An Interactive Approach to Collectively Resolving URI Coreference

An Interactive Approach to Collectively Resolving URI Coreference

ws .nju.edu.cn An Interactive Approach to Collectively Resolving URI Coreference Saisai Gong, Wei Hu, Gong Cheng, Yuzhong Qu [email protected] Contents ws .nju.edu.cn Background Related Work Overview of our Approach

Evolvement of Individual Partition Computing Consensus Partition Evaluation Conclusion [email protected] 2 Background ws .nju.edu.cn owl:sameAs http://advogato.org/person/timbl/foaf.rdf#me

http://www.w3.org/People/Berners-Lee/card#i URICoreference http://dblp.l3s.de/d2r/resource/authors/Tim_Berners-Lee [email protected] http://data.semanticweb.org/person/tim-berners-lee http://dbpedia.org/resource/Tim_Berners-Lee 3 Related Work

ws .nju.edu.cn Fully automatic approaches OWL semantics Similarities between descriptions Self training Automatic approaches remain far from prefect (see F errara et al. 2013 ) [email protected] 4

Related Work (Cont.) ws .nju.edu.cn Semi-automatic approaches Active learning Micro-task crowdsourcing Assumptions made by semi-automatic approaches Users act as oracle One single right answer Not always hold Users may have different opinions Disagreement among users happen

Distinguish a user's individual URI coreference from the m ass Resolve disagreement among users [email protected] 5 Our Approach iReC ws .nju.edu.cn iReC: an interactive approach to resolve collectively URI coreference with user involvement Basic idea: achieve a good partition of the URI univer se Maintain individual partition for each user

Form consensus partition aggregated from individual ones Evolve partitions through user interaction Two goals Alleviate user involvement Reflect the collective power of masses [email protected] 6 Overview of our Approach [email protected] ws .nju.edu.cn

7 Candidate Selector ws .nju.edu.cn Generating Candidates Find potential coreference from various sources owl:sameAs links existing resolution services such as sameas.org, keyword-based entity search engines such as Falcons Object Searc h the user's individual partition the consensus partition

Merge URIs belonging to the same equivalent class into a c andidate entity [email protected] 8 Learning Binary Classifier ws .nju.edu.cn To reduce user involvement Learning model: averaged perceptron (See Collins 02) Online learning algorithm

y sign ( w x b ), y {1, 1} Learn individual classifier both online and offline, lea rn global one offline [email protected] 9 Learning Binary Classifier ws .nju.edu.cn Training data Online : latest URI pairs from user feedback Offline training examples

Positive : URIs pairs from equivalent classes Negative URI pairs from user feedback URI pairs from different equivalent classes sharing types URI pairs Falcons search result [email protected] 10 Learning Binary Classifier ws .nju.edu.cn Training algorithm

x Feature : the cartesian product of the two candidates' pr operties Feature value: for each property pair, compute maximum si milarity of the given two properties values URIs: vsim=1 iff identical or in equivalent class Numeric literals: vsim=1 iff difference less than threshold Boolean literals: vsim=1 iff value equal Other literals: Jaccard similarity [email protected]

11 Learning Binary Classifier ws .nju.edu.cn Training algorithm [email protected] 12 Selecting Most Beneficial Candidate ws .nju.edu.cn

Combine individual classifier and global one by their weights (_+ = 1) Confidence of coreference based on margin w x b The larger the absolute value of margin is, the higher the co nfidence is Uncertainty: the absolute value of margin Select candidate with minimum absolute value of mar gin [email protected]

13 Comparative Snippets ws .nju.edu.cn To facilitate user interaction Coreferent (non-coreferent resp.): values of discrimin ative property pairs signicantly similar (dissimilar res p.) Discriminability of property pairs: absolute values of weight in combined classifier [email protected] 14

Comparative Snippets ws .nju.edu.cn Compute maximum weighted matching on the biparti te graph from property pairs Get topk property value pairs based on maximum sim ilarity of property values [email protected] 15 Computing Consensus Partition

ws .nju.edu.cn Minimize disagreements between individual partition s In our approach, using symmetric difference distance Maximizing NP-complete [email