📄 A Formal Computational Semantics and Pragmatics of Speech Acts

Summary

Werner addresses a gap in speech act theory. Earlier formal semantics, such as Montague and situation semantics, covered only assertions. Searle and Vanderveken's account of illocutionary force leaves its central notions, the point and the direction of fit, as undefined primitives. The paper argues that in Searle's six-class taxonomy the illocutionary point simply duplicates the class label. It also argues that the sincerity condition is not part of force and that direction of fit is explained circularly. It further rejects Habermas's idea of generalizing truth conditions into satisfaction conditions, since that idea confuses presuppositions with meaning.

As a replacement, the author models each participant's cognitive state as a triple: information, intention (sets of strategies) and evaluation. A speech act's meaning is the change it makes to an idealized hearer's and speaker's state, assuming the message is accepted. Force is then defined as the set of subrepresentations the act is meant to transform. Each of Searle's classes is recast this way, including effects on institutional roles.

The work was done for the WISBER dialogue system and is pitched as groundwork for distributed AI.

Key points

• The analysis is built on Searle's taxonomy: assertives, directives, commissives, expressives, declarations and representative declaratives. That taxonomy uses the dimensions of point, direction of fit, psychological state and propositional content.
• Two of these dimensions add nothing to the classification. Point maps one-to-one onto act type. The speaker's want (W) in directives is shown irrelevant with the example of a general ordering troops against his own wishes.
• Each participant has a representational state R = (I, S, V): information, intentional and evaluative substates. A pragmatic operator, Prag, maps each utterance to updates of these substates, and it depends on the participant's role.
• The class-by-class effects differ by which substate changes:
  • Assertives update the hearer's information and the hearer's model of the speaker's beliefs.
  • Directives update the hearer's intentions.
  • Commissives update the speaker's own intentions.
  • Declarations and representative declaratives additionally alter composite institutional intentions. The latter also require that the speaker's information already supports the content.
• Context: the work comes from the WISBER project (Hamburg, with Nixdorf, Siemens and others) and was presented at COLING 1988.

Interesting ideas

• I1. Define the force of an utterance by which parts of the participants' mental state it is meant to change, not by a label; this makes speech act types operational and testable.
• I2. Separate ideal competence from actual performance: the meaning is the effect on an idealized hearer who accepts the message, not on any real reader's reaction.
• I3. Keep nested models, such as the hearer's model of the speaker's beliefs or intentions, distinct from the agent's own state, so sincerity becomes a checkable relation between the two.
• I4. Let the same utterance produce different updates depending on the social role of each participant, with institutional declarations spreading to everyone in the role structure.
• I5. Treat sincerity, legitimacy and rightness as presuppositions checked separately from the meaning itself, instead of merging them into satisfaction conditions.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): crate prag could tag each speech act with the update targets it implies, such as hearer-beliefs, hearer-intentions or speaker-commitments, and keep that list as the explanation of the act.
• I2 (world): the world of the text could apply speech act updates under an explicit "accepted" assumption and mark them as such, so a refusal or a lie can later be modelled as a deviation.
• I3 (world): fable and tale entities could hold their own state and a model of other characters' states, letting deception or false promises be detected as mismatches.
• I4 (frames): for "You're fired" or a verdict, prag could look up the roles of the participants and emit role-specific consequences, such as an employee no longer being paid.
• I5 (answer): a reader could answer "what did X claim" from the meaning and answer "was X honest or entitled" from a separate presupposition check, keeping both answers explainable.