CPSC 503 Computational Linguistics

1
CPSC 503Computational Linguistics

• Discourse and Dialog
• Lecture 14
• Giuseppe Carenini

2
Finish form (22/10)

• Word Sense Disambiguation
• Word Similarity
• Semantic Role Labeling

3
Semantic Role Labeling Example
Some roles..
Employer
Employee
Task
Position

• In 1979 , singer Nancy Wilson HIRED him to open
  her nightclub act .
• Castro has swallowed his doubts and HIRED
  Valenzuela as a cook in his small restaurant .

4
Supervised Semantic Role Labeling

• Typically framed as a classification problem
  Gildea, Jurfsky 2002
• Train a classifier that for each predicate
• determine for each synt. constituent which
  semantic role (if any) it plays with respect to
  the predicate
• Train on a corpus annotated with relevant
  constituent features

These include predicate, phrase type, head word
and its POS, path, voice, linear position and
many others
5
Semantic Role Labeling Example
ARG0
issued, NP, Examiner, NNP, NP?S?VP?VBD, active,
before, ..
predicate, phrase type, head word and its POS,
path, voice, linear position
6
Supervised Semantic Role Labeling (basic)
Algorithm

1. Assign parse tree to input
2. Find all predicate-bearing words (PropBank,
   FrameNet)
3. For each predicate. apply classifier to each
   synt. constituent

Unsupervised Semantic Role Labeling
bootstrapping Swier, Stevenson 04
7
Knowledge-Formalisms Map(including probabilistic
formalisms)
Understanding
Generation
State Machines (and prob. versions)
Morphology
Syntax
Rule systems (and prob. versions)
Semantics

• Logical formalisms
• (First-Order Logics)

Pragmatics Discourse and Dialogue
AI planners (MDPs Markov Decision Processes)
8
Today 27/10

• Brief Intro Pragmatics
• Discourse
• Monologue
• Dialog

9
Semantic Analysis
Sentence
Meanings of grammatical structures

• Syntax-driven and Lexical
• Semantic Analysis

Meanings of words
Literal Meaning
I N F E R E N C E
Common-Sense Domain knowledge
Further Analysis
Discourse Structure
Intended meaning
Context
Pragmatics
10
Pragmatics Example

• (i) A So can you please come over here again
  right now
• (ii) B Well, I have to go to Edinburgh today
  sir
• (iii) A Hmm. How about this Thursday?

What information can we infer about the context
in which this (short and insignificant) exchange
occurred ?
11
Pragmatics Conversational Structure

• (i) A So can you please come over here again
  right now
• (ii) B Well, I have to go to Edinburgh today
  sir
• (iii) A Hmm. How about this Thursday?

Not the end of a conversation (nor the beginning)

• Pragmatic knowledge Strong expectations about
  the structure of conversations
• Pairs e.g., request lt-gt response
• Closing/Opening forms

12
Pragmatics Dialog Acts
(i) A So can you please come over here again
right now (ii) B Well, I have to go to
Edinburgh today sir (iii) A Hmm. How about this
Thursday?

• A is requesting B to come at time of speaking,
• B implies he cant (or would rather not)
• A repeats the request for some other time.

• Pragmatic assumptions relying on
• mutual knowledge (B knows that A knows that)
• co-operation (must be a response triggers
  inference)
• topical coherence (who should do what on Thur?)

13
Pragmatics Specific Act (Request)
(i) A So can you please come over here again
right now (ii) B Well, I have to go to
Edinburgh today sir (iii) A Hmm. How about this
Thursday?

• A wants B to come over
• A believes it is possible for B to come over
• A believes B is not already there
• A believes he is not in a position to order B to

Pragmatic knowledge speaker beliefs and
intentions underlying the act of requesting
Assumption A behaving rationally and sincerely
14
Pragmatics Deixis
(i) A So can you please come over here again
right now (ii) B Well, I have to go to
Edinburgh today sir (iii) A Hmm. How about this
Thursday?

• A assumes B knows where A is
• Neither A nor B are in Edinburgh
• The day in which the exchange is taking place is
  not Thur., nor Wed. (or at least, so A believes)

Pragmatic knowledge References to space and time
wrt space and time of speaking
15
Today 27/10

• Brief Intro Pragmatics
• Discourse
• Monologue
• Dialog

16
Discourse Monologue

• Monologues as sequences of sentences have
  structure
• Tasks Text Segmentation and Rhetorical
  (discourse) parsing and generation

(like sentences as sequences of words)

• Key discourse phenomenon referring expressions
  (what they denote may depend on previous
  discourse)
• Task Coreference resolution

17
Discourse/Text Segmentation(1)

• State of the art
• linear (unable to identify hierarchical
  structure)
• Subtopics, passages

• UNSUPERVISED
• Key idea lexical cohesion (vs. coherence)
• There is not water on the moon. Andromeda is
  covered by the moon.

