This paper is about explaining politeness in terms of speech-act-like conversational moves, particularly in the kinds of task-based dialogue that we find in virtually all human-computer interaction. I present examples from various dialogues in which conversational behaviors manifested as politeness-like actions arise out of the specific conversational situation and are apparently intended to achieve goals not directly related to the conversants’ nominal task. The paper does not seek to explain why the politness is expressed the way it is, that is, how the locution embodies the illocution.
Brown and Levinson (1987) produced the deepest and most comprehensive account of perspective of politeness in the conversation-analytic literature. What they sought were linguistic universals that would explain politeness phenomena across languages, cultures and domains. Such universals, which presume Grice’s theory of conversational implicature, would have enormous value. For example, as Brown and Levinson point out, much of the difference between the nominal meaning and the implications of an expression can be attributed to politeness. This analysis is, I think, correct. But it is not of great help to engineers of conversational systems. The politeness model helps to explain some fairly subtle language phenomena while the problems of producing first-order language remains difficult for dialogue systems.
Brown and Levinson followed Goffman’s notion that politeness behaviors in conversation are motivated by consideration of positive and negative face. This motivation, they suggest, results in positive and negative politeness that is manifested through a set of conversation strategies, such as “Use in-group identity markers.” While I agree with Brown and Levinson with respect to the strategies they enumerate, in this paper I will argue that the instances of the strategies are produced not by general conversational or social forces such as face but rather by much more specific goals that result in the carrying out of activities such as group maintenance even in the context of highly specified domain tasks. In other words, many politeness behaviors can be explained without resort to linguistic universals by looking at the pragmatics of the contextualized conversation acts.
The consequences of this analysis are that dialogue systems (a) may be more likely to be able to interpret and produce the richer set of conversational acts that characterize human-human conversation but (b) are likely to have to build more sophisticated task models that account for more than nominal domain tasks. This will particularly be the case for multi-party interaction involving a mix of human and computational agents.
2. The Implicit-Task Model
In a previous paper (Novick, 1996) I noted that observation of real, task-oriented conversations indicates that utterances, exchanges and whole conversations typically include actions that realize a mix of goals. In the case of conversations in business meetings, for example, we know that the “task” entails a complex of goals and activities. While the nominal task may involve things like setting sales policies, episodes of activities related to the theme typically are interspersed with other kinds of supporting activities, including group formation or maintenance. In other words, there is an explicit task (the nominal, domain task) and a set of additional, implicit tasks. This means that, when looking at transcripts of even heavily “task-oriented” conversations, it is normal to find conversation acts that appear to be off-topic. But these extra acts are only off-topic in the sense that they don’t directly apply to some externally assigned explicit task. The conversants’ implicit tasks are just as real a part of their interaction; we usually have to conjecture, though, about the goals underlying the implicit tasks. For example, in a business meeting called to address a particular topic, the participants may supplement the topic with other, opportunistic interaction that takes advantage of the meeting’s time, place and participants (Wynn & Novick, 1995).
So here I want to argue that the actions that we perceive as politeness are the locutionary representation of illocutionary acts that have as their goal the establishment or maintenance of the relationship between the conversants. This is based on Brown and Levinson. but seeks to go beyond notions of “face” as a linguistic or cultural universal and instead seeks the reasons for the actions of the conversants in the specific contexts of their conversations. If this is true, then it should be possible to show that conversants have specific, situated intentions that are realized either as entire acts or as a kind of meta-act associated with the production of an act for the nominal task.
I will explore this claim through the presentation of examples from a variety of task-based corpora, including the HCRC map-task corpus, air-traffic control dialogues, and the OGI letter-sequence dialogues. In each case, politeness behaviors can be attributed to specific, contextually motivated goals of the conversants.
3. Map-Task Example
Figure 1 presents an example taken from the HCRC map-task corpus (HCRC, undated). In the map task, both conversants have maps of an imaginary area. One conversant, called the giver, is supposed to direct the other conversant, called the follower, along a path with reference to landmarks present on their maps. The maps may have missing or inconsistent landmarks, so the task often involves resolution of apparent conflicts.The transcript identifies the conversant (“Giver” or “Follower”), the utterance number, the HCRC’s coding of the speech act, and the utterance itself. I have followed the HCRC’s segmentation, but however I’ve left out the overlap notation.
One of the interesting things about the map corpus is that the conversants knew each other already. That is, they had an existing social relationship outside the experiment. So in addition to the nominal map task they also brought other goals to the conversation, such as maintenance of this relationship. This maintenance task, implicit but present, manifests itself in utterance M196 as “It’s my fault.”
