A Computational Model of Small Talk

Timothy Bickmore

MAS 962 Discourse and Dialog for Interactive Systems

13 May, 1999

Abstract

Small talk is a valuable mechanism for managing the channel of communication and interpersonal distance between a user and an interface agent. A computational approach for an agent to engage in small talk is presented, including treatment of when the agent should begin small talk, when it should transition out of small talk into task talk, and how it should behave while engaged in small talk. A preliminary implementation for an agent acting as a real estate salesperson is described, and results from sample runs are presented and discussed.

1. Introduction

Human social encounters between individuals who have never met or are unfamiliar with each other are usually initiated by small talk in which "light" conversation is made about topics of general interest (e.g., weather, current events) or in which personal experiences, preferences, and opinions are shared. Even in business or sales meetings, it is customary to begin with some amount of small talk before "getting down to business" (at least in American culture). The purpose of this small talk is primarily to build rapport and trust among the interlocutors, and to establish their reputations (Dunbar, 1996). As a discourse function, small talk can be used to manage the channel of communication between participants (the phatic function of discourse; Jacobson, 1960) as well as the interpersonal distance between them (Tannen, 1984).

Software agents have even more of a credibility problem than human nonacquaintances do. In addition to overcoming the problems of unfamiliarity and general technophobia, users must establish the extent of an agent's capabilities and intelligence before conducting business with it. While this is not as much of an issue for entertainment applications, it is important for any information dissemination or advice-giving tasks and crucial in sales tasks.

One of the primary purposes of rapport-building in general is to establish that our interlocutor is, in some way, "just like us," that is, shares our basic values and beliefs, as well as sharing the ability to communicate. One way for a software agent to overcome the credibility problem is to embody the agent in the interface, that is, provide it with a human appearance and lifelike behavior. The Rea project at the MIT Media Lab (Cassell, et al, 1999) has as its goal the construction of an embodied, multi-modal real-time conversational interface agent. Rea implements the social, linguistic, and psychological conventions of conversation to make interactions as natural as face-to-face conversation with another person. Rea's task domain is real estate sales, a domain intentionally selected to provide opportunities for both task-oriented and social conversational frames of interaction. Within a typical sales encounter, it is important for Rea to engage in some amount of small talk to build rapport with the user before going into the "sales" conversational frame. The rapport may convince users that Rea does in fact understand their communicative behaviors, as well as convincing them that she knows what she is talking about in the real estate domain.

A second mechanism for establishing credibility and rapport is small talk. The effectiveness of small talk on the user's perception of a software agent has been demonstrated in text-only "chatterbot" systems such as ELIZA (Weizenbaum, 1966). There are many documented cases in which users have established on-going relationships with such systems, even when they knew that the system did not understand what they were typing (Turkle, 1995). Thus, a software agent that is both embodied and capable of small talk will have a much better chance of establishing a rapport with a user than a software agent that has neither of these features. This rapport, in turn, may actually allow the system to function more effectively, by enabling the user to rely on his/her natural social and linguistic skills.

A real-estate sales situation is a unique type of service encounter, in which the customer may feel uncomfortable (wanting to avoid commitment, i.e., threats to negative face) and the salesperson is not only trying to make the customer feel at ease for the purpose of information gathering but also to attempt to develop a long-term relationship with them. The basic questions that need to be addressed by a computational model of small talk in this context are:

When should the computer agent should begin small talk?
When should it transition out of small talk into real-estate task talk? and
How it should behave while engaged in small talk?

I begin this report by briefly reviewing the discourse literature related to small talk and work on agents which have been designed to engage in small talk. I then present an approach to having a computer agent engage in small talk with a user, and describe an implementation and results of using this approach. Finally, I discuss lessons learned from developing this model and future directions for research and development in this area.

2. Related Research

Although the topic of small talk cuts across many sub-disciplines of discourse, there has been very little written about this genre explicitly. I review the work that has been done in this area and then briefly summarize related work on the more general topic of casual conversation. Similarly, very little work has been done on the development of computational approaches to small talk. The implemented systems which come the closest to supporting small talk are the "chatterbot" systems (ELIZA and her descendants), and the techniques used and results from these are described.

2.1. Discourse Literature on Small Talk

In one of the earliest writings on the topic, Malinowski (1923) coined the term phatic communion to describe "a type of speech in which the ties of union are created by a mere exchange of words," that is, speech whose purpose is primarily to establish and maintain the social bonds of the interlocutors, rather than to convey information. Jakobsen (1960) used Malinowski's term in a slightly different way to describe the function of maintaining an open channel of communication between interlocutors (the "phatic" function) in his well-known "channel" model of communication.

