Transcript NLG - Indian Institute of Technology Kharagpur

Natural Language
Generation
An Introductory Tour
Anupam Basu
Dept. of Computer Science &
Engineering
IIT Kharagpur
Summer School on Natural Language Processing and Text Mining 2008
Language Technology
Meaning
Natural Language
Understanding
Natural Language
Generation
Text
Text
Speech
Recognition
Speech
Synthesis
Speech
Speech
What is NLG?
Thought / conceptualization of the world
------ Expression
The block c is on block a
The block a is under block c
The block b is by the side of a
The block b is on the right of a
The block b has its top free
The block b is alone ………
Conceptualization

Some intermediate form of representation
ON (C, A)
ON (A, TABLE)
ON (B, TABLE)
RIGHT_OF (B,A) …….
What to say?
Conceptualization
Block
Is_a
C
On
Is_a
A
B
Right_of
What to say?
What to say ? How to say ?
Natural language generation is the
process of deliberately constructing a
natural language text in order to meet
specified communicative goals.
[McDonald 1992]
Some of the Applications

Machine Translation

Question Answering

Dialogue Systems

Text Summarization

Report Generation
Thought / Concept  Expression

Objective:


Input:


produce understandable and appropriate texts
in human languages
some underlying non-linguistic representation
of information
Knowledge sources required:

Knowledge of language and of the domain
Involved Expertise

Knowledge of Domain



Knowledge of Language


Lexicon, Grammar, Semantics
Strategic Rhetorical Knowledge


What to say
Relevance
How to achieve goals, text types, style
Sociolinguistic and Psychological Factors

Habits and Constraints of the end user as an information
processor
Asking for a pen

have(X, z)
not have (Y,z)

want have (Y,z)
Situation
Why?
Goal

What?
Conceptualization

How?
Expression
ask(give (X,z,Y)))
Could you please give me a pen?
Some Examples
Summer School on Natural Language Processing and Text Mining 2008
Example System #1: FoG





Function:
 Produces textual weather reports in English and French
Input:
 Graphical/numerical weather depiction
User:
 Environment Canada (Canadian Weather Service)
Developer:
 CoGenTex
Status:
 Fielded, in operational use since 1992
FoG: Input
FoG: Output
Example System #2: STOP





Function:
 Produces a personalised smoking-cessation leaflet
Input:
 Questionnaire about smoking attitudes, beliefs, history
User:
 NHS (British Health Service)
Developer:
 University of Aberdeen
Status:
 Undergoing clinical evaluation to determine its effectiveness
STOP: Input
SMOKING QUESTIONNAIRE
Please answer by marking the most appropriate box for each question like this: 
Q1 Have you smoked a cigarette in the last week, even a puff?
YES 
Please complete the following questions
Please read the questions carefully.
Q2
Home situation:
Live

alone
NO

Please return the questionnaire unanswered in the
envelope provided. Thank you.
If you are not sure how to answer, just give the best answer you can.
Live with

husband/wife/partner
Live with
other adults

Q3
Number of children under 16 living at home
Q4
Does anyone else in your household smoke? (If so, please mark all boxes which apply)
husband/wife/partner 
other family member 
others 
Q5
………………… boys
Live with
children
How long have you smoked for? …10… years
Tick here if you have smoked for less than a year

………1……. girls

STOP: Output
Dear Ms Cameron
Thank you for taking the trouble to return
the smoking questionnaire that we sent you.
It appears from your answers that although
you're not planning to stop smoking in the
near future, you would like to stop if it
was easy. You think it would be difficult
to stop because smoking helps you cope with
stress, it is something to do when you are
bored, and smoking stops you putting on
weight. However, you have reasons to be
confident of success if you did try to
stop, and there are ways of coping with the
difficulties.
Approaches
Summer School on Natural Language Processing and Text Mining 2008
Template-based generation
• Most common technique
• In simplest form, words fill in slots:
 “The train from Source to Destination will
leave platform number at time hours”
Most common sort of NLG found in
commercial systems
Pros and Cons


