L. Ardissono, G. Boella and L. Lesmo
Dipartimento di Informatica - Università di Torino
Corso Svizzera n.185 - 10149 Torino - Italy
Fax: +39 - 011 - 751603; Phone: +39 - 011 - 6706711
e-mail: {liliana,guido,lesmo}@di.unito.it
In this paper, we propose a logical description of the mechanisms which cause a speech act to be impolite, and of how the indirect expressions may prevent speakers from offending their partners. We specifically focus on conventional indirect speech acts, providing a formal framework to recognize the beliefs underlying them and the way how the possible offenses produced by communicative actions may be blocked by using politeness techniques.
1. Introduction
It is widely accepted that
utterances are actions performed by agents
to change their partners' beliefs and intentions
[4].
However, illocutionary acts must also satisfy social goals, related
with politeness, and this second aspect has not yet received a satisfactory
treatment. Indirect speech acts are the paradigmatic example:
speakers often use them when the direct forms may appear impolite.
Indirect speech acts have been explained in terms of rational behavior
([16]; [7]) and related to
the felicity conditions of the intended speech
acts. More recently, the conventionality of these phenomena has been addressed
in [11]. Furthermore, the impact of an agent's character
on the selection of speech acts has been studied in the AI literature
about social behavior ([15]). Finally, the variety
of ways to perform a speech act has been exploited in [19]
for interactive entertainment, and in HCI to adapt a system's behavior to the
user's needs ([18]).
Brown and Levinson [5] proposed an informal
explanation of the politeness strategies
adopted in different natural languages
in terms of social goals and means-ends reasoning.
However, there is still no formal explanation of how indirect
expressions contribute to politeness, and which parameters influence the
different dialog situations.
In this paper,
we propose a logical description of the mechanisms which
cause a speech act to be impolite, and of how conventional indirect
speech acts may prevent speakers from offending their partners.
We have developed our work on the Italian language: this implies a
particular point of view, especially so far as the linguistic politeness
forms are concerned, but also with respect to
the underlying beliefs and reasoning; in fact,
as it is widely recognized, politeness phenomena are culturally specific.
In the paper, we report in Italian only the examples
idiosyncratic to our language.
2. The notion of face
Brown and Levinson [5] base their theory on two
assumptions that explain why
humans often restrain from a ``highly rational maximally efficient"
communication mode:
The conventional indirect speech acts (on record acts in [5]) should be distinguished from other more indirect forms (off record acts), like the presequences (e.g. Have you got a pen?) for two main reasons; first, these indirect speech acts have achieved a conventional character (however, as [5] observe, there ``are rational bases for conventions"). Second, they explicitly refer to the action the speaker wants the partner to execute.
3. Representation of speech acts
Following Brown and Levinson's ideas,
in [3], we introduced a Speech Act library which
describes direct and indirect ways to perform illocutionary acts.
We describe the speech acts in a recursive hierarchy of actions, related to
one another by decomposition and generalization links.
The illocutionary acts specialize into different forms, including
indirect ones. Figure 1 shows a portion of the Speech Act
library and contains some direct and indirect requests: Direct-request
is generated by a surface speech act in imperative mood and is appropriate
if the speaker does not bother about the hearer's face
(Give me a pen); instead,
Can-indirect-req and Int-indirect-req are executed by questioning the
hearer about his capability (or willingness) to do the action (Could you
open the door?): this is represented by the occurrence of Ask-if in its
decomposition.
Ask-if is also used to ask real questions: in fact, it occurs in the
decomposition of Obtain-info (not shown in the figure;
e.g. What time is it?).
Since questions
themselves have direct and indirect decompositions,
more and more indirect forms can be produced recursively:
interrogative forms are also recursively performed by means of Ask-if.
So, progressively more polite speech acts can be obtained, e.g. a request
can be
even performed by means of surface statements as
I would like to ask you if you could lend me some money.
Figure 2 shows the Request speech act, concerning an action act of the hearer, which is referred to in the superficial form of the utterance. Request has the following effect:
Sh(h, s, Cint(s, h, Int(s, Int(h, do(h, act)))) (1)The intuitive meaning of this effect is that both the speaker and the hearer share the knowledge that the speaker has the communicative intention that the speaker intends that the hearer intends to perform act.
4. Where do the threats come from?
After the production of an utterance, the interlocutors mutually know that
the speaker has the intentions described in the effects of
the performed speech act (provided that the speech act succeeded).
