Transcript Towards Semantics Generation

Towards Semantics Generation
Third stage presentation of M.S project
Ashish Almeida
03M05601
Guide
Prof. Pushpak Bhattacharyya
7/8/2015
1
Motivation
• Goal: semantic role labeling
• To commonly used functional element in
English. (34% (source: Penn tree-bank))
• To act as both preposition and as infinitival
marker.
• PRO was not considered before in
semantic labeling
7/8/2015
2
Roadmap
•
•
•
•
•
•
•
Problem
UNL*
Linguistic analysis
Attachment solution
Dictionary creation
Implementation
Conclusion
7/8/2015
3
Current work (third stage)
•
•
•
•
Organization of attributes
Analysis of to-infinitive
PRO-handling and resolution
Acquisition of attributes for dictionary
7/8/2015
4
Problem
• Semantics generation for sentences involving
lexeme to
• Three problems
– Identifying the proper part of speech (POS)
– Attachment ambiguity resolution
– Handling PRO
• Focus
Only [V-N-to-N/V] frames considered.
Document specific dictionary used
7/8/2015
5
UNL*
give(icl>do)
agt
John(icl>person
)
obj
@entry.@past
gol
Mary(iof>person)
flower(icl>flora)
• UNL
• UWs
• Relations
7/8/2015
6
Differentiating POS
• Identify to-preposition phrase from toinfinitival clause
• … gave papers to the judge
- to is followed by a determiner
• … increases to 25 million rupees
- to is followed by a number
• … to cooks.
- to is followed by a plural noun
7/8/2015
7
Differentiating POS
… to-infinitival
• …to go…
- to is followed by a base verb
• … to clearly write…
- to is followed by adverb followed by base verb.
7/8/2015
8
Attachment algorithm
For Prepositional phrases
7/8/2015
9
Example
• John gave a flower to Mary.
– Verb gave expects to
– Noun flower does not expect to
– Apply case 3
– Attach ‘to Mary’ to gave
• Final UNL:
7/8/2015
10
To infinitival clauses
• Example
1a. He promised me [to come for the party].
1b. Hei promised me [PROi to come for the party].
promise  subject controlled pro
2a. They forced Mary [to give a party].
2b. They forced Maryj [PROj to give a party].
force  object controlled pro
7/8/2015
11
UNL representation
Theyi promised Mary [PROi to give a party].
7/8/2015
12
Attachment algorithm table
for to-infinitival clauses
7/8/2015
13
PRO resolution
Example
a. He ordered us [to finish the work].
b. He ordered usi [PROi to finish the work].
Steps
1. fetch PRO type fom dictionary entry of order
2. Resolve all relations within clause
- [PROi to finish the work]
3.
4.
Relate the clause to verb order
Finally replace the PRO with actual UW
7/8/2015
14
Semantic relations
• Filled using the Levin’s verb classes.
• No semantically role resource available
• Stored in dictionary along with argument
information
7/8/2015
15
Sentence having to
System
Detect part of speech
To-infinitive
To-preposition
Find attachment site
Find attachment site
Decide type and existence of PRO
Resolve pro
Find semantic relation
Find semantic relation
Coindex the PRO
7/8/2015
UNL expressions
16
Dictionary
• All words must be present in dictionary
• Structure
[letter] “letter(icl>document)” (N,INANI,PHSCL) <E,0,0>
headword
7/8/2015
Universal word
Attributes
17
Dictionary: Acquisition of attributes
New attribute needed to apply the algorithm
• Argument structure information
• Semantic relations
• PRO control property of verbs
• Oxford, WordNet
• Penn Treebank
• Beth Levin’s verb classification
7/8/2015
18
from WordNet
• Sentence frames for verbs
• Example
• For verb want
– They ____ him to write the letter.
For the verb promise
– Somebody ----s somebody to INFINITIVE
7/8/2015
19
from Oxford dictionary
• Oxford advanced learners dictionary (OALD)
provides partial frames wherever applicable
• Examples
effort noun
…… 2 [C] ~ (to do sth) an attempt to do sth especially when it is
difficult to do: to make a determined / real / special effort to finish
on time …..
force verb
make sb do sth 1 [often passive] ~ sb (into sth / into doing sth) to
make sb do sth that they do not want to do SYN COMPEL …•
[VN to inf] I was forced to take a taxi because the last bus had
left. • She forced herself to be polite to them. …
7/8/2015
20
from Penn Treebank
• Syntactically annotated corpus
• Example
• Algorithm to extract this property
7/8/2015
21
Organizing attributes
• WordNet noun ontology explored.