Pros

Conceptually simple

No specialized knowledge needed

Can be tailored to a domain with good performance
Cons

Not general

No variation in style – monotonous

Not scalable
Modern Approaches

Rule Based approach

Machine Learning Approach
Some Critical Issues
Summer School on Natural Language Processing and Text Mining 2008
Context Sensitivity in Connected
Sentences

X-town was a blooming city. Yet, when the hooligans
started to invade the place, __________ . The place
was not livable any more.

the place was abandoned by its population

the place was abandoned by them

the city was abandoned by its population

it was abandoned by its population

its population abandoned it……..
Referencing
John is Jane’s friend. He loves to swim with
his dog in the pool. It is really lovely.
I am taking the Shatabdi Express
tomorrow. It is a much better train than
the Rajdhani Express. It has a nice
restaurant car, while the other has nice
seats.
Referencing
John stole the book from Mary, but he was
caught.
John stole the book from Mary, but the fool
was caught.
Aggregation
The dress was cheap.
The dress was beautiful
The dress was cheap and beautiful
The dress was cheap yet beautiful
I found the boy. The boy was lost.
I found the boy who was lost
I found the lost boy.
Sita bought a story book. Geeta bought a story
book.
???? Sita and Geeta bought a story book.
???? Sita bought a story book and Geeta also
bought a story book
Choice of words (Lexicalization)
The bus was in time. The journey was fine.
The seats were bad.
The bus was in perfect time. The journey
was fantastic. The seats were awful.
The bus was in perfect time. The journey
was fantastic. However, the seats were not
that good.
General Architecture
Summer School on Natural Language Processing and Text Mining 2008
Component Tasks in NLG

Content Planning
=== Macroplanner

Document Structuring

Sentence Planner


=== Microplanning
Aggregation ; Lexicalization; Referring Expression
Generation
Surface Form Realization

Linguistic realization; Structure Realization
A Pipelined Architecture
Document
Planning
Document Plan
Microplanning
Text Specification
Surface
Realization
An Example
Consider two assertions
has (Hotel_Bliss, food (bad))
has (Hotel_Bliss, ambience (good))
Content Planning selects information ordering
Hotel Bliss has bad food but its ambience is
good
Hotel Bliss has good ambience but its food is
good
has (Hotel_Bliss, food (bad))
Sentence Planning
choose syntactic templates Subj
choose lexicon
Entity
bad or awful
food or cuisine
good or excellent
Aggregate the two propositions
Generate referring expressions
It or this restaurant
Ordering
A big red ball OR A red big ball
Have
Obj
Feature
Modifier
Realization
correct verb inflection Have  Has
may require noun inflection (not in this case)
Articles required? Where?
Conversion into final string
Capitalization and Punctuation
Content Planning

What to say




Data collection
Making domain specific inferences
Content selection
Proposition formulation


Each proposition  A clause
Text structuring


Sequential ordering of propositions
Specifying Rhetorical Relations
Content Planning Approaches

Schema based (McKeown 1985)



Application of operators (similar to Rule
Based approach) --- Hovy 93


Specify what information, in which order
The schema is traversed to generate discourse
plan
The discourse plan is generated dynamically
Output is Content Plan Tree
Discourse
Detailed
view
Demograph
Name
Summary
Group
nodes
Age
Blood
Sugar
Care
Content Plan
Plan Tree Generation
 Ordering – of Group nodes
 Propositions


Rhetorical relations between leaf nodes

Paragraph and sentence boundaries
Rhetorical Relations
ENABLEMENT
MOTIVATION
MOTIVATION
You should ... I’m in ... The show ...
EVIDENCE
It got a ...
You can get ...
Rhetorical Relations
Three basic rhetorical relationships:
 SEQUENCE
 ELABORATION
 CONTRAST
Others like
 Justification
 Inference
Nucleus and Satellites
Contrast
I drive my Maruti 800
Elaboration
I love to collect
classic cars
N
My favourite car is
Toyota Innova
Target Text
The month was cooler and drier than
average, with the average number of
rain days, but the total rain for the
year so far is well below average.
Although there was rain on every day
for 8 days from 11th to 18th,
rainfall amounts were mostly small.
Document Structuring in
WeatherReporter
The Message Set:
MonthlyTempMsg ("cooler than average")
MonthlyRainfallMsg ("drier than average")
RainyDaysMsg ("average number of rain days")
RainSoFarMsg ("well below average")
RainSpellMsg ("8 days from 11th to 18th")
RainAmountsMsg ("amounts mostly small")
Document Structuring in Weather
Reporter
SEQUENCE
ELABORATIO
N
ELABORATIO
N
MonthlyTmp
Msg
Monthly
RainfallMsg
CONTRAST
RainyDays
Msg
CONTRAST
RainSoFar RainSpell
Msg
Msg
RainAmounts
Msg
Some Common RST Relationships