Schneider (1988) did the first extensive study on the genre of small talk, at least as it occurs in several common situations involving chance meeting of strangers. His review of the literature provided two basic functions for small talk; one addresses positive face needs (fulfilling a socialization function, as in Malinowski's phatic communion) and one which addressed negative face needs (avoiding the embarassment or "danger" of silence with a stranger). Schneider transcribed over 30 conversations and performed a detailed characterization of small talk conversations. The typical small talk conversation consists of a series of conversational segments, each with the following sequence of moves:

The exchange usually starts with a query from the dominant interactant (e.g., shopkeeper in a service encounter), I shall call 'A'. Topics generally progress from deictic (referring to something in the shared context of the interlocutors) to general (something that all interlocutors should be able to discuss, such as the weather) to personal, and the query is phrased so that it is easy for the listener ('B') to agree if the query is a statement of opinion. Schneider describes nine subtypes of first moves.
The second move consists of B responding to the query. One of the principles of small talk is a modification of Grice's maxim of Quantity in that one should always respond with more information than was explicitly asked for. For example, responding to "Isn't it a wonderful day?" with "Yes." would often be seen as a flouting of this maxim within a small talk frame.
The third move consists of A responding to B's information. Schneider describes five types of third moves: echo question ("Where are you off to?", "Glasgow.", "Glasgow?"); check-back ("Oh, Glasgow, huh?"); acknowledgement ("I see."); confirming an expected response ("Glasgow."); and positive evaluations ("That's great.").
The remainder of the moves within a small talk segment are what Schneider refers to as "tail moves", which consist of zero to eleven moves of "idling behavior", such as "Mhm", "Yes.", "Really.", "Yea", etc. This apparently serves Jakobsen's phatic function while the interlocutors search for a new topic to discuss.

A more recent analysis of small talk was performed by Coupland and Coupland (1992) in which they provided test subjects with an initial query that was ambiguous as to whether it was a prompt for small talk or task-oriented talk ("How are you?" in the context of a health interview). They found that interviewees responded with a range of responses from the purely phatic ("Fine, thank you.") to the purely task-oriented ("I have asthma."). The responses in-between these extremes used various forms of hedging (face threatening act mitigation strategies as in Brown and Levinson (1987)) such as relativized appraisals ("Good for my age."). The Couplands' conclusion was that small talk is not a discrete frame of interaction (as characterized by Schneider) but rather a style of interaction which fulfills the function of "phaticity," and that all talk has varying levels of phaticity based on the current needs of the interlocutors to manage their interpersonal distance.

That casual conversation should be used to build rapport is also well documented in the popular press (e.g., RoAne, 1997), and literature in sociobiology (e.g., Dunbar, 1996) and social psychology (e.g., Laurenceau JP, Barrett LF, Pietromonaco, 1998). Equilibrium models of intimacy in social psychology are particularly interesting in that they account for the fact that individuals seek to achieve a fixed level of intimacy in a given interaction, and if this equilibrium is disturbed (e.g., by topic change in the conversation) they will seek to restore the overall balance of intimacy through other channels (e.g., eye contact, smiling, and physical proximity).

2.2. Chatterbot Systems

Chatterbot systems are capable of carrying on text-based "casual" conversations with users. Most of these systems are direct descendants of ELIZA (Weizenbaum, 1966) and primarily use regular expression pattern matching against the user's input to directly index canned responses. The success of these systems depends upon having a large enough set of responses that the system never repeats itself, and crafting the responses in such a way as to be ambiguous enough that the user can draw their own relevance relations among them.

More recent chatterbot systems use various indices in addition to pattern matching to determine the system's response. For example, the system described in (Mauldin, 1994) used a current topic index which is updated based on an activation network. Rousseau and Hayes-Roth (1997), describe a system in which response indices include the character's affective state.

Chatterbot systems also utilize several "tricks" to give the illusion of intelligence and fluency (Mauldin, 1994), including:

Maintaining initiative by asking questions.
Including portions of the user's input in the system's response.
Changing the level of the conversation (e.g., "Why do you ask that?").
Rigidly continuing a topic.
Launching into a new topic.
Making controversial or humorous statements.

3. Approach

Within the real estate sales domain, small talk is embedded within the larger task-oriented context of house-selling. That is, it is not a random phenomenon, but one which occurs at specific times for identifiable reasons, primarily the management of interpersonal distance between the agent and the customer. In order to perform task-oriented dialog, a discourse planner must be used which can decompose and sequence high-level goals (such as "sell house") into communicative speech acts (such as "ask the customer if the house is large enough") which can be executed.