• The top level labels used as attributes.
• Example:
7/8/2015
22
English to UNL system
Partial UNL
expression
Input
sentence
WordNet
OALD
UNL
expression
Penn tree-bank
• Rule base
7/8/2015
23
Implementation
•
•
•
•
POS Identification
Finding Attachment site
Creating Relation
PRO insertion
• Post processing
– Resolve the co-reference.
7/8/2015
24
Identification of POS
Pattern to detect to infinitive:
-to followed by verb in base form
:{:::}{^TO_INF_NEXT:+TO_INF_NEXT::}(#TO,TO_INF)(BLK)(VRB,V_1)P40;
IF (The left analysis window (indicated by {}) is on any word)
AND
(The right analysis window is on a word which does not have a TO_INF_NEXT i.e.
look ahead is not performed yet. )
THEN
Select the next sequence of words such that they will satisfy the conditions as –
pick the word to corresponding to infinitival-to (indicated by attributes #TO and
TO_INF)
AND pick a space (indicated as BLK)
AND pick a verb which is in its simple form (indicated by V_1)
AND add the property TO_INF_NEXT to the word in the right analysis window
7/8/2015
25
Attachment rules
• Do noun attachment
– Move ahead when on frame [V][N]-P-N
R{VRB,#_TO_AR2:::}{N,#_TO:::}(PRE,#TO)P60;
• Create goal relation
– gol(uw1, uw2)
<{VRB,#_TO_AR2,#_TO_AR2_gol:::}{N,TORES,PRERES::gol:}P25;
7/8/2015
26
Handling PRO
1. Produce a “PRO” element in UNL with appropriate
relation. (Enconverter)
:{VRB,SUB_PRO:::}"[[SUB_PRO]]:N,SUB_PRO,
#INSERTED::"(VRB,TO_INFRES,^PRORES)P30;
2. Relate it to the verb of the infinitive clause
semantically. (Enconverter)
>(VRB){N,SUB_PRO::agt:}{VRB,VOA,TO_INFRES:
+PRORES,+SUB_PRORES::}P40;
3. Substitute a referred UW in the place of PRO.
(Post editor)
7/8/2015
27
Replace PRO
Example
They promised Maryi [PROi to give a party].
agt (promise(icl>do).@entry.@past, they:0A)
gol (promise(icl>do).@entry.@past, Mary(iof>person))
obj (promise(icl>do), :01)
agt :01(give(icl>do), sub_PRO:0C)
obj :01(give(icl>do), party(icl>function))
After post processing
agt :01(give(icl>do), they:0A)
7/8/2015
28
Evaluation
• Preparation of test sentences
• Source : Penn Treebank, edict
concordencer and Oxford
• Dictionary
– Automatic dictionary generator
– Editing and corrections
– Appending extra attributes.
7/8/2015
29
Results
To
Total number of sentences(200)
Preposition Infinitive
sense
sense
100
100
Number of sentences where correct sense of to is
detected
97
93
Number of sentences with correct attachment/UNL
80
72
7/8/2015
30
Conclusion
• Automatic acquisition of attributes is
effective.
• Correct Semantic representation is crucial.
– Helps in applications like information retrival,
generation to other language, question
answering
7/8/2015
31
References
• Grimshaw, Jane: Argument Structure. The MIT Press, Cambridge,
Mass. (1990)
• Mohanty R.K., Almeida A., Srinivas S., Bhattacharyaa P.: The
complexity of OF, ICON, Hydrabad, India. (2004)
• UNDL Foundation: The Universal Networking Language (UNL)
specifications version 3.2. (2003) http://www.unlc.undl.org
• Resources
–
–
–
–
–
–
7/8/2015
OALD
WordNet
Penn Tree bank
DDG
Concordance search on Brown corpus
Beth Levin’s verb classes
32
! Thank you
7/8/2015
33