• Discourse segments tend to be lexically cohesive
• Cohesion score drops on segment boundaries

18
Discourse/Text Segmentation(2)

• SUPERVISED
• Binary classifier (SVM, decision tree,)
• make yes-no boundary decision between any two
  sentences

• features
• Cohesion features (e.g., word overlap, word
  cosine)
• Presence of (domain specific) discourse markers
• News good evening, I am.., joining us now is
• Real estate ads is previous word phone number?

19
Sample Monologues Coherence
House-A is an interesting house. It has a
convenient location. Even though house-A is
somewhat far from the park, it is close to work
and to a rapid transportation stop.
It has a convenient location. It is close to
work. Even though house-A is somewhat far from
the park, house-A is an interesting house. It is
close to a rapid transportation stop.
20
Corresponding Text Structure
House-A is an interesting house.
CORE-1
CONCESSION-1
EVIDENCE-1
It has a convenient location.
it is close to a rapid transportation stop
it is close to work
Even though house-A is somewhat far from the park
decomposition
ordering
rhetorical relations
21
Text Relations, Parsing and Generation

• Rhetorical (coherence) Relations
• different proposals (typically 20-30 rels)
• Elaboration, Contrast, Purpose

• Parsing Given a monologue, determine its
  rhetorical structure Marcu, 00 and 02

• Generation Given a communicative goal e.g.,
  convince user to quit smoking generate
  structure
• Next class

22
Reference
Language contains many references to entities
mentioned in previous sentences (i.e., in the
discourse context/model)

• I saw him
• I passed the course
• Id like the red one
• I disagree with what you just said
• That caused the invasion

• Two tasks
• Anaphora/pronominal resolution
• Co-reference resolution

23
Reference Resolution
Terminology Referring expression NL expression
used to perform reference Referent entity that
is referred

• Types of referring expressions
• Indefinite NP (a, some, )
• Definite NP (the, )
• Pronouns (he, she, her,...)
• Demonstratives (this, that,..)
• Names

• Inferrables
• Generics

24
Pronominal Resolution Simple Algorithm

• Last object mentioned (correct gender/person)
• John ate an apple. He was hungry.
• He refers to John (apple is not a he)
• Google is unstoppable. They have increased..

• Selectional restrictions
• John ate an apple in the store.
• It was delicious. stores cannot be delicious
• It was quiet. apples cannot be quiet

• Binding Theory constraints
• Mary bought herself a new Ferrari
• Mary bought her a new Ferrari

25
Additional Complications

• Some pronouns dont refer to anything
• It rained
• must check if verb has a dummy subject

• Evaluate last object mentioned using parse
  tree, not literal text position
• I went to the GAP which is opposite to BR.
• It is a big store.

GAP, not BP
26
Focus

• John is a good student
• He goes to all his tutorials
• He helped Sam with CS4001
• He wants to do a project for Prof. Gray

He refers to John (not Sam)
27
Supervised Pronominal Resolution
Corpus annotated with co-reference relations (all
antecedents of each pronoun are marked)

• What features ?

(U1) John saw a nice Ferrari in the parking
lot (U2) He showed it to Bob (U3) He bought it
28
Need World Knowledge

• The police prohibited the fascists from
  demonstrating because they feared violence.
• vs
• The police prohibited the fascists from
  demonstrating because they advocated violence.

Exactly the same syntax!

• Not possible to resolve they without detailed
  representation of world knowledge about feared
  violence vs. advocated violence

29
Coreference resolution

• Decide whether any pair of NPs co-refer
• Binary classifier again

anaphor
NPj
antecedents

• What features?
• Same as for anaphora specific ones to deal with
  definite and names. E.g.,
• Edit distance
• Alias (based on type e.g., for PERSON Dr. or
  Chairman can be removed)
• Appositive (Mary, the new CEO, .

30
Today 27/10

• Brief Intro Pragmatics
• Discourse
• Monologue
• Dialog

31
Discourse Dialog

• Most fundamental form of language use
• First kind we learn as children

Dialog can be seen as a sequence of communicative
actions of different kinds (dialog acts) - (DAMSL
1997 20)
32
Dialog two key tasks

• (1) Dialog act interpretation identify the user
  dialog act

• (2) Dialog management (1) decide what to say
  and when

33
Dialog Act Interpretation

• What dialog act a given utterance is?

• Surface form is not sufficient!

