Pragmatics

1
Pragmatics

• How do we use language to communicate?
• Dave Inman

2
Outline

• 1. Introduction
• 2. Goals and plans
• 3. Example of a "primitive plot" plan
• 4. Frames and scripts
• 5. A "dog" frame
• 6. A "restaurant script"
• 7. Presuppositions
• 8. Rules of discourse
• 9. Why don't these work?

3
Introduction

• This talk looks at what knowledge we use in
  everyday discourse, and how we might use this to
  help a computer understand natural language.
• The kinds of knowledge we will consider here are
  
• Goals and plans
• Frames and scripts
• Presuppositions
• Rules of discourse
• Grice's maxims

4
Introduction - Goals and plans

• e.g. Email dialogue analysis
• If we know what the goal of communication is, or
  the plan of the communicator, we should be able
  to limit the range of possible interpretations.
  If an email requests a meeting, the reply should
  accept, change or deny the request, for example.

5
Introduction - Frames and scripts

• e.g. Shank's restaurant script
• If we can describe the use of language in terms
  of frames (like an object description) or scripts
  (like a film script) then we may be able to match
  the natural language against one of these
  objects, and gain a lot of default expectations

6
Introduction - Presuppositions

• "Have you stopped beating your dog."
• Language often points to the unstated commonly
  understood information. We may be able to use
  inference to find out what this information is.

7
Introduction - Rules of discourse

• Grice's maxims
• Here we can gain information by knowing that
  unless there are "funny circumstances" we tend to
  use language in predetermined ways. For example
  unless we want to deceive we tell the truth. Such
  rules should allow us to gain insights if the
  rules are followed, or to determine what the
  funny circumstance might be if the rules are not
  followed.
• Why do you think this is?

8
2. Goals and plans

• If we discover the goal of a speaker for example,
  it should make interpretation of their
  uttererances easier.
• "How much is a train ticket on the 3pm to
  Bristol?"
• "Nothing" would be inappropriate response if no
  3pm train
• "The next train leaves at 330 and costs 30" if
  plan of speaker is known
• Once a plan is known we can fill in the gaps with
  defaults even if we have no information
• We need "plan templates" for "primitive plots"

9
3. Goals and plans primitive plots

• Problems
• You get fired and need a job
• Your cheque bounces and you need to find money
• You dog dies and you need companionship
• Success
• You ask for a raise and get it
• You fix a flat tyre
• You need a car so you steal one
• Failure
• Your marriage proposal is declined
• You can't find you wallet
• You can't get a bank loan
• Loss / Perseverance / Hidden blessing / Mixed
  blessing etc

10
4. Goals and plans Frames and scripts

• Frames and scripts act as templates of
  expectation.Work by Schank and others
• They have slots and values
• Slots can be
• Compulsory
• Optional
• Default
• Procedures
• Frames are used for concepts
• Scripts are used for sequences of actions / frames

11
5. A "dog" frame

• Make a conversation Frame

12
6. A "restaurant script"

• Write an exam script

13
Goals and plans

• Scene 1 Entering
• Customer enters restaurant
• Customer looks for table
• Customer decides where to sit
• Customer sits down
• Scene 2 Ordering
• Scene 3 Eating
• Scene 4 Leaving
• Fill in the sections above.
• Compare your results to your neighbours.
• What problems do you have here?

14
7. Presuppositions

• "Have you seen John's car?"
• - we should infer that John has a car
• "Was it John who stole the money?"
• - we want to infer that money was stolen, NOT
  John stole it
• - acceptable reply is "No it was Bill"
• - unacceptable reply is "No it was the car"

15
Rules of discourse Grices Maxims

• Be brief
• Be honest
• Be relevant
• Be clear

16
Rules of discourse Grices Maxims

• Be brief
• Give as much info as needed, but not more
• "Who is the President of the USA?"
• "Clinton" - OK
• "A guy who has a wife, 2 daughters, 3 ...called
  Clinton" - Too much

17
Rules of discourse Grices Maxims

• Be honest
• Unless we have a goal of humour, escape
  conviction etc
• If you were a criminal would you lie to
• "Do you take sugar in your coffee?" - to a
  partner
• "What is your annual income?" to the police

18
Rules of discourse Grices Maxims

• Be relevant
• "Can you pass the salt?"
• is not a relevant question in most situations, so
  it must be a request

19
Rules of discourse Grices Maxims

• Be clear
• unless you have reasons / objectives not to be
  such as a politician, answering a student's
  question about exam questions

20
9. Why don't these work?

• Goals and plans
• Frames and scripts
• Presuppositions
• Rules of discourse
• Write down 1 reason why each of these may not
  work.
• Compare with your neighbour.

21
9. Why don't these work?

• Goals and plans
• How many plots do you need?
• How do adapt one plot to make a similar but
  slightly different one?
• How do you match an utterance against a plot?

22
9. Why don't these work?

• Frames and scripts
• How do you acquire a new script for a novel
  situation
• (a) Create a new one - maybe we have too many
• (b) Update an existing one - which one?
• How do we add new knowledge to an existing script
• e.g. being short changed to a restaurant script -
  which other scripts do we need to update?
• shop /garage / cinema /theatre
• what else?
• How do you match an utterance against a frame /
  script?

23
9. Why don't these work?

• Presuppositions
• How do you deal with
• "How many students does Audrey teach?"
• If she is on sabbatical leave we cannot
  presuppose she teaches students.
• - acceptable reply is "She does not teach now"
• - unacceptable reply is "None" this is
  accurate but breaks rules of being relevant

24
9. Why don't these work?

• Rules of discourse
• How do we decide what is "normal" and what is a
  "funny circumstance" ?
• How many "funny circumstances" are there?
• If we have enough "funny circumstances" what
  value are these guidelines?

25
Conclusions

• Language is like the tip of an iceberg.
• Context is like the picture on the box.
• How do we represent context? e.g. CYC
• Can we make do with restricted domains?
• Can we do NLP by example?