Transcript Preposition Phrase Attachment in English Language Analysis

Preposition Phrase
Attachment in English
Language Analysis
Ashish Almeida
03M05601
PP attachment

John read the report on new technologies.
read
John
*
the
report
John
on
the
report
on
new
technologies
new
technologies
PP attached to VP
PP attached to NP
July 17, 2015
read
CFILT,IIT Bombay
2
Same Structure: different roles

Ram ate rice with a spoon.
-instrument
ins(eat(icl>do).@past.@entry, spoon(icl>tool))

Ram ate rice with Sita.
-co-agent
cag(eat(icl>do).@past.@entry, Sita(iof>person))
July 17, 2015
CFILT,IIT Bombay
3
UNL and EnConvertor





UNL is an intermediate language for
representing meaning of natural language.
EnConvertor is a language independent parser
Rules are written for analysis
UW Dictionary is created for analysis
PP-attachment is handled in EnCo.
July 17, 2015
CFILT,IIT Bombay
4
More about UNL
UNL graph of John eats rice with a spoon
eat(icl>do)
@ entry.
@ present
ins
obj
agt
spoon(icl>artifact
)
rice(icl>food
)
John(iof>person)
July 17, 2015
CFILT,IIT Bombay
5
Adjuncts


It provides extra information in a sentence.
It attaches to the verb.
Examples:
Ram came home.
Ram came home on Monday.
 Sita is sleeping.
Sita is sleeping in the room.

July 17, 2015
CFILT,IIT Bombay
6
Temporal Prepositional Phrases

Represents temporal information with the help of an object
e.g. on Sunday , for two days, at 5 p.m., on Diwali,
in ice age, after the meeting, till 6 O’ clock,
over two hours, beyond midnight
Prepositions that can show temporality
at
over
through
in
from during
on after between
July 17, 2015
before
for
since
by
beyond
till
until
to
inside
into
within
of
CFILT,IIT Bombay
7
Time attributes
To identify the type of time word
e.g.

Attribute
Examples
TIME
Today, March
UNIT
Day, second
EVENT
Morning, Dinner
TIM_TOKEN
am, pm
SECOND*
12, 15
DAY*
Sunday; 23
POFDAY
Dinner, Sunset
WEEK
Week
ZDIM
Start, End, Middle
July 17, 2015
CFILT,IIT Bombay
8
UNL of Temporal PPs
Different UNL generation in two different cases
1. Delete the preposition
e.g. come at noon
2.
tim(come , noon)
Retain the preposition
e.g. come before noon
tim(come, before)
obj(before,noon)
July 17, 2015
CFILT,IIT Bombay
9
Mapping from Prepositions to UNL
relations
Preposition
UNL Relations
Attributes of the Arguments
at
tim
[TIME,TIM_TOKEN]
in
tim
[N,TIME,MONTH]
[N,TIME,YEAR]
[N,POF_DAY]
on
after
tim
tim-obj
[N,TIME,DAY]
before
tim-obj
[TIME,TIM_TOKEN]
[N,TIME]
[N,EVENT]
July 17, 2015
[TIME,TIM_TOKEN]
[N,TIME]
[N,EVENT]
CFILT,IIT Bombay
10
Rules

Rules for at-PP
- Applies to “at 6 pm”
;delete at
DL(VRB){PRE,#AT:::}{TIME,TIM_TOKEN:
+ATRES,+PRERES,+pTIM::}P22;
;create relation tim
<{VRB:::}{ATRES,PRERES,pTIM::tim:}P20;
July 17, 2015
CFILT,IIT Bombay
11
Testing




Wall Street Journal (WSJ) corpus is used.
Sentences from Oxford advanced learner’s
Dictionary are also tested.
WSJ has V-N-P-N four-word sentence
fragments
All temporal cases are tested for correctness of
UNL
July 17, 2015
CFILT,IIT Bombay
12
Results
#Total PPs
20801
#Temporal PPs
1326
#Cases of correct
UNL
1112
Average accuracy
83.9 %
•Errors are mainly due to
mistakes in corpus and
inaccuracies in UW
dictionary
July 17, 2015
CFILT,IIT Bombay
13
Associative of-PP

NP: the comedy of Shakespeare
mod(comedy(icl>abstract thing), Shakespeare(iof >person))

NP: the eyes of the boy
pof(boy(icl>person), eye(pof>body).@pl)

AP: guilty of an offence
obj(guilty(aoj>thing), offence(icl>abstract thing))

NP: book of Ram
pos(book(icl>concrete thing), Ram(iof>person))
July 17, 2015
CFILT,IIT Bombay
14
Partitive of-PP

Here, in case of N1-OF-N2 the semantic head is the N2.
The first NP indicates a quantity
Examples
1.
2.
3.
4.

cup of tea
bag of oranges
bundle of sticks
a pinch of salt
qua(tea, cup)
Kind construction
e.g. that kind of people mod(people, kind)

Kind-type words : kind, type, sort, variety etc.
July 17, 2015
CFILT,IIT Bombay
15
Argument structure (AS)

Argument structure specify the structural frame into
which a verb can be fitted.
For example,
*Ram saw.
This is unacceptable as verb see has strict
subcategorisation feature (+ _NP). That is verb see
takes NP as object. Thus a valid sentence is
Ram saw Sita.