Elaboration: The satellite presents more details about the
content of the nucleus

Contrast: The nuclei presents things, which are similar in
some respects but different in some other relevant way.

Multinuclear – no distinction bet. N and S

Purpose: S presents the goal of performing the activity
presented in the nucleus

Condition: S presents something that must occur before the
situation presented in N can occur

Result: N results from S
Planning Approach
Save
Document
Choose
Save
option
A dialog
box
displayed
Select
Folder
The system
saves the
document
Type
Filename
Click
Save
Button
Dialog box
closed
Planning Operator
Name: Expand Purpose
Effect:
(COMPETENT hearer(DO-ACTION ?action))
Constraints:
(AND (get_all_substeps ?action ?subaction)
(NOT (singular list ?subaction))
Nucleus:
(COMPETENT hearer (DO-SEQUENCE ?subaction))
Satellite:
(((RST-PURPOSE (INFORM hearer (DO ?action)))
Expand Subactions
Effect:
(COMPETENT hearer (DO-SEQUENCE ?actions))
Constraints:
NIL
Nucleus:
(for each ?actions (RST-SEQUENCE
(COMPETENT hearer (DO-ACTION ?actions))))
Satellites:
NIL
Purpose
Sequence
Choose Folder
Choose
Save
Dialog
Box
Opens
Result
Discourse

To save a file




1. Choose save option from file menu
A dialog box will appear
2. Choose the folder
3. Type the file name
4. Click the Save button
The system will save the document
Rhetorical Relations – Difficult to infer
Johh abused the duck
The duck buzzed John
1.
2.
3.
4.
John abused the duck that had buzzed
him
The duck buzzed John who had abused it
The duck buzzed John and he abused it
John abused the duck and it buzzed him
On Clause Aggregation
Summer School on Natural Language Processing and Text Mining 2008
Benefits of Aggregation

Conciseness


Cohesion


Same information with fewer words
We want a semantic unit – not a jumble of
disconnected phrases
Fluency


Less effort to read
Unambiguous and acc. to communication
conventions
Complex interactions

Aggregation adds to fluency
The patient was admitted on Monday and
released on Friday.
 Someone ate apples. Someone ate
oranges
 Someone, who ate apples also ate oranges
Aggregation Operators
Category
Operators
Resources
Surface
markers
Interpretive
Summarization
Inference
Common
sense
knowledge
Ontology
Referential
Ref. expr.
Generation
Quantified
expression
Ontology
Discourse
Each, all
both some
Syntactic
Paratactic
Hypotactic
Syntactic rules
Lexicon
And, with,
who, which
Lexical
Paraphrasing
Lexicon
Interpretive
John punched Mary
Mary kicked John => John fought with Mary
John kicked Mary
Not always meaning preserving
Note use of Ontology
John kicked Mary + John punched Mary
=/>
John fights with Mary
Referential Aggregation

Reference Expression generation
The patient is Mary [name].
The patient is female [gender]
The patient is 80 years old [age].
The patient has hypertension [med.history]
How much info in
one sentence?
The patient is Mary. She is an 80 year old
female. She has hypertension.
Reference ( Quantification)

John is doing well
Mary is doing well  All the patients are
doing well

Note the use of background knowledge


The patient’s leftarm
The patient’s right arm  Each arm

Note the use of Ontology

Syntactic Aggregation

Paratactic: Entities are of equal syntactic status
Ram likes Sita and Geeta
Main operator is co-ordinating conjunction

Hypotactic: Unequal status
NP modified by a PP
Ram likes Sita, who is a nurse
Lexical Aggregation

In hypotactic aggregation, the satellite propositions are
modified.