To solve the problem of when the agent should transition into and out of small talk, I introduce a scalar measure of interpersonal closeness (IPC), which starts at zero (representing complete strangers) and increases with the perceived closeness between the agent and the customer. Certain communicative speech acts (such as asking about personal information) can have a minimum required IPC specified as part of their felicity conditions, and thus can only be executed when the customer and agent have achieved some level of familiarity with each other.

I also introduce small talk as a discrete frame of interaction (in addition to task, greet, and farewell frames). Within the small talk frame, IPC is gradually increased the longer the interlocutors engage in small talk. Task-oriented request speech acts can only be executed within the task frame. Given this framework, a dialog planner can plan to engage in small talk in order to reach its task-oriented goals, and can interleave task talk and small talk as needed to reach its goals as efficiently as possible.

The question of how dialog should be organized within the small talk frame is provided by Schneider's characterization of small talk segments and moves. The agent normally issues the initial query in each segment, allows the user to respond, provides evaluative feedback, and the performs a random amount of idling behavior.

Mixed initiative interaction is provided for by allowing the customer to ask the initiating query within the small talk frame, and transitioning immediately to the task frame if they introduce a task-oriented topic.

4. Implementation

An computational model of the above approach was implemented in CLIPS to function as a text-based conversational real estate agent. A simple linear dialog planner was implemented which decomposes the goal of selling a house into speech act subgoals for obtaining information and commitment from the customer. Each of these speech acts has a specified IPC requirement before it can be executed.

Interaction begins in the greet frame then automatically transitions into the small talk frame (per ritual convention). Interaction proceeds in the small talk frame until IPC has been increased to the point at which task-oriented speech acts can be executed or the customer mentions a task topic, at which time the task frame is started. Within the task frame, task-oriented speech acts are executed until either the agent's goals are satisfied, causing a transition to the farewell frame, or no more speech acts can be performed at the current IPC level, causing a transition back into the small talk frame.

Within the small talk frame, the dialog is conducted with segments and moves modeled according to Schneider. The agent normally starts a new segment and performs the odd-numbered moves, but the user is also free to do so, and the agent will perform the even-numbered moves. As each segment is concluded IPC is incremented. ELIZA-style pattern-response pairs are used to generate evaluative (third move) and query (second move) responses.

5. Results

Attachment 1 presents a complete run of the system, with the agent taking the initiative in all cases. Conversation moves from the greet frame to cycles of small talk and task frames until the agent's goals are satisified, and then the farewell frame is activated.

An example of the user taking initiative in the small talk frame is shown in Attachment 2, and an example of the user causing an immediate switch into the task frame through introduction of a task topic is shown in Attachment 3.

6. Conclusion

Small talk is an important mechanism for managing both the channel of communication and interpersonal distance between a user and an interface agent, and are thus an important part of an agent's social intelligence, especially in certain types of applications such as automated sales agents. The computational model developed in this project demonstrates that small talk can be used as part of the deliberately planned communicative behavior of a conversational agent to help it achieve its task-oriented goals. Small talk also provides an opportunity for such agents to build the illusion of intelligence through unconstrained mixed-initiative conversation on a wide range of topics (using chatterbot techniques) before transitioning into narrow task-oriented dialog.

Integrating small talk functionality into an embodied conversational character provides numerous challenges and a fertile area of research. Some of the more difficult and interesting problems include:

Characterization of real-estate small talk. It is possible that the small talk which occurs in real-estate encounters is qualitatively different from the small talk which occurred in the scenarios that Schneider characterized. A detailed study of this phenomenon should be conducted on naturally-occurring data.
Better evaluation of interpersonal closeness. The current heuristics used to dynamically assess IPC are very crude and unprincipled. More refined and psychologically-plausible methods of estimation should be used. In applications involving the use of embodied interface agents, proxemics may also provide a valuable information source for this.
Mode of small talk. Once casual conversation gets beyond a particular level of closeness, it expands to include many additional types of moves including conversational storytelling and joke telling. Although much work has been done on how these moves are integrated into an ongoing conversation (e.g., Jefferson (1978)), not much has been done on the criteria an agent uses to select among different available modes to achieve their communicative goals.
Phaticity. Finally, the result of the Coupland and Coupland (1992) paper cannot be ignored. Small talk does not exist in discrete frames only; all talk can be characterized by varying levels of phaticity. Techniques for recognizing the degree of phaticity underlying the customer's utterances and generating appropriate linguistic devices to effect the agent's phatic goals are significant research problems.

References

Brown, P. and Levinson, S. (1987). Politeness: Some universals in language usage. Cambridge University Press.

Cassell, J., Bickmore, T., Billinghurst, M., Campbell, L., Chang, K., Vilhjálmsson, H., Yan, H. (1999). "Embodiment in Conversational Interfaces: Rea", to appear in CHI99, Pittsburgh, PA.