In the dialogue leading up to this utterance, the conversants are experiencing a mild conflict: the location of the yacht club doesn’t seem to match between the conversants’ maps. As a result G tries to defuse the conflict by taking the blame for the apparent conflict. This is an instance of Brown and Levinson’s Negative Strategy 6, "Apologize,” except that here G isn’t necessarily trying to mitigate impinging on F’s negative face. Rather, G has directly both short- and long-term interests in a cooperative solution to the apparent conflict.
The HCRC coded this utterance as unclassifiable because it didnot fit into their coding scheme, which was based on acts associated with the nominal task. Here the implicit task intrudes. There follows a brief subdialogue on this theme, in which both conversants participate, apparently to reassure each other that each is acting in good faith and that the problem, whatever it may be, lies outside their relationship.
4. ATC Examples
My next two examples are taken from the OGI air-traffic control (ATC) corpus (Ward, Novick & Sousa, 1990). The corpus consists of a transcript of 30 minutes of radio communications between an air-traffic controller and pilots. In this case, the air-traffic controller was handling what is called the approach, which mainly involves transitioning aircraft from high-altitude cruise to low-altitude set-up for landing at an airport. In the transcript, the approach controller interacted with 18 aircraft, for a total of 275 utterances. The conversations between the controller and the various aircraft are interleaved. One of the aircraft was a military flight, and only the controller’s side of the conversation was available; for this reason, this flight has been excluded from this analysis. Thus I analyzed 17 (interleaved) dialogues with a total of 263 utterances.
Typical utterances involve routine clearances. For example, utterance 243 is from Approach to American 845 (punctuation added for clarity):
American eight forty five you’re one one miles from Laker. Maintain four thousand five hundred till established on the localizer. Cleared ILS two-eight right approach.
In this example, American 845 is a commercial flight, Laker is the name of a navigation point on the approach path to Portland International Airport, 4500 is an altitude, the localizer is a navigation radio beam used for instrument landings, and being cleared for the ILS 28R approach means that the aircraft is approved to fly the path that leads to landing on runway 28R. A principal reason for the formulaic nature of the interaction is that the language to be used in ATC is prescribed by federal regulation. Consequently, when the conversants infrequently depart from the prescribed forms, such utterances hold great interest. In this corpus, of the 17 full or partial conversations analyzed, five contained examples of politeness, almost always expressed through the words “thank you.” One additional conversation contained a routine use of “good-bye” that was not analyzed further.
Figure 2 shows an example of a typical use of “thank you” in these dialogues. In this example, a private aircraft, Bonanza 64A, requests a lower altitude, apparently due to the presence of other aircraft that are flying toward a standard approach to a runway at Portland International Airport. Although directly safety-related, this request seems to be non-routine. So the the pilot of Bonanza 64A is a situation where he is making a special request to the controller, and it’s possible that he may have to make other requests later. As a result, even though the pilot and controller are not maintaining a pre-existing relationship, the pilot is careful to acknowledge the controller’s help.
Conversely, in another exchange, the controller asks a pilot of a private plane for help in understanding weather conditions. This request adds to the pilot’s workload, so the request is non-trivial. The controller depends on the cooperation of pilots both for responding to control clearances and for providing information about conditions, so he has a distinct long-term interest in maintaining good relations with pilots, particularly with respect carrying out these kinds of supplemental conversations that are not strictly part of the prescribed task.
Overall, the conversants in the dialogues‑whether pilot or controller‑almost always use explicit politeness after receiving the results of a request. Moreover, these are the only forms of politeness shown in the in corpus, aside from end-of-dialogue saluations. The corpus contains six requests. Of these, five were later concluded with “thank you.” In a sense, the requesting party incurs a debt, and the thank-you concluding the exchange represents an instance of Brown and Levinson’s Negative Strategy 10, “Go on record as incurring a debt, or as not indebting H.” In the sixth case, the request involve lifting a speed restriction that had been imposed by the controller. While the pilot does not conclude the exchange with a “thank you,” he does make the request using hedged language: “Can we pick it up, (just a) little bit?”. This can be seen as an example of Brown and Levinson’s Negative Strategy 2, “Question, hedge.”
This corpus gives us two pieces of evidence supporting the view that politeness behaviors are determined by the task context rather than set of linguistic universals. The first piece of evidence is that these fairly similar domain-task requests were supported with different politeness behaviors. The second piece of evidence involves comparing domain-task requests with other requests in the ATC corpus. As I noted above, five of six domain-level requests were followed with thank-yous and the sixth involved mitigating language. But the corpus also includes five other, non-domain-level requests, namely requests for clarification. The ATC channel is low-bandwidth and noisy, and there’s usally a lot of noise in the cockpit, too. So the conversants sometimes have to ask for clarificationof what was just said. This corpus contained four cases of requests for clarifications, and none of these was followed with a thank-you or other politeness behavior. Note that the relationships between controller and pilot were the same, so the use of politeness had to spring from the particular context created by the domain request rather than general rules of politeness.
