8 November 2003 PP attachment problem1 Prepositional Phrase Attachment Problem 03M05601 Ashish Almeida.

8 November 2003 PP attachment problem 1

Prepositional Phrase Attachment Problem

03M05601

Ashish Almeida

8 November 2003 PP attachment problem 2

Overview

– Introduction to NLP– Analysis in UNL system– Prepositional phrase attachment problem– Proposed method to handle this problem

8 November 2003 PP attachment problem 3

Motivation

• Analysis involves many complex problems• Prepositional phrase attachment problem is

one such difficult problem.• If solved, improve the quality of information

extracted manifold• No existing system solves the problem

8 November 2003 PP attachment problem 4

Tasks involved in NLP

Analysis and generation

Text Meaning

NL understanding

NL generation

8 November 2003 PP attachment problem 5

Phases in NLP

• Morphological analysis• Syntactic analysis• Semantic analysis

• Discourse integration• Pragmatic analysis

8 November 2003 PP attachment problem 6

Is NL Compositional ?

• Compsitional expression– Meaning of the whole from meaning of

parts

e.g. strong tea

- rich tea

day by day

- all the time

8 November 2003 PP attachment problem 7

Analysis

Morphological + Syntactic + Semantic analysis

• All these phases are dependent on each other.

• Interactive Vs modular approach

• Analysis in UNL system - interactive

8 November 2003 PP attachment problem 8

UNL …• UNL is Interlingua

e.g. Ram ate rice with spoon.

agt

obj

spoon(icl>artifact)John(iof>person)

rice(icl>food)

eat(icl>do)@ entry@ present

ins

8 November 2003 PP attachment problem 9

UNL expresion

UNL Expression for Ram ate rice with spoon.agt(eat(icl>do).@past.@entry, Ram(iof>person))

obj(eat(icl>do).@past.@entry, rice(icl>food))

ins(eat(icl>do).@past.@entry, spoon(icl>tool))

Relation AttributesUWs

agt(eat(icl>do).@past.@entry, Ram(iof>person))

8 November 2003 PP attachment problem 10

Analysis in UNL

• Enconverter– Natural Language to UNL– Handles one sentence at a time– Predicate preserving parser– Kind of Turing machine

• Components– Dictionary : lexical units, uw, semantic attributes– Rule base : head movement rules, relation resolving rules

• Working – Uses dictionary and rule bases to process the sentence.

8 November 2003 PP attachment problem 11

Prepositional Phrase Attachment Problem

• Type of Structural ambiguity in a sentence

on new technologies.PP

the reportNP

readVP

JohnNP

Verb attachment

Noun attachment

8 November 2003 PP attachment problem 12

Prepositional Phrase Attachment Problem…

• Noun attachment Vs verb attachment

e.g. John read the report on new technologies. read

John the report

on

new technologies

read

John the report on

new technologies

*

8 November 2003 PP attachment problem 13

Establishing semantic relation

Same structure-different semantic relation

e.g.1. Ram ate rice with spoon. ……instrument

The UNL for this sentence is

ins(eat(icl>do).@past.@entry, spoon(icl>tool))

2. Ram ate rice with Sita. ……co-agentThe UNL for this sentence is

cag(eat(icl>do).@past.@entry, Sita(iof>person)) 

8 November 2003 PP attachment problem 14

Difficult problem

• PP attachment problem is simpler or no problem for human being

- who use world knowledge to process it.

• This world knowledge is not available to machines.

e.g. travel by night …time

travel by bus …instrument

8 November 2003 PP attachment problem 15

Different sites of attachment

– The search for the policy is going on.– The test will be held at the end of August.– In August 1947, India became free from British

rule.– Wilson received a medal from the commanding

officer at a farewell party.

• There is no restriction on how far the PP can lie from the word to which it relates.

8 November 2003 PP attachment problem 16

Affinity with preceding phrase

• The preposition of gets attached to a noun phrase or a verb phrase immediately preceding it.

– They were involved in the murder of a 90-year-old woman.– It was begun last week by the crew of a giant crane-barge.– He died of an overdose of sleeping pills– The system will be tailored to meet the need of the political

party.

8 November 2003 PP attachment problem 17

Existing methods

• generate mod-obj combination for almost all PP relations– E.g He came according to his promise.

agt(come(icl>do)@past.@entry, he)*mod(come(icl>do)@past.@entry, :01)obj:01(according to, promise(icl>abstract thing))mod:01(promise(icl>abstract thing),he)

• Tags introduced manually to resolve phrase boundaries – E.g. It delineates <p>the scope of phrases</p> before

<p>conversion of the sentence</p>.

8 November 2003 PP attachment problem 18

Related work

• Statistical learning methods used• Wordnet is used to find relations between words • Analysis of corpus is required• Not all aspects of problem considered

• The hypothesis does not apply to all cases

“PP attachments obey the principle of locality”

8 November 2003 PP attachment problem 19

Observations• Prepositions frequency is calculated from British

National Corpus• Classified into 2 parts

– Simple Preposition – Ambiguous prepositions

Frequency Preposition Poly. count

29391 of 718214 in 109343 to 814 by way of 116 by means of 1

8 November 2003 PP attachment problem 20

Addition to Semantic Attributes hierarchy

• Semantic attributes required to disambiguate • Addition required, if existing attributes fail to classify• necessary condition

– the attributes should be able to classify the semantically separate structures as separate entities.

e.g. the train for Delhi ….to()

the price for the Hill Road pool ….mod()

8 November 2003 PP attachment problem 21

Inclusion of preposition in UNL expression

• a picture on the wall plc(picture, wall).

• The cat walked across the street.– Wrong UNL

*plc ( walk, street )-cat walked along the street

-cat walked across the street – Correct UNLplc (walk, :01)obj:01(across, street)

8 November 2003 PP attachment problem 22

Classification based on syntax structure

• Sentences have different syntactic structure• Parsing the depends on surface structure

- Active-passive, transitive-di-transitive, present-past participles etc.

[ Verb + for + Noun phrase]v-pur He was waiting for the rainy day.v-pur He applied for a certificate.  [ Noun phrase + for + Noun phrase] n-mod The search for the policy is going on.n-mod He pays the price for his indulgence.

• Classification based on syntax pattern

8 November 2003 PP attachment problem 23

Classification based on semantics• Deciding factors

– Syntax, attributes, preposition, subcategorisation frame(for verbs)

Partial list of preposition on and its possible semantic relation

Relation Example sentenceON

plc a picture on a wallins to travel on the bustim He came on Sundayseq Report to reception on arrivalmod a book on South Africains She played a tune on her guitarplc You can get me on 0181 530 3906

8 November 2003 PP attachment problem 24

Updating rule base• Simpler if the classification is perfect.• Issues involved

– Priority, proper specification

Two rules showing difference in priority – specific to general

Comment ;N/abs for N/abs ;search for policy

delete preposition for

DL(N,ABS){PRE,#FOR:::}{N,ABS:+PRERES,+FORRES,+pPUR::}P25;

Comment ;V FOR N-UNIT-QUARES ;suspend for 2 days

Delete preposition for

DL(VRB){PRE,#FOR:::}{N,UNIT,TIM,QUARES:+PRERES,+FORRES,+pDUR::}P30;

8 November 2003 PP attachment problem 25

Conclusion

• World knowledge is realized in terms of semantic attributes.

• Phrasal verbs are not considered• Idiomatic constructs are not handled

- e.g. day by day

all the time