The Maths score was 99.8%
99.8% is a record high score
The Maths score was 99.8%, a record high score
(apposition modification)



The Maths score was a record high score of 99.8%

A dog used by police  A police dog
Rise sharply  shoot
Drop sharply  plunge


Rhetorical Relations and Hypotactics
Use of cue operators
RR: Concession
He was fine
He just had an accident
Although he had an accident he was fine
RR: Evidence
My car is not Indian
My car is a Toyota
My car is not Indian because it is a Toyota
RR: Elaboration
My car is not Indian
My car is expensive
My expensive car is not Indian
Hypotactic Operators

If propositions do not share any common entity, the
operator can simply join using cue phrase
N:Tom is feeling cold
S:The window is open
Tom is feeling cold because the window is open

Cause
If the linked propositions share common entities, the
internals of the linked propositions undergo modifications
N: The child stopped hunger S: The child ate an apple
[Purpose]
To stop hunger, the child ate an apple.
Two stage transformation:
RR: Evidence
N: Tom was hungry
S: Tom did not eat dinner
Replace Tom in N by ‘he’
Apply Rule 1
Because Tom did not eat dinner, he was
hungry
Corpus study to Rules [Example RR: Purpose
N: Lift the cover S: Install battery]
%
Example
To-infinitive
59.6
To install battery, lift
the cover
For-Nominalization
7.5
Lift the cover for
battery installation
For-Gerund
2.5
Lift the cover for
installing battery
By-pupose
10
Install battery by
lifting cover
So-Tat Purpose
8.4
Lift cover so battery
can be installed
Syntactic constructions for realizing
Elaboration relations
Verbosity
M-direction
Examples
R-Clause
Short
Before
An apple which
weighs 3 oz
Reduced R-Clause
Shorter
Before
An apple
weighing 3oz
PP
Shorter
Before
An apple in the
basket
Apposition
Shortest
Before
An apple, a
small fruit
Prenominalization
Shortest
After
A 3 oz apple
Adjective
Shortest
After
A dark red
apple
Lexical Constraints

Except for R-clause and Reduced R-clause, transforming a
proposition into an apposition, an adjective or a PP requires
that the satellite proposition be of a specific syntactic type (
a noun, an adj or a PP respectively).
N: Jack is a runner.
S: Jack is fast.
Jack is a fast runner
Fast and runner has a possible qualifying relationship.
Qualia Structure (Pustejovsky 91)
Constraints

Linear Ordering

Paratactic

Years 1998,1999 and 2000
 Not Years 1999, 1998 and 2000

Hypotactic

Uncommon orderings between premodifiers create
disfluencies
 A happy old man ---- An old happy man
Linear Ordering and Scope of Modifiers
Problem
when multiple modifiers modify
the same noun


Decide the order
Avoid ambiguity
Ms. Jones is a patient of Dr. Smith, undergoing heart
surgery
Old men and women should board first
Women and old men should board first
Linear Ordering of Modifiers

A simplex NP is a maximal noun phrase that includes premodifiers such as determiners and possessives, but not
post-nominals such as PPs and R-Cls.

A POS tagger along with FS grammar can be used to
extract simples NPs.

A morphology module transforms plurals of nouns,
comparative and superlative adjectives into their base form
for frequency count.

Regular expression filter to remove concatenations of NPs


Takeover bid last week
Metformin 500 milligrams
Three stages of subsequent analysis

Direct Evidence

Modifier sequences are transformed in ordered
pairs

Well known traditional brand name drug




Well known < traditional
Well known < brand name
traditional < brand name
Three possibilities

A < B ; B< A; B=A (no order)

For n modifiers nC2 ordered pairs

Form a w X w matrix where w is the
number of distinct modifiers.

Find Count[A,B] and Count[B,A]

For small corpus binomial distribution of
one following the other is observed.

