Discourse and Pragmatics

1
Discourse and Pragmatics

• Dr. Björn Gambäck
• SICS Swedish Institute of Computer Science AB
• Stockholm, Sweden

2
DUMAS

• Dynamic
• Universal
• Mobility
• for Adaptive
• Speech Interfaces
• Contact dumas_at_sics.se

3
DUMAS Facts

• Development and research project on multilingual,
  adaptive speech interfaces
• Demo application AthosMail
• (e-mail with a dialogue interface over mobile
  phone)
• 3-year project
• started October 2001
• finishes September 2004
• Budget 3,56 million euro (ca 50 million birr)
• 8 members in 4 different countries

4
Motivation

• Interactive applications are a technological and
  commercial possibility
• However
• Incapability to process structured text
• Limited conversational abilities
• Limited user models

5
Objectives

• Dynamic intelligent interaction capabilities
• Universal various users, situations and formats
• Mobility mobile applications
• for
• Adaptive user habits and preferences
• Speech text and speech, multilinguality
• interfaces agent-based architecture

6
Athos Speech Interfaces

• are adaptable to the situation, domain and user
• can be used for various mobile applications
• exploit learning techniques to adapt to a
  particular users preferences
• can handle and produce both speech and text in
  several languages
• are instances of a flexible agent architecture

7
Key Research Areas

• Agent-based development platform
• Athos architecture for speech applications
• Agents, evaluators, managers
• Exploration of machine learning techniques
• Reinforcement learning user interaction
  strategies
• Vector-based methods user priorities, interests
• Dialogue management
• Scenario-based corpus collection in 3 languages
• Dialogue act classification
• Annotation, AG format

8
Key Research Areas, cont.

• Text parsing and analysis
• Robust and multilingual analysis
• User modelling
• Preferences and characteristics of the user
• Presentation techniques
• Provide information in a pleasant and intelligent
  way
• Input handling techniques
• Speech errors, multilingual input

9
Talking to AthosMail

• U I want to listen to my messages.
• A You have two new messages, one from Anders
  about plans and one from Markku about Integration
  Issues.
• U Lets hear the one from Anders.
• A Message content Ill pick the children up
  from the daycare centre at four today so if we
  should discuss changes its good if we do it
  right away. Anders. End of message.
• U Read the message from Markku.
• A Message content Dear all, the latest version
  of the User Modelling Agent is available on the
  CVS. Please download it and make sure it works
  with...
• U Stop!
• Can I hear the one Anna sent yesterday instead?

10
Linguistic Meaning (Gottlob Frege, 1848-1925)

• Giving an account of linguistic meaning
  describing the meanings of complete sentences
• Explaining the meaning of a sentence S
  explaining under which conditions S is true
• Explaining the meanings of other units describe
  how they contribute to Ss meaning

11
Linguistic Meaning (Donald Davidson, 1915- )

• The truth values of the sentences are determined
  by their syntactic structure.
• The meanings of component words is all that a
  theory of meaning for a language can deliver.

12
Linguistic Meaning (Hans Kamp, 194?- )

• A theory of meaning must also say things about
  interpretation.
• A speakers grasp of the meaning of a language
  depends on his ability to interpret sentences he
  hear.
• Truth and interpretation are intimately
  connected.

13
Discourse Representation Structures (DRSs)

• (instead of using first-order representations)
• DRSs are obtained through the application of
  certain rules to the input sentences.
• These rules do not look just at the current
  sentence, but also at DRSs that already has been
  built.

14
Components of DRSs

• a list of discourse referents
• a list of conditions
• If d1, , dn are discourse referents (n gt 0)
• and c1, , cm (m gt 0) are conditions then
• is a DRS.

d1, , dn
c1 cm
15
Indefinites

• A woman snores

x
WOMAN(x) SNORE(x)
X discourse referent from the NP a woman The
VP adds the condition SNORE(x)
16
Proper Names

• Vincent does not die

x
xVINCENT ?
DIE(x)
Proper names also introduce discourse referents
17
Universal Quantification

• every man snores
• (?x(MAN(x) ? SNORE(x))

x
?
MAN(x)
SNORE(x)
18
Reference

• Relationship between linguistic elements (words,
  etc.)
• and the non-linguistic world of experience.
• Indicates which things in the world are talked
  about.
• The same expression can refer to different things
• (e.g. your left ear)
• Two different expression can have the same
  referent
• (e.g., the Morning Star and the Evening Star)

19
Reference, cont.

• Anaphoric reference
• the element referred to has been mentioned
  before
• Antecedent
• the element referred to by the anaphor

• John went to the cinema.
• He goes there often.