Coupland, J. and Coupland, N. (1992). "How are you?: Negotiating phatic communion." Language in Society 21, pp. 207-230.

Dunbar, R. (1996). Grooming, Gossip, and the Evolution of Language. Harvard University Press.

Garner, A. (1980). Conversationally Speaking. Lowell House, Los Angeles.

Goffman, E. (1983). Forms of Talk. University of Pennsylvania Press.

Jakobson, R. (1960). "Linguistics and poetics." In Sebeok 1960. Pp. 350-77.

Jefferson, G. (1978). "Sequential Aspects of Storytelling in Conversation". Studies in the Organization of Conversational Interaction, pp. 219-248. Academic Press.

Laurenceau J., Barrett L., and Pietromonaco P. (1998). "Intimacy as an interpersonal process: The importance of self-disclosure, partner disclosure, and perceived partner responsiveness in interpersonal exchanges." Journal of Personality and Social Psychology, 74: (5) 1238-1251.

Malinowski, B. (1923). "The problem of meaning in primitive languages." Supplement to C. Ogden and I. Richards The meaning of meaning. London: Routledge and Kegan Paul. Pp. 146-152.

Mauldin, M. (1994). "Chatterbots, Tinymuds, and the Turing Test: Entering the Loebner Prize Competition. Proceedings of AAAI 94.

Polanyi, L. (1985). Telling the American Story: A Structural and Cultural Analysis of Conversational Storytelling. Ablex Publishers, Norwood, New Jersey.

Reeves, B. and Nass, C. (1996). The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places. Cambridge University Press.

RoAne, S. (1997). What do I Say Next? Talking your way to business and social success. Warner Books, New York.

Schneider, K. (1988). Small Talk: Analysing Phatic Discourse. PhD Thesis, Philipps-Universitat Marburg, West Germany.

Tannen, D. (1984). Conversational Style: Analyzing talk among friends. Ablex Publishing, Norwood, New Jersey.

Turkle, S. (1995). Life on the Screen: Identity in the Age of the Internet. Touchstone, New York.

Weizenbaum, J. (1966). Eliza -a computer program for the study of natural language communication between man and machine. Communications of the ACM, 9, 26-45.

Sample Run of Computational Small Talk System

--------------- initiate Frame ---------------

R: Hello.

U: Hi there.

--------------- smalltalk Frame ---------------

R: Isn't this a great neighborhood?

U: I think it's wonderful.

R: you think it s wonderful?

U: yes.

R: Isn't this living room fabulous?

U: I think it's pretty.

R: you think it s pretty?

U: yep.

R: yep?

--------------- task Frame ---------------

R: Is this home in the right location for you?

U: Yes.

R: Do you like this style of home?

U: Yes, it's nice.

R: Is this home big enough for you?

U: Yes, I think so.

--------------- smalltalk Frame ---------------

R: Where are you living now?

U: Cambridge.

R: cambridge?

U: Not far.

R: How big of a family do you have?

U: Just myself.

R: just yourself?

U: yes.

R: Do you have little ones on the way?

U: nope.

R: nope?

U: no.

--------------- task Frame ---------------

R: Do you have the down payment?

U: yes

R: Do you have any income?

U: yes, sufficient.

--------------- smalltalk Frame ---------------

R: Boston is great place to live, don't you think?

U: It's ok.

R: it s ok?

U: Yea.

R: How about this weather?

U: Wonderful.

R: wonderful?

U: Yes.

R: yes?

--------------- task Frame ---------------

R: Do you want to buy this home?

U: Yes.

** Game over. Rea won. **

--------------- farewell Frame ---------------

R: Goodbye.

Example of User Initiative in Small Talk Frame

--------------- smalltalk Frame ---------------

R: How about this weather?

U: It's gorgeous.

R: it s gorgeous?

U: Don't you think it's a wonderful Spring?

R: Maybe.

U: Yea?

R: yea.

U: I just love the flowers, don't you?

R: Yes.

U: Yea?

R: yea.

U: : Hmm.

R: Boston is great place to live, don't you think?

U: Yes, it is.

R: yes it is?

U: Yea.

R: yea?

R: How big of a family do you have?

U: Just me.

R: just you?

U: yes.

--------------- task Frame ---------------

Example of User Switching from Small Talk to Task Talk

--------------- initiate Frame ---------------

R: Hello.

U: hi

--------------- smalltalk Frame ---------------

R: Isn't this a great neighborhood?

U: Yes, it's great.

R: yes it s great?

U: I'm looking for a house with a pool.

--------------- task Frame ---------------

…