Transitivity
Again from corpus
A < B and B< C ? A < C
Long, boring and strenuous stretch
Long strenuous lecture

Clustering: Formation of equivalence classes of
words with same ordering with respect tp other
premodifiers
John is a 74 year old hypertensive diabetic
white male patient with a swollen mass in
the left groin
John is a diabetic male white 74 year old
hypertensive patient with a red swollen
mass in the left groin
Other Constraints

For conjunctions




Moral: Same syntactic category?



John ate an apple and an orange (NP and NP)
John ate in the morning and in the evening (PP and PP)
X John ate an apple and in the evening (NP and PP)
John and a hammer broke the window ???
He is Nobel Prize winner and at the peak of his career.
Others: Adj phrase attachment, PP attachment
etc.
Conjunctions
Summer School on Natural Language Processing and Text Mining 2008
Three interesting types

John ate fish on Monday and rice on
Tuesday (non-constituent coordination)

John ate fish and Bill rice (gapping)

Right node raising

John caught and Mary killed the spider
A Naïve Algorithm
1.
Group propositions and order them
according to similarities
1.I sold English books on Monday
2.I sold Hindi books on Wednesday
3.I sold onion on Monday
4.I sold Bengali books on Monday
((1,3,4),2) OR ((1,4),3,2) OR…..
2. Identify recurring elements
3. Determine sentence boundary
4. Delete redundant elements
Still Funny Scenarios
The baker baked. The bread baked.
  The baker and the bread baked.

I don’t drink. I don’t chew tobacco.
  I don’t drink and chew tobacco.


==What should the constraints be?
Morphological Synthesis

Inflections depending on tense, aspect, mood, case, gender,
number, person and familiarity.

A typical Bengali verb has 63 different inflected forms (120
if we consider the causative derivations)
Exceptions

Synthesis Approach

Classification of words based on Syllable
structure [19 classes for Bengali verbs]

Paradigm tables for each of the classes

Table-driven modification of the words

Exceptions treated separately.
Noun Morphology Synthesis

Different rules are used to inflect qualifiers and headwords
The rule to inflect proper noun as a headword in a particular SSU
IF (headword type = proper noun AND the SSU to which the headword
belongs = kAke AND the last character of root word = ‘a’),
THEN
Rule1: headword = headword + “ke”
rAma  rAmake
IF (Verb1==verb2 AND the Conjunction = Ebong AND SSU2 to which the
headword belongs = kakhana AND the last character of root word = ‘a’)
THEN
Rule1: headword = headword –’a’.
Rule2: headword = headword +’o’.
Aaem gfkal bl /K/leClam ybL Aajo /Klb.
Headword : Aaj + o
Verb Morphology Synthesis
•
Depends upon TAM option.
Category Identification +Table lookup
Category Identification: Structure of root verb: X * VC * $. where: X= Any
Character, V= vowel, C=constant and $ € { Ø, a, A, oYA }.
$

a*
A
oYA
V
a
A
ha [haoYA]
(to happen)
khA [khAoYA]
kara [karA]
(to do)
jAna [jAnA]
(to know)
karA[karAno]
(do, causative)
saoYA [saoYAno]
(undergo, causative)
jAnA [jAnAno]
(to inform)
khAoYA [khAoYAno]
(to feed)
(to eat)
i
e
o
u/au
di [deoYA]
(to give)
likha [lekhA]
(to write)
dekha [dekhA]
(to see)
so [so;oYA]
(to lie down)
tola [tolA]
(to pick)
ni~NrA [ni~NrAno]
dekhA
[dekhAno]
(to show)
tolA [tolAno]
deoYA [deoYAno]
(give, causative)
(pick, causative)
ghumA [ghumAno]
(to sleep)
(lie, causative)
so;oYA [so;oYAno]
Table Look Up
 The Table Lookup
Stage:
i) Pr  Present
ii) Pa  Past
iii) Sim  Simple
iv) Per  Perfect
v) Co  Continuous
vi) Ind  Indicative
vii) Neg  Negation.
?Questions?
Summer School on Natural Language Processing and Text Mining 2008