E.g., Im having problems with the homework

• Statement - prof. should make a note of this,
  perhaps make homework easier next year
• Directive - prof. should help student with the
  homework
• Information request - prof should give student
  the solution

34
Automatic Interpretation of Dialog Acts
State Machines (and prob. versions)
Morphology
Cue-based
Syntax
Rule systems (and prob. versions)
Semantics

• Logical formalisms
• (First-Order Logics)

Pragmatics Discourse and Dialogue
Plan-Inferential
AI planners
35
Plan Inferential (BDI) Pros/Cons

• Dialog acts are expressed as plan operators
  involving belief, desire, intentions

• Powerful uses rich and sound knowledge
  structures -gt should enable modeling of subtle
  indirect uses of dialog acts

36
Cue-Based Key Idea

• Words and collocations
• Please and would you -gt REQUEST
• are you and is it -gt YES-NO-QUESTIONs

Prosody Loudness or stress yeah -gt AGREEMENT
vs. BACKCHANNEL

• Conversational structure
• Yeah following PROPOSAL -gt AGREEMENT
• Yeah following INFORM -gt BACKCHANNEL

37
Cue-Based model (1)

• Each dialog act type (d) has its own
  micro-grammar which can be captured by N-gram
  models

Annotated Corpus

• Lexical given an utterance W w1 wn for each
  dialog act (d) we can compute P(Wd)

• Prosodic given an utterance F f1 fn for each
  dialog act (d) we can compute P(Fd)

38
Cue-Based model (2)

• Conversational structure Markov chain

• 1

Annotated Corpus

• 1

• .8

• .3

• 1

• .2

• .2

• .3

• .5

• .7

Combine all info sources HMM
N-gram models!
39
Cue-Based model Summary

• Start form annotated corpus (each utterance
  labeled with appropriate dialog act)

• For each dialog act type (e.g., REQUEST), build
  lexical and phonological N-grams

• Build Markov chain for dialog acts (to express
  conversational structure)

40
Dialog Managers in Conversational Agents

• Examples Airline travel info system,
  restaurant/movie guide, email access by phone

• Tasks
• Control flow of dialogue (turn-taking)
• What to say/ask and when

41
Dialog Managers
State Machines (and prob. versions)
Morphology
FSA
Syntax
Rule systems (and prob. versions)
Semantics
Template-Based
Pragmatics Discourse and Dialogue

• Logical formalisms
• (First-Order Logics)

BDI
MDP
AI planners (and prob. versions)
42
27/10 Probably stop here
43
FSA Dialog Manager system initiative

• xxx

44
Template-based Dialog Manager (1)

• GOAL to allow more complex sentences that
  provide more than one info item at a time

S How may I help you? U I want to go from
Boston to Baltimore on the 8th.
Slot Optional questions From_Airport
From what city are you leaving? To_Airport
Where are you going? Dept-Time When do
you want to leave? Dept-Day

• Interpretation Semantic Grammars, semi-HMM,
  Hidden-Understanding-Models (HUM)

45
Template-based Dialog Manager (2)

• More than one template e.g., car or hotel
  reservation

• User may provide information to fill slots in
  different templates

• A set of production rules fill slots depending on
  input and determines what questions should be
  asked next

E.g., IF user mention car slot and most of air
slot are filled THEN ask about remaining car
slots.
46
Markov Decision Processes 02

• Common formalism in AI to model an agent
  interacting with its environment.
• States / Actions / Rewards

• Application to dialog
• States slot in frame currently worked on, ASR
  confidence value, number of questions about
  slot,..
• Actions questions types, confirmation types
• Rewards user feedback, task completion rate

47
BDI Dialog Manager
S1 How may I help you? U1 I want to go to
Pittsburgh in April. S2 And, what date in April
do you want to travel? U2 Uh hmm I have a mtg.
there on the 12th.
REQUEST
ACKNOWLEDGE
REQUEST
INFORM

• Sys to understand U2 needs model of
  preconditions, effects, decomposition of
• meeting event (precon be there)
• fly-to plan (decomp book-flight, take-flight)
• Take-flight plan (effect be there)

48
BDI Dialog Manager
S1 How may I help you? U1 I want to go to
Pittsburgh in April. S2 And, what date in April
do you want to travel? U2 Uh hmm I have a mtg.
there on the 12th.
REQUEST
ACKNOWLEDGE
REQUEST
INFORM

• Sys to generate S2 needs model preconditions of
• Book-flight action (agent knows departure date
  and time)

Integrated with logic-based planning system

• Generating an utterance plan generation
  (possibly) satisfying multiple goals

• Understanding an utterance plan recognition
  (recognize multiple goals)

49
Designing Dialog Systems User-Centered Design

• Early Focus on User and Task e.g., interview the
  users

• Build Prototypes Wizard-of-Oz (WOZ) studies

• Evaluation

Iterative Design
50
Next Time Natural Language Generation

• Read handout on NLG
• Lecture will be about an NLG system that I
  developed and tested