Speech Acts and Goal Directed Dialogue

(a lightly-edited version of some ancient CIS 530 lecture notes by Mark Steedman)


The idea of a speech-act is familiar. It is the use of an utterance by a speaker to make a hearer change their state of mind in some way, such as to make them bring to mind a referent (the speech act of reference) or to induce an expectation in them (say by the speech act of promising). It is an idea which lies on the rhetorical side of the dichotomy between logic and rhetoric. Speech act theory therefore has a specialised vocabulary of its own for talking about the way in which speech acts depend for their effect on the precise state of mind of the speaker and hearer (hereafter, S and H).

First, a distinction is drawn between Propositional Content and Illocutionary Force. The distinction is between the proposition that the utterance involves, such as that John sits down, and the kind of speech act that it involves, such as stating, questioning or ordering, which are signalled by the declarative, interrogative and imperative constructions "John sits down," "does John sit down?," "sit down, John!." But, because we are no longer dealing with logical meaning alone, but are considering the world-representations of speaker and hearer as well, the concept of illocutionary force alone is not enough. Consider what happens if S utters the sentence ``You are going to London on Tuesday?'' as a question, as one may in English, but H mistakes it for a statement, and indignantly disagrees. What was the speech-act -- a question, or a statement? Clearly in some sense it was both. From S's end it was intended to provoke the answer Yes or No from H -- it just didn't work. From H's end it feels like S intended to make H think something which he happens to know is false. For this reason, a distinction is drawn between illocutionary force -- the speech-act that S actually intends -- and illocutionary uptake -- the speech act that H thinks S intends. Even when illocutionary force and uptake are in line with each other, as when S utters the statement ``Your shoelace is undone'' with the IF or intention of making H believe that fAct, and H correctly takes up or recognises that speech act, H may still not believe S. That is, the speech act may not have the intended effect on H. We refer to this effect as the perlocutionary effect of the utterance -- the effect that actually results in H's mind.

Austin, (19XX), who invented the term Speech Act, distinguished a considerable variety of kinds of speech act. Since that time there have been several taxonomies, or proposals to express this considerable variety in terms of a smaller number of ``primitive'' speech acts or classes of speech act. (For example, Promising, Naming, and the like are all rather like statements or ``informatives'', whereas Ordering, Forbidding, Questioning and Requesting are more like orders, or ``directives''.) The details of these taxonomies (whether there are three primitive speech-acts, or nineteen, or one) matter less than the fact that all analyses of speech acts involve the recognition by H of an intention on the part of S.

Certain speech acts are explicitly marked using special words and constructions. These speech acts are ``conventional''. For example, it is a convention of the English language that the tag ``please'' marks requesting and that subject-auxilliary inversion (sometimes) marks questioning, just as it is a convention of the English language that the word ``frog'' identifies frogs, and of Dutch that the unrelated word ``kikker'' identifies the same. But (and this is the most interesting thing about speech acts) many, if not most, speech acts are not conventionally conveyed at all, but arise from what Grice calls ``conversational implicatures''. That is, to use Searle's (1975) term, they are ``indirect''.

Indirect Speech Acts and Conversational Implicatures

One can bring about the same effect as requesting someone to get off your foot using a statement such as ``You are standing on my foot''. One can also bring about the same effect as a request for the salt by using a question like ``Can you reach the salt?''. These are not ``conventionalised'' requests -- there seems to be in principle an unlimited number of such roundabout ways of getting people off your foot, or getting them to pass the salt. Searle wants to say that they are ``indirect'' requests -- that is, utterances that succeed in requesting via some other literal or conventional meaning. Grice (who actually wrote this paper some ten years before its publication) describes the utterances which are literally statements or questions as ``conversationally implicating'' or implying the requests. Both Searle and Grice imply that conversational implicatures do a great deal of work in conversation. Every utterance is surrounded by a cloud of conversational implicatures. They arise because both speaker and hearer have certain assumption, principally that both are cooperating in having a conversation. Grice enshrines this observation in a principle that can be paraphrased as follows:
  (This statement of the principle is modified somewhat to take account of certain converasational situations---such as courtroom testimony---in which such goals as staying out of jail may limit the degree of cooperation characteristic of a witnesses contributions. This is to circumvent certain criticisms of Grice raised by Sacks and by Levinson. Note that if we left out the word ``conversational'' the definition would cover pretty much any cooperative activity.)

Conversational implicatures are like a kind of invisible glue or scaffolding that sticks individual utterances together into coherent conversation -- cf Searle's 1975 analysis (p61-63) of the exchange:

-- where the assumption that B's remark is cooperative and therefore relevant is crucial to A's understanding (on the basis also of knowledge about the world, how long films and studying last) that B's utterance counts as an indirect refusal of A's proposal. See also Searle's analysis (p73-74) of the indirect request Can you pass the salt. A crucial ingredient in Searles analysis is the participants understanding of the relation of preconditions like ability to pass the salt (which is questioned in can you pass the salt) and intention to get the salt (stated in such indirect SAs asI want you to pass the salt). There is some technical terminology here that doesnt matter much -- the former sort of precondition are called ``preparatory conditions'', the latter ``sincerity conditions''.

