Natural Language Understanding

1
Natural Language Understanding

• Do you understand the following sentence?
• Mary ate spaghetti with George.
• What do we mean by understanding a sentence?

2
Understanding (II)
When the balloon touched the light bulb, it broke
3
Understanding (III)
When the balloon touched the light bulb, it
broke. This caused the baby to cry. Mary gave
John a dirty look and picked up the baby. John
shrugged and picked the balloon.
Questions,Please.
Balloon
Which one got broken?
Baby
Who cried?
Was anyone angry?
?
Did John care?
?
4
Understanding (IV)
When the balloon touched the light bulb, it
broke. This caused the baby to cry. Mary gave
John a dirty look and picked up the baby. John
shrugged and picked the balloon.
Questions,Please.
Yes?
Was the light bulb hot?
Yes?
Was the balloon inflated?
?
Was the balloon exploded?
?
Was John concerned?
Who was responsible for the baby crying?
John
Yes?
Was the baby scared?
5
Knowledge Structures

• Got to have some internal representations for
  sentence meaning
• Got to have some kind of inference generation
• Got to do sentence analysis- I.e. syntax and
  semantics analysis
• Who knows what?

6
Sentence Analysis (Serial Flow of Control)
Input sentence
Grammar
Syntactic Analysis
Parse tree
Semantic Analysis
Features
Semantic representation
Pragmatic analysis
????
Inferences
7
Is serial flow of control enough?

• John took her flowers.
• A stranger took her money.
• Semantics and context are used to resolved the
  syntax ambiguities.

8
Historical Developments

• 1972 Terry Winograd, PhD thesis Understanding
  Natural Language
• 1972 Eugene Charniak, PhD thesis Toward A Model
  of Children Story Comprehension

9
Understanding Natural Language (Winograd)

• Fifteen years ago, a program SHRDLU demonstrated
  that a computer could carry on a simple
  conversation about a blocks world in written
  English. Its success led to claims that NLP had
  been solved and predictions that within a short
  time conversations with computers would be just
  like those with people.

10
Natural Language Understanding (Winograd)

• With years of hindsight and experience, we now
  understand better why the early optimisim was
  unrealisitc. Language, like many human
  capabilities, is far more intricate and subtle
  than it appears on first inspection Winograd,
  1987

11
Comments from Hubert Dreyfus

• .. By 1970, AI had turned into a flourishing
  research program, thanks to a series of
  microworld successes, such as Winograds SHRDU,
  Evans Analog Problem Program and Winstons
  program which learned concepts from examples.
• Then rather suddenly, the field ran into
  unexpected trouble. It started, as far as I can
  tell, with the failure of Charniaks attempts to
  program childrens story understanding. It turned
  out to be a much harder problem than one expected
  to formulate a theory of common sense. It was
  not, as Minsky had hoped, just a question of
  cataloging a few hundred thousand facts.

12
Charniaks Thesis

• An earlier version of the model described in this
  thesis was computer implemented and handled two
  story fragments, about a hundred sentences. The
  problems involved in going from natural language
  to internal representation were not considered,
  so the program does not accept English, but an
  input language similar to the internal
  representation is used (Charniak, 1972)

13
Two veins of research

• Problem-driven research
• Basic research for the long haul. Given the
  difficulties inherent in understanding language,
  what techniques might be of use to us in
  surmounting these difficulties.
• Technology-driven research
• Research of near-term applications. Given the
  current state-of-art, what applications are
  appropriate for the existing technologies?

14
When the balloon touched the light bulb, it
broke. This caused the baby to cry. Mary gave
John a dirty look and picked up the baby. John
shrugged and picked the balloon.

• Observations (1) 7 explicit information are
  given
• (2) There are implicit
  information in the text
• The balloon was original inflated.
• The light bulb was hot.
• The balloon exploded.
• The explosion made a loud noise.
• The baby was scared.
• The loud noise scared the baby.
• Mary picked up the baby to comfort it.

