Speech acts (dialogue acts, conversational moves)

A tagset for utterances. For computer dialogue, it's necessary to have a system which can tag utterances automatically.

Problem: "indirect speech acts" . Some requests:
· "Could you tell me if Delta has a hub in Boston?"
· "I want to fly from Boston to San Francisco"
· "May I get a lunch on flight UA21 instead of breakfast?"

"Plan-inferential" approach

uses inference to relate utterances to desired outcomes. Uses "BDI" (belief, desire, intention) models.

Action schemas:
· preconditions (must be true before the action can take place)
· effects (become true when the action is performed)
· body (substeps of the action)

Speech acts can be represented as action schemas. For instance, to request something:
· Precondition: speaker wants the hearer to do it
· Effect: hearer wants to do it
· Body: hearer believes that the speaker wants the hearer to do it

Add some inference rules, such as that if someone wants something they want its effects.

Cue-based approach

The above approach is very expensive. A cheaper approach takes into account cues that are in the utterance.
· Words/collocations (e.g. "please…"); can be "learned" automatically by using training corpora
· Prosody (e.g. rising pitch); can also be trained from acoustic features like pitch, slope, energy, duration
· Conversational structure (discourse context): things after questions are probably answers