For Grice, the Cooperative Principle subsumes four ``Maxims'' of conversation as corrolaries or consequences. He calls them Quantity, Quality, Relation and Manner. They amount to the maxims ``Give an appropriate amount of information'', ``Be truthful and do not mislead'', ``be relevant'', and ``be clear''. We might paraphrase them a little more formally as folows:

  1. Quantity: Adjust the amount of information in your contribution to the state of knowledge of the hearer.
  2. Quality: Do not utter falsehoods or half-truths.
  3. Relation: Make your contribution relevant to the goals of the con- versation.
  4. Manner: Make your contribution easy to understand given the state of knowledge of the hearer.
Again, the taxonomy is less interesting than the things you can do with it. (For example, Deirdre Wilson has pointed out that all of the maxims can be subsumed under Relation. Notice also that the maxims only really hold for conversations where no other active goals conflict with the cooperative principle, and may not be adhered to in stituations like courtroom testimony.) Grice points out that certain utterances take on their meaning from the blatant flouting of one or another of these maxims or principles. That is, their literal meaning is blatantly untrue, irrrelevant or whatever. Many such utterances are examples of the classical rhetorical categories, such as Ironical or Sarcastic, or Metaphorical utterances. (Cf Grice, p52).

For example, if someone who has been done an injury says to their injuror ``You're a fine friend'' (Grice, p53), the utterance has its effect, according to Grice, precisely because the hearer knows that the speaker cannot mean what they say, and must therefore mean the opposite. (Why the opposite though?)

Metaphorical utterances like ``You are the cream in my coffee'' (Grice, p53) also have their effect by virtue of the fact that the hearer recognises that the speaker cannot literally mean what they say, and therefore looks for another meaning. (but how does H know H isnt faced with irony, as in the last example?). Other examples, such as meiosis (rhetorical understatement) and hyperbole (rhetorical overstatement) are in Grice.

Similarly, the utterance ``Lovely weather we are having'' as a response to a tactless remark might be thought of as achieving an act of admonishment by flouting Relation.

Two Proposals

1. Conventional Implicature

(7) #P -> !P
(8) #You're a fine friend -> ! You're a fine friend
Gordon and Lakoff 1971:

(9) #query(P) -> request(P) (10) #"Can you pass the salt?" -> "Please pass the salt"
(Not explanatory).

2. Plan based Inference

``Can you pass the salt'' (Searle 1975)

1. X has asked if I am able to pass the salt
2. I am able to pass the salt
3. X must know this already
4. X must have an ulterior motive.
5. being able to P is an ``enabling condition' for actually P-ing.
6. X has a plan involving my actually passing the salt.
7. My passing the salt is a method for causing X to have the salt.
8. X's goal is to have the salt
9. This is a request to pass the salt.

Exercise: Do a Searle analysis for ``I want you to pass the salt'' and ``Why not pass the salt''. How many ways of indirectly requesting the salt are there?

A ``puzzle'': Why arent any of the following ways of indirectly requesting the salt?
1. # You want to pass the salt.
2. # Do I want you to pass the salt?
(Because they are dysfunctional as direct speech acts.)

A Generalization: You can only make an indirect Speech act by first making a suc- cessful ``direct'' speech act. Understanding proceeds via literal meaning or sense- semantics. It is not an independent parallel process.

Against Speech act Theory

Grice's theory (and its elaboration by Searle) are exciting, and for the first time bring rhetorical uses of language within the scope of computational linguistic and psychological theory. However, there are a number of ways in which they might be thought less than satisfactory.

First, it is always unclear whether certain components of the theory are mere descrip- tions of what happens, or principles of the language understanding process itself. One example is the maxims themselves. At times Grice (and Searle) talk as if, when faced with ``You're a fine friend'' we match it up against the maxims, note that it violates Quality, and use that as the clue to look for the hidden meaning accordingly. But that isnt necessarily the way it works. On the reasonable assumption that we match all other peoples statements against what we already know, and in view of the fact that we already know that we haven't behaved like a friend, the utterance could achieve its effect of making us think of our guilt without the maxim of Quality ever being referred to.

Similarly, one often feels that the speech acts themselves are merely descriptions of what the conversation does, rather than being things that the hearer has to identify. For example, by step 7 of the ``Let's go to the movies'' exchange, the crucial point, that B cant go to the movie has been got across. What further good does it do A to work out that this is the same effect that would have been achieved by the speech act of refusing? Similarly, in Searles ``Can you pass the salt'' example, by step 7, Y has worked out the crucial fact that X wants the salt. What further purpose is fulfilled by going on to work out that this is what would have resulted from a request? Ones doubts are brought to a head by the ``Problems'' for Speech act theory that are courageously revealed by Searle on p75-79, and in particuar by the fact that it is a problem for this analysis that, while you can do indirect requesting by questioning a p``preparatory condition'', such as willingness (do you want to pass the salt?), or by stating a ``sincerity condition'' (I want you to pass the salt ), but not vice versa (*Do I want you to pass the salt? *You want to pass the salt.) That such elementary predictions are ``problems'' suggests that simething is wrong. In particular, it suggests that ideas like ``speech act'', and ``request'' have got too far separate from their realisation and the knowledge of speaker and hearer that they are bound up with. Clearly , those latter are not effective indirect requests because they are not effective direct speech acts, and so can have no side effects at all, let alone those which accomplish the effects of requesting.