• John went to the cinema.
• He goes there often.

20
Pronouns

• Deictic
• accompanied by a deictic act
• (e.g., pointing a finger)
• Anaphoric
• referring to some item mentioned elsewhere
• (the antecedent)

21
Pronoun Resolution

• Salience (recency)
• John has a Fiat.
• Mary has a Ford.
• Betty likes to drive it.
• Selectional restrictions
• John has a Fiat.
• Betty likes it.
• Betty likes him.

22
John watches Big Brother.It fascinates him.

• John watches Big Brother.

x
x y
S
JOHN(x) x watches Big Brother
john watches Big Brother
JOHN(x) BIG BROTHER(y) x watches y
VP
NPmale
NPhum
V
John
x
watches
Big Brother
y
watches Big Brother reducible condition
A proper name introduces a reference marker
23
John watches Big Brother.It fascinates him.

• It fascinates him.

x y
x y u v
x y u
S
JOHN(x) BIG BROTHER(y) x watches y it fascinates
him
JOHN(x) BIG BROTHER(y) x watches y u y u
fascinates him
JOHN(x) BIG BROTHER(y) x watches y u y v x u
fascinates v
VP
NPhum
NPmale
V
It
u
fascinates
him
v
A pronoun introduces a reference marker and a
condition ?? where ? is a suitable marker
24
Text Coherence

• John hid Bills car keys. He got angry.
• John hid Bills car keys. He was drunk.
• John hid Bills car keys. He likes spinach.
• Coherence relations
• Result S1 ? S2
• Explanation S2 ? S1
• (and many more)

25
Dialogue

• Turn-taking
• Utterances
• Dialogue Acts
• GREETING
• REQUEST
• QUESTION
• ANSWER
• COMMAND
• Span one or more utterances

26
Conversational Implicature (Grice)

• Quantity
• Be exactly as informative as required
• Quality
• Be true
• Relevance
• Be relevant
• Manner
• Be perspicuous (avoid ambiguity, etc)

27
Dialogue Structure

• Hi! GREETING
• Hi there, whats up? GREETING, QUEST
• Not much. Did you see the game? ANSWER, QUEST
• Yeah, but they were lousy. CONFIRM, STMNT
• Mmm CONFIRM
• Ok, see you later. CLOSING
• Later, dude! CLOSING

28
Dialogue Management

• System driven (system initiative)
• Prompting
• Slot-filling (frames/templates)
• Finite-state automaton
• User driven
• Mixed-initiative
• (goal oriented)

29
Human-Computer Conversation(Wilks Catizone)

• CONVERSE
• Top-down control of conversation (scripts)
• Large-scale linguistic resources (dictionaries)
• Catherine, 26-year old editor
• Loebner Prize winner 1997

30
Conversational InterfacesAdvances and
Challenges (Zue Glass)

• Mixed-initiative
• Learning
• Robustness
• Dialogue management
• Misunderstandings
• Portability

31
User Modelling in AthosMail

• Record user characteristics and actions
• Enable the system to tailor its responses
• Give expectations of
• user vocabulary
• likely next actions

32
Construction of Semantic Representations

• Three basic principles
• Lexicalization
• try to keep semantic information lexicalized
• Compositionality
• pass information up compositionally from
  terminals
• Underspecification
• Dont make a choice unless you have to
• (the interpretation of ambiguous parts is left
  unresolved)

33
Underspecification

• A meaning ? of a formalism L is underspecified
• represents an ambiguous sentence in a more
  compact manner than by a disjunction of all
  readings
• L is complete Ls disambiguation device
  produces all possible refinements of any ?
• Example
• consider a sentence with 3 quantified NPs
• (with underspecifed scoping relations)
• L must be able to represent all 23! 64
  refinements
• (partial and complete disambiguations) of the
  sentence.

34
Phenomena for Underspecification

• local ambiguities
• e.g., lexical ambiguities, anaphoric or deictic
  use of PRO
• global ambiguities
• e.g., scopal ambiguities, collective-distributive
  readings
• ambiguous or incoherent non-semantic information
• e.g., PP-attachment, number disagreement

35
Lexical ambiguity
Some English words with many senses (from
Merriam-Webster Pocket Dictionary) Word Category
Senses go verb 63 run verb 35 way
noun 31 do verb 30 form noun
24 take verb 24 dead adjective 21
36
Underspecified Semantic Representations

• Reyle Underspecified Discourse Representation
  Structures
• Bos Labelled Underspecified Discourse Structures
  
• Object Language Kamps Discourse Representation
  Structures
• Underspecified w.r.t. scope of quantifying
  expressions
• One underspecified representation describes
  several DRSs