Inferences
15
Charniaks Focus

• Charniaks interest in childrens stories was
  centered on the problem of inference generation.
• Mid-to-late 1970s Most of the wok devoted to
  identifying knowledge structures that could spawn
  inferences.
• Era of Strong methods

16
Script A device for inference Generation

• Proposed for inference generation (first
  knowledge structure)
• Designed to encode stereotypic event sequences.
• e.g. I went to a movie last night
• You understand that
• I must have had money to buy a ticket
• The ticket was purchased at the theatre.
• I may have had to wait in line for a bit before I
  can go into the theater.
• Once inside theater, I could have bought popcorn,
  candy, or ice cream.
• I exchanged the ticket with an usher who gave me
  a stub back.
• .

17
The Balloon Script

• Blow-up balloon Pump-up balloon
• by mouth with
  helium
• Tie balloon
• Balloon whithers Balloon Balloon
• away explodes
  flies away

18
Plans and Goals

• Script is not enough.
• Other knowledge structures such as plans and
  goals were proposed.
• They are a level of abstraction that goes beyond
  scripts but which still allows us to characterize
  stereotypic situations.(1977)

19
Plot units (Lehnert 1981)Summaries for narratives
Plot Unit Graph
Multiple Level abstractions
Affect State map
Script Applications
Analysis of Plans
Analysis of Goals
20
Thomas and Albert
Thomas and Albert respected each others
technical judgement and decided to form a company
together. Unfortunately, Thomas learned that
Albert was notoriously absentminded, whereupon he
instisted that Albert have nothing to do with the
proposed companys finances. This angered Albert
so much that he backed out of their agreement,
hoping that Thomas would be disappointed.
21
John and Mary
John and Mary loved each other and decided to be
married. A month before the wedding, John
discovered that marys father was secretly
smuggling stolen art through Venice. After
struggling with his conscience for days, John
reported marrys father to the police. Mary
understood Johns decision, but she despised him
for it nevertheless, and she broke their engament
knowing that he would suffer.
22
The Gift of the Magi
Della and her husband, Jim, were very poor.
Nevertheless, since Christmas was approaching,
each wanted to give something special to the
other. Della cut off and sold her beautiful hair
to buy an expensive watch fob for Jims heirloom
gold watch. Meanwhile, Jim sold his watch to buy
some wonderful combs for Dellas hair. When they
found out what they had done, they were sad for a
moment, but soon realized that they loved each
other so much, nothing else mattered.
23
Thematic abstraction units (Dyer 1983)Summaries
for narratives
24
End of Era Knowledge Structures
Rich diversity of the knowledge structures
requires lots of things
Boris system (1982)
Script
22
1
3
2
25
Commitment start to shift
We Need
Uniform Knowledge Representations
Uniform inference Mechanism
Elegant Control mechanism
26
Marker Passing
Looking for new work on homogeneous inference
generation Quillian(1968) intersection search
algorithm Rieger(1974) Memory
27
Memory
Sentence S1
Sentence S2
input
input
Memory
inference
inference
S1
S2
inference
Inference meet at the same node implies a
causal chain
28
Memory
Balloon in contact with A hot object
Balloon touches the light bulb
Balloon being Contact With a hot object
Balloon coming into contact with the light bulb
that is turn on
Balloon Explode
Balloon coming into contact with the light bulb
that is turn on And hot
Balloon Break
29
memory

• Not working on knowledge-based framework
• No script
• There are 16 inference classes that was
  responsible for the propagation of inferences (If
  there were knowledge, it would be buried inside
  lisp code somewhere)
• Emphasize on search to create model
• Generates false intersections
• His thesis advisor proposed Scripts because of
  the flaw in Memory

30
Faustus (Norvig 1987)
Based on extensive amounts of Knowledge Use
Marker passing algorithm described by simple
grammar
31
Faustus (II)
Light bulb
Obj 2
Physcont
Obj 1
balloon
Prior state
Object ?
Causal Relation
Object ?
Post event
Breaking
32
Faustus (Inheritance)
Breaking
Broken object
?
Exploded object
exploding
?
Inflated balloon exploding
inflated balloon