However, the utterance also carries other information: first,
the execution of any speech act shows that the speaker believes that its
preconditions hold (i.e. a speech act presupposes its felicity conditions, see
[16]);
second, it shows the speaker's intention that the hearer bothers to interpret
his words, since the success of an illocutionary act needs the receiver's
uptake [4].
As stated before, we
suppose that every agent has a face, defined according to
Brown and Levinson's description.
The expression of these beliefs and intentions may threaten some aspects of
the hearer's face, like his freedom to perform actions
without imposition, his beliefs, or his self-image.
The expression of an intention
that an agent does something is not by itself offensive. However, when an
utterance is addressed to an agent, he performs other inferences: in fact,
speakers know that agents need a reason to act (a reason to adopt the
speakers' intentions). In particular, given the speaker s's utterance,
the hearer h must discover the reason why s believes that
h will act.
The fact that there must be such a reason
can be obtained by means of the following rule:
Sh(h, s, Cint(s, h, Int(s, Int(h, do(h, act)))))
Bel(h, Bel(s, reason-to-act(h, s, act)))
But there are different reasons to adopt the goal of performing an action, even if the agent is not the beneficiary of the action (in the following, Bel-do-for(h, s, act) Bel(h, Int(s, Int(h, do(h,act))))). Among them, we consider the following reasons to act:Bel-do-for(h,s,act) goal(h, indebted(s,h))
[Bel-do-for(h,s,act)
beneficiary(h,do(h,act)) do(h,act)] indebted(s,h)
Other reasons to act may be introduced, but we will focus on reason (c). We suppose that all the reasons to act are mutually exclusive; therefore, if one of them is shared in the dialog context, the others are false.
In order to fully understand the beliefs underlying a speech act, the
receiver of an utterance
has to identify which of the above listed reasons to act is assumed
by the speaker:
in some cases, the interaction
context provides the hearer with such information (e.g. if the interactants are
involved in a shared plan, they are cooperating and assumption (a) holds
in the context); in other cases, the speaker makes explicit the reason why the
hearer should act; finally, there
are cases where the reasons to act must be inferred by the hearer himself.
We introduce a general rule that a hearer triggers when he interprets
a new illocutionary act produced by a speaker s:
Bel(s, act' (Bel-do-for(h, s, act') Int(h, do(h, act'))))
I.e. if the speaker s believes that the hearer h has a reason to act,
then by default s believes that, if h believes that s
intends him to act, h intends to act (the ``'' notation
denotes default inference).
We chose a default rule because the hearer's interpretation process
has a defeasible character [1].
So, if the context supports another explicit reason for h to
act, or s's utterance provides any clue about this,
then h will not conclude that the speaker relies on h's
dependency on s's willingness.
The belief in the conclusion of the default rule (D)
could threat the hearer's freedom to act without imposition: the
imposition occurs when the intended action is burdensome to
the hearer who has to perform it, and it is not in his favor.
In this case, the hearer will be offended:
Bel(h,{Bel(s, act (Bel-do-for(h,s,act) Int(h, do(h, act))))
act'[Bel-do-for(h,s,act') Bel(h,beneficiary(do(h,act'),h) burdensome(h,act'))]})
offended(h,s) (2)
The offended(h,s) predicate means that agent h has been offended by s because his face has been violated by s.
Certainly, a speech act may convey information that could be offensive; on the other hand, the same means can be exploited to convey other information that inhibits any undesired default conclusion on the hearer's part. For this reason, the conventional indirect speech acts are realized by means of other illocutionary acts: in this way, the threats carried by their effects can be challenged by other information, inferred from the instrumental (surface) speech act. In the following, we will consider some illocutionary acts and discuss how the potential threat they carry can be prevented by expressing them in indirect forms. Our analysis is not intended to be exhaustive, but it aims at reducing to the same unifying principles several indirect forms used to express politeness.
4.1 Requests
Requests are the clearest example of face threatening speech acts: in fact,
they explicitly express the speaker's intention that his partner performs an
action, and this fact triggers the default rule (D); moreover, the requested
actions have usually to be performed for the sake of the speaker; so,
requests clearly satisfy the antecedent of the implication (2).
One way to make the conclusion of the rule (D)
inconsistent is to explicitly provide an
alternative, not offensive reason to act; another way is to allow the
inference that the speaker is not certain that the hearer intends to
perform the requested action act1.
