Inclusion of Temporal Semantics over Keyword-basedLinked Data Retrieval
Md-Mizanur Rahoman, Ryutaro Ichise
June 7, 2013
Outline
Introduction
Linked dataMotivationProblem and probable solution
Proposed Method
Query text processingSemantic query
Experiment
Conclusion
Md-Mizanur Rahoman, Ryutaro Ichise | 2
Introduction
Linked data
Represent knowledge using simpletechnique like subject, predicate, objectCan be presented by graph-like structureUse loose data publishing strategy
data publisher can publish data usingtheir own data schema
Can hold temporal feature related data
date, time or event related information
Iikka Paananen
music_artist
....
December,
29, 1960
birthDate
profession
Michael Jackson
Indiana
29th August
1958
birthDate
profession
deathDate
birthPlace
....
birthPlace
deathDate
2009-06-25
Md-Mizanur Rahoman, Ryutaro Ichise | 3
Motivation
Temporal feature influences linked dataretrieval
Various challenge among temporalfeature related information retrieval
Link data hold data heterogeneityLink data allow all kind of temporalfeature presentation strategies
Very few study over temporal featurerelated linked data information retrieval
Iikka Paananen
music_artist
....
birthDate
profession
Michael Jackson
IndianabirthDate
profession
deathDate
birthPlace
....
birthPlace
deathDate
2009-06-2529th August
1958
December,
29, 1960
Md-Mizanur Rahoman, Ryutaro Ichise | 4
Problem & Solution
Retrieval of temporal feature related information
Problem
Difficult in adaptation over linked data perspective
Solution
Convert all temporal features to a common formatAdapt a keyword-based linked data QA system [Rahoman, et al., 2012]for the formatted data
Md-Mizanur Rahoman, Ryutaro Ichise | 5
QA system [Rahoman, et al., 2012]
Take ordered keyword as inputUse some templates and tries to relate part of the linked dataset
Construct templates for each two adjacent keywordsMerge templates, if input keywords are more than two
Examplefor keywords: music artist and birth date
templates relation over dataset result
?
music_artist
birthDate
?
music_artist
?
birthDate
?
.. ..
Iikka Paananen
music_artist
....
birthDate
profession
Michael Jackson
IndianabirthDate
profession
deathDate
birthPlace
....
birthPlace
deathDate
2009-06-2529th August
1958
December,
29, 1960
Iikka Paananen ...Michael Jackson ...
Md-Mizanur Rahoman, Ryutaro Ichise | 6
Adaptation of temporal semantics
Assumption
Question hold temporal feature indicating word called signal word[Saquete, et al., 2009]Signal word prior keywords considered as question focus keywords(Q-FKS)Signal word follower keywords considered as question restrictionkeywords (Q-RKS)Example
question: music artist birth date on 29th December, 1960signal word: onQ-FKS: {music artist, birth date}Q-RKS: {on 29th December, 1960}
Md-Mizanur Rahoman, Ryutaro Ichise | 7
Temporal semantics over linked data
Proposed systemQuery text processing
Setup ordering key between Q-FKS and Q-RKS [Saquete, et al., 2009]and, then format Q-RKS related temporal feature to a common format(i.e., TIMEX3)
Semantic queryExecute QA system [Rahoman, et al., 2012], format temporal featurerelated output to TIMEX3 and then filter formatted output
Phase 1: Query Text Processing
Phase 2: Semantic Query
Final Output
Query Keywords
Ordering Key Generator
QA System
with Time Filter
Q-FKS, Ordering Key,
Formatted Time
Time Formatter
Q-FKS, Ordering Key, Q-RKS
Md-Mizanur Rahoman, Ryutaro Ichise | 8
Ordering key
Order temporal feature attachment between Q-FKS and Q-RKSaccording to the signal word [Saquete, et al., 2009]
Signal word Ordering keyIn/On Q-FKS = Q-RKSAfter Q-FKS > Q-RKSBefore Q-FKS < Q-RKS... ...
Help filtering Q-FKS output by restricting temporal feature ofQ-RKS
Example
Input keywords: music artist, birth date, on 29th December, 1960Signal word: onQ-FKS: {music artist, birth date}Q-RKS: {on 29th December, 1960}Ordering key: Q-FKS = Q-RKS
Md-Mizanur Rahoman, Ryutaro Ichise | 9
Query text processing
Ordering key generatorDivide input keywords according to signal wordPut ordering key between Q-FKS and Q-RKS
Time formatterUse Stanford parser [Chang, et al., 2012] over Q-RKS and generateTIMEX3 formatted temporal values
Phase 1: Query Text Processing
Phase 2: Semantic Query
Final Output
Query Keywords
Ordering Key Generator
QA System
with Time Filter
Q-FKS, Ordering Key,
Formatted Time
Time Formatter
Q-FKS, Ordering Key, Q-RKS
Md-Mizanur Rahoman, Ryutaro Ichise | 10
Semantic query
QA system with time filterExtract temporal feature related outputConsist 3 steps:
Step 1: Execute QA system over Q-FKSStep 2: Format part of temporal feature related output to TIMEX3Step 3: Filter Q-FKS related formatted output considering ordering keyand TIMEX3 value of Q-RKS
Phase 1: Query Text Processing
Phase 2: Semantic Query
Final Output
Query Keywords
Ordering Key Generator
QA System
with Time Filter
Q-FKS, Ordering Key,
Formatted Time
Time Formatter
Q-FKS, Ordering Key, Q-RKS
Md-Mizanur Rahoman, Ryutaro Ichise | 11
Semantic query
Example
Q-FKS: {music artist, birth date}Q-RKS: {on 29th December, 1960}Formatted value of Q-RKS: {1960-12-29}Ordering key: {Q-FKS = Q-RKS}
Execution of QA system with time filter
Step Result1 Iikka Paananen ... December,29,1960
Michael Jackson ... 29th August,19582 Iikka Paananen ... 1960-12-29
Michael Jackson ... 1958-08-293 Iikka Paananen ... 1960-12-29
Md-Mizanur Rahoman, Ryutaro Ichise | 12
Experiment
Experimental Data
Question Answering over Linked Data 1 (QALD-1) open challengedata
Consist natural language questionsSorted out for questions which relate temporal feature in answering
DBPedia test case - 4 questionsMusicBrainz test case - 18 questions
Input
Ordered input keywords (with singal word)
Md-Mizanur Rahoman, Ryutaro Ichise | 13
Performance of proposed QA system
Check average recall, average precision and average F1 measure foreach participant dataset
Participant # of Performance of proposed systemquestion set questions
Recall Precession F1 MeasureDBPedia QALD-1 4 1.000 1.000 1.000MusicBrainz QALD-1 18 0.765 0.765 0.765
DBPedia dataset achieve gold-standardPerformance drop over MusicBrainz dataset
QA system not able to generate Q-FKS related information
Successfully adapts
signal keyword, ordering key and Stanford parser over temporal featurerelated linked data information retrieval
Md-Mizanur Rahoman, Ryutaro Ichise | 14
Conclusion
Our study
Adapt temporal semantics over an keyword-based linked data retrievalframeworkReduce data heteroginity by converting all temporal value to acommon formatShow implementation result for real linked implementation
Future work
Want to introduce more shophisticated keyword matching sincecurrent system depends on exact matching
Md-Mizanur Rahoman, Ryutaro Ichise | 15