5. Letter-Sequence Example
My concluding example is from the OGI letter-sequence corpus (Novick, Hansen & Lander, 1994). The corpus consiss of a set of task-oriented laboratory dialogues. The experimenters gave subjects the assignment of collaborating face-to-face in the letter sequence task, a task designed to mimic the structure of more complex collaborative tasks in which participants have incomplete knowledge. Each pair of subjects was given cards containing a sequence of 16 letters and blanks such that they could reconstruct the entire sequence by pooling their knowledge. They were asked to put the cards out of sight and to work together to reconstruct the complete sequence from memory. Figure 3 shows a sample of two subjects’ letter sequences.
The corpus contains two dialogues each from four pairs of subjects. The dialogues contain from 93 to 236 utterances each. Of the total of 1240 utterance, three contain the word “sorry.” One of these was after a cough. The other two relate to problems that the conversants are having with the domain tasks.
Figure 4 presents one of these episodes, in which the conversants are reaching the end of a pass through the sequence. (I note that utterances 47 and 49 appear to be missing. In fact, they correspond to non-verbal behaviors that are not relevant to this analsyis.) L says Z, which she believes is the last letter, and sits expressionless as R then continues with more letters, saying “okay, L E B Y.” R is unaware that they have misunderstood each other. L initiates a repair because she thought they had reached the end of the sequence and R had then added additional letters. So L says, “Wait,” laughing. “I’m confused now.” When R notices L’s hesitation, she asks “no?” basically asking if that’s not right, either. Then she apologizes, looking at L for explanation. Apparently R was fairly confident of her “L E B Y” part and is a little confused about what she has done wrong.
The evidence suggess that it is this disjuncture between R’s original view of the state of the dialogue and her updated view based on L’s initiation of a repair that leads her to apologize. As in the example from the map-task corpus, this is an instance of Brown and Levinson’s Negative Strategy 6, “Apologize.” The conversation act in U57 is not part of the domain task. Rather, it addresses some other, implicit goal presumably dealing with the relationship between the subjects. The subjects still haven’t completed the domain task (as is clear from the fact that they both recognize in this exchange that there’s been a problem), and so they’ll have to continue to cooperate at least for a little while longer.
In each these examples, I have attempted to show how in each case the conversants’ politeness behaviors can be related to non-nominal-task goals that were appropriate to and thus were realized in the situations through Brown and Levinson’s politeness strategies. The examples all reflected what Brown and Levinson called negative politeness. I note, though, that the map-task and letter-sequence domains both include many examples of positive politeness. I expect that these examples could be similarly linked to specific dialogue contexts and implicit task.
These goals underlying the implicit tasks in the examples appear to be reasonable common, so they might be relatively easy to represent in spokent-language systems.It is not immediately evident that the human participant in routine human-computer dialogue would expect their machine partner to share this goal, though. Nevertheless, as humans gain experience with increasingly sophisticated dialogue systems, they may look forward to conversing with agents that take account of the context beyond that of the nominal task. For example, pilots might find an automated ATC system to be cloying if it routinely used explicit forms of politeness for all exchanges, but might find the same system to be more usable if it employed and could understand politeness as a follow-up to domain requests.
Finally, the kind of politeness I’ve discussed in this paper is not particularly sophisticated. The examples all display, as near as I can tell, a politeness that is sincere. The implicit-task model can start to account for this form but is not capable of addressing more complex forms such as the use of politeness to express irony or sarcasm. Moreover, the analysis addresses only the underlying acts and not their forms of expression.
Brown, P., and Levinson, S. (1987). Politeness. Cambridge: Cambridge University Press.
Human Communications Research Center, University of Edinburgh, “Welcome to the HCRC Corpus Interface Demo,” http://www.ltg.ed.ac.uk/~amyi/maptask/demo_actions.html.
Novick, D. (1997). Simple actions, complex acts, Working notes, AAAI 97 Symposium on Communicative Action in Humans and Machines, Massachusetts Institute of Technology, November 8-10, 1997.
Novick, D., Hansen, B. and Lander, T. (1994). Letter sequence dialogues, Technical Report No. CS/E 94-007, Oregon Graduate Institute of Science and Technology.
Ward, K., Novick, D., and Sousa, C. (1990). Air traffic control communication at Portland International Airport, Technical Report CS/E 90-025, Oregon Graduate Institute of Science and Technology.
This paper was published as Novick, D. (2000). Politeness as actions of an implicit task. Proceedings of the Third International Workshop on Human-Computer Conversation, Bellagio, Italy, July, 2000, 124-129.