In fact, h would believe
Bel(s, Bel-do-for(h, s, act1) Int(h, do(h, act1)))
if he believes that the antecedent of the nested implication is true, while its
consequent is not believed (i.e. Bel(s, Bel-do-for(h, s, act1)) and
Bel(s, Int(h, do(h, act1)))).
However, this antecedent is certainly believed, because it
derives from the illocutionary effect of the request:
So, s has to show that he does not hold the belief that the consequent of the implication is true, by means of another illocutionary act that conveys this information. Note that, since h has to interpret s's utterance by exploiting the plan library in figure 1, in the intention recognition process h identifies that s has performed a request (although in an indirect way). So, h understands that s has the goal that h performs act1. However, the fact that Bel(s, Int(h, do(h, act1))) does not conflict with the goal of s that h does act1. In fact, as [14] has highlighted, plans do not necessarily support the expectation of their success. In order for s to act, it is sufficient that he does not believe his action will certainly fail (i.e. Bel(s, Int(h, do(h, act1)))). See [8] for a discussion on the reasons for choicing this solution.
In the following, we will examine some alternative ways of performing indirect requests:
In particular, s may perform a Request and show that he does not believe the consequent of the implication by exploiting a question about h's willingness to do act; e.g.:
Hai voglia di darmi un passaggio? [Do you want to give me a lift?]
The precondition of the question is: Bel-if(s, Int(h, do(h, act))). From s's lack of knowledge, h can infer that s does not believe neither that h intends, nor that he does not intend to do act.
Mi passi il sale? [Do you pass me the salt?]
Int(a, do(a, act)) cando(a, act) (3)
where cando represents the joint conditions that a has the capability of
doing act, believes that the applicability conditions of act hold,
and the intention to do act is reconciled with his other intentions.
When s asks h whether h can do an action, s displays that he does
not believe that h intends to do act; e.g. consider:
Puoi darmi una penna ? [Can you give me a pen?]
Bel(s, Int(h, do(h, act))) must be false: in fact, (3) is a mutually believed axiom and s does not believe that the consequent cando(h, act) is true (the precondition of the question is Bel-if(s,cando(h, act))).
Non puoi prestarmi la macchina, vero? [You can't lend me your car, can you?]
In fact, this type of question implies that the speaker believes that the propositional content of the question is false, and that he is just checking this belief; so it highlights that the Bel(s, cando(h, act)) formula is false, since Bel(s, cando(h, act)).
This approach also explains why questions that take a positive answer for granted have an impolite effect.
Non puoi darmi un passaggio? [Can't you give me a lift?]
Instead of hedging the possible offense, the speaker highlights some condition related with it.
Notice that, as described by axiom (2), the potentially offensive reason to act is harmless if the object of a request is in favor of the hearer. This happens for example in the case of offers:
Take my car!
In some contexts, the above imperative form may be more polite than some indirect request, like:
Could you possibly take my car?
In fact, if the speaker performs a polite speech act, he shows that he does not consider the referred action totally positive for the addressee (otherwise, he would not need to exploit any politeness means).
4.2 Questions
Also questions can bring speakers into trouble; this fact is proved by the
existence of indirect forms, like:
which are often used instead of direct questions.
On the other hand, it seems that speakers must balance the potential threats to the hearer's face with the fact that, in some cases, indirect questions are not well formed; this fact depends on the information asked for. Consider, for example:
Do you know where you have been yesterday evening? (5)This sentence can be only interpreted as the first component of a presequence [12], possibly followed by B's reply Where?.
From the politeness perspective, indirect questions are similar to
requests: in fact, a question has the effect that the
speaker and the hearer share the knowledge that the speaker intends that the
hearer answers the question (see Ask-if in Figure 1):
therefore, a question looks like a request that the hearer informs the speaker about an information p, which is the propositional content of the question itself.
An indirect question is usually more polite than its direct version, because the speaker prevents the application of the default conclusion (D), that would threat the hearer's freedom to answer. In particular:
Note that also the literal question of example (4) could be an imposition to h, but in general it is not; the reason is that it is a request to inform s about h's knowledge, and this action of informing is less impinging than the one specifically directed to proposition p. On the other hand, this constraint makes the indirect question unfeasible if it is shared between the interactants that the hearer knows the answer, as in (5).