AS of see is (NP _ NP)
July 17, 2015
CFILT,IIT Bombay
16
Adjunct and Complements
1.
He gave a book to Ram.
give (NP _ NP to-PP)
- without to Ram sentence is unacceptable
- to Ram is complement
2.
He gave a book to Ram on Sunday
- without on Sunday sentence is acceptable
- on Sunday is adjunct

Similarly, nouns and adjectives take complements
July 17, 2015
CFILT,IIT Bombay
17
More on AS

Ram accused Sita of cheating

AS is (NP _ NP of-PP)
UNL for the sentence frame with the verb accuse
(agtNP _ objNP rsnof-PP)

Whereas for adjuncts, the case relations differ.

1.
2.
3.
July 17, 2015
He saw the girl through the window.
He saw the girl with anger.
He saw the girl in the library.
CFILT,IIT Bombay
18
Dictionary entry



He gave a book to Ram.
The lexicon will have entry of gave which provides AS
information.
[gave] {} “give(icl>do)” (VRB,VOA,VOA-PHSL,
#_TO, #_TO_GOL,PAST) <E,0,0>;
PP in the sentence can be identified as a complement or not.
This solves the attachment in some cases.
July 17, 2015
CFILT,IIT Bombay
19
Verb attachment of ‘of-PP’

Three possible cases
NP1
NP2
V attaches to NP1
V attaches to NP2
(A)
July 17, 2015
V
V
V
NP1
NP2
V attaches to NP1
NP1 attaches to NP2
(B)
CFILT,IIT Bombay
NP1
NP2
V attaches to NP2
NP2 attaches to NP1
(C)
20
of-PP attachment cases
... remind him of Gita
Case A
... saw the book of physics
Case C
July 17, 2015
CFILT,IIT Bombay
Case B
... drank a cup of milk
21
Four cases of attachment

Attachment in presence or absence of attribute of
1
2
3
4
Attributes of
V
V,OF
V,^OF
V,OF
V,^OF
Attributes of
NP1
N,OF
N,OF
N,^OF
N,^OF
Attachment of
NP2
NP1
NP1
V
NP1
•
July 17, 2015
CFILT,IIT Bombay
22
Rules
;Noun attachment
R{VRB,#_OF:::}{N,#_OF:::}(PRE,#OF)P60;
;Noun attachment
R{VRB,^#_OF:::}{N:::}(PRE,#OF)P60;
;Verb attachment, first resolve the immediate object
<{VRB,#_OF,#_OF_OBJ:::}{N,^#_OF::obj:}
(PRE,#OF)P30;
July 17, 2015
CFILT,IIT Bombay
23
Testing process




British National Corpus (BNC) is used.
V-N-of-N type sentences are extracted from
corpus.
AS information is added to verbs and nouns
using Oxford Advanced Learner’s Dictionary
and Beth Levin’s verb classes.
AS information is merged into the Dictionary.
July 17, 2015
CFILT,IIT Bombay
24
Results
Total
Correct
Incorrect
V-attachment
N-attachment (all)
7
493
7
439
0
54
Associative cases
411
362
49
Partitive cases
Total
82 (16%) 77
500
446
5
54
July 17, 2015
CFILT,IIT Bombay
25
To sum up




The PP attachment problem is successfully divided in
two parts – complement PPs and adjunct PPs
This division clearly delineates the problem so that their
analysis does not conflict.
Analysis of complements will be mostly driven by rich
attribute set from lexicon
Whereas analysis of adjuncts will be driven by rules
July 17, 2015
CFILT,IIT Bombay
26
Contribution






Deciding the overall strategy of analysis
Providing computation insights in linguistic
analysis
Design of attribute set/introducing new
attributes
Design and implementation of rules for EnCo.
Testing of sentences.
Dictionary corrections/modifications
July 17, 2015
CFILT,IIT Bombay
27
Future work



Focus will be on complements.
Also, that-clause, to and -ing infinite clause
handling, PRO detection will be tried in similar
fashion.
Automatic/semi-automatic acquisition of AS
information for dictionaries will be tried out.
July 17, 2015
CFILT,IIT Bombay
28