Discussion Topics

Earlier, reference was said to be a speech act. As such, it should be possible to use referring expressions to do indirect speech acts, and it is. Consider the indirect consequences of referring to the wife of a friend as ``that woman you live with''. (CF. Grice's ``Miss X produced a series of sounds that corresponded to the score of `Home, Sweet Home''').

How does the hearer of Grice's example of metaphor know that it's metaphor, and not, for example, irony?

In real conversation, is there such a thing as a purely direct speech act?

What is the relation of Grice's Maxims to the Cooperative Principle. Why are there four? Would three do? How would you react to a proposal to add a fifth, such as the Maxim of Entertainment (``Be amusing'')?

Plan-based Speech act Understanding

See papers by Allen et al and Cohen et al. See paper by R.Power. and that by Cassell et al at http://www.cis.upenn.edu/steedman/papers.html

(from Power 1979)

The Universe of Discourse consists of John and Mary on opposite sides of a closed door.

1. JOHN is OUT, MARY is IN, the DOOR is SHUT.
2. JOHN can see, MARY cannot.
3. JOHN BELIEVES that if you move, nothing changes. MARY BELIEVES that if you move, and the door is open, you change position.
4. JOHN has the goal of getting IN. MARY has no goals.

1. JOHN: Mary.
3. JOHN: I want to suggest a Goal.
5. JOHN: Will you help me get in?
7. JOHN: Shall we make a plan?
9. JOHN: Yes.
11. JOHN: Go ahead.
13. JOHN: No.
15. JOHN: OK.
17. JOHN: Yes.
19. JOHN: Go ahead.
21. JOHN: Yes
23. JOHN: Yes.
25. JOHN: Go ahead.
27. JOHN: No

Note the non-hierarchical nature of the dialogue. Johns initial plan gets interrupted by Mary and is never reactivated. It is the goal of the conversation that persists and keeps it on track.

1. JOHN: Mary
3. JOHN: I want to explain something.
5. JOHN: If you move, nothing happens.
7. JOHN: I see.

After further planning in order to find a way of getting the door open, they construct a plan and John changes position to IN by moving. But its not that simple...

1. JOHN: Mary.
3. JOHN: I want to tell you something.
5. JOHN: I have moved.
7. JOHN: Lets assess the result of my action.
8. MARY: OK.
9. JOHN: I have changed position.
11. JOHN: Yes
13. JOHN: Go ahead.
15. JOHN: I disagree. I have changed position.
17. JOHN: I am now in.

Power in 1979 Linguistics gives a complete commentary on this example. Further work: Houghton and Isard 1984. Cassell et al 1994.

Houghton and Isard

Houghton and Isard refined Powers' model in terms of more elaborated models of conversationally relevant information like having the other particulants attention. (Power's model relies on speech acts alone to ensure attention, a fact that gives his dialogue an unnecesarily stilted character.

H&S break the task up into three levels characterized by different kinds of planning, kinds of goals and knowledge sources.

H&S propose a notion of Interaction Frames at Level 2, resembling Power's ``games''. They recognize four types:
  1. Make Known: Impart information to H.
  2. Find Out: Obtain Information from H.
  3. Get Done: Get H to do S a favour.
  4. Get Attention: Get H's attention.
Each Information Frame is characterized by four types of content:
  1. Type of goal the interaction is a method for.
  2. Tests for the preconditions of the interaction.
  3. Procedures for carrying out the non-communicative aspects of the interaction (such as searching memory) on the part of either S or H.
  4. Type of repley expected from H (such as confirmation).
To use an IF, an agent checks the preconditions. If successful, check that other is attending (if not, call Get Attention), and Make Known the goal.

Example: The situation is as before, except that there are three differently coloured doors. This is because other issues of reference (``in'', ``the door'') are also being addressed.

  1. Fred: Doris.
  2. Doris: Yes.
  3. Fred: I want to be in.
  4. Doris: I see
  5. Fred: Could you push the yellow door?
  6. Doris: No, because the bolt isn't up.
  7. Fred: Is there a bolt that is up?
  8. Doris: No
  9. Fred: How do you get a bolt to move?
  10. Doris: You get it to be in, then you slide it.
  11. Fred: Could you slide the green bolt?
  12. Doris: OK.
  13. Fred: Push the door
  14. Doris: OK
(At this point, with the green door open, Fred moves and his goal is achieved.) Cassell, Pelachaud et al. 1994 See the paper by them at: http://www.cis.upenn.edu/ steedman/papers.html