of reducing its dependence on the user. Back to text
5The models of emotion used in both systems are based on the theory presented in Ortony, Clore, and Collins 1988. Back to text