There is another reason that makes these indirect question forms unfeasible: since the speaker expresses doubts about the hearer's knowledge, he risks to contradict his positive self-image as an agent competent on the topic. For this reason, if the asked information concerns a topic which the hearer believes to be considered competent on, the speaker must face a trade-off between the risk of imposition and that of challenging the hearer's self-image.
There is a further source of threats from questions, which we only mention: when questions concern ``private'' information, indirect forms are not likely to have any effects:Do you know your credit card number?
Please, could you tell me your credit card number?
In fact, they are intrinsically invasive; therefore, politeness techniques can not cancel their offensive power.
4.3 Statements
Statements suffer from two main problems: the imposition
on the hearer (who should update his own beliefs with those conveyed by
the statement) and the risk to annoy him, trying to make him believe
something he already knows, or he does not agree with.
Note that not all the surface statements have the same illocutionary effect: some are genuine statements, performed to make the hearer believe some information (e.g. the answers to questions); others just aim at focusing on some common ground, shared by the speakers; this second type is often used in argumentations. In order to explicitly distinguish the second statement type from the possibly offensive first one, the speaker can negate one of the constraints of an ``informative statement'', Bel(s, Bel-if(h,p)), which prevents speakers from communicating already known information; e.g.: ``As you know, humans are mortal, Socrates is a human, so ...''
However, if the speaker s intends to inform the hearer h about something
h might disagree with, s has to give up the statement and replan a weaker
one: in fact, h's ``positive face'' could be threatened if s imposed h
to update his own beliefs about the world in a way h does not agree with.
In order to avoid this imposition, s can weaken his statement, adding to it
an expression of doubt, by means of adverbs like ``maybe, probably'', and so
forth. In alternative, he can put the statement in a subjective perspective, by
adopting the goal that the h knows that s believes the conveyed
information:
I believe that Mary had gone home.
Such a statement makes mutually believed that the speaker s wants the hearer h to update his beliefs with information regarding s's beliefs:
Sh(h, s, Cint(s,h, Int(s, Int(h, do(h, UpdateBels(h, Bel(s, p)))))))The communication of a subjective belief does not threaten the hearer's face, because it only requires that h updates his beliefs regarding the speaker's beliefs, and these second-level beliefs are by nature uncertain; so, h's face cannot be threatened, even if he discovers that he previously had a wrong belief about s's beliefs.
5. Conclusions
In this paper, we have provided a rational model of
indirect speech acts in communication.
The aim of this work is to show how
linguistic phenomena that have achieved a conventional character
can be explained at a deeper level. We do not claim that
the kind of reasoning illustrated above is performed every time
a hearer interprets an utterance: in general inferences can be considered
as almost ``precompiled''.
Notice that this ``precompilation" step depends on the environment, i.e. the
culture and the conventions the agents live within. So, some forms (even
in case they are reasonable from a rational point of view) could not have
arisen in a particular linguistic community. Moreover, also the more basic
notions of face, offence, and so on, are in some cases culture-dependent,
so that the definition of our informal predicates (indebted, offended, etc.)
could vary from one culture to another. In general,
our analysis regards western cultures and in particular the
Italian one.
We have explained why indirect speech acts can be used to convey a meaning as a side effect of a literal speech act, beyond the main illocutionary effect. In particular, indirect speech acts exploit different surface forms to smooth the inferences conveyed by a speech act by carrying them together with other communicative information. This information has the goal to prevent the hearer's reasoning from reaching any offensive interpretation of the speech act.
In this paper, we do not analyze how the belief that the face has been threatened leads to changes in the emotional state of the hearer (e.g. offence). [13] provide a computational model of how the beliefs about the content and intent of utterances indirectly result in changes to goals, emotions and interpersonal relationships.
Notice that our framework builds on some subjective notions,
like the face, cooperativity, the burden of actions, and so forth.
These notions are used to decide
whether a speech act is polite or not, and they are relativized to the opinion
of the agent receiving the speech act.
We have limited our discussion to conventional indirect speech
acts since we claim that off record acts exploit a different strategy,
based on
more complex reasoning, like the ability to exploit ambiguity
of interpretation and the assumption that the partners are able to perform plan
recognition inferences.
In [2], we explain off record forms, like those
occurring in presequences (have you got a cigarette?), by means of the
problem solving activity that agents perform while they plan and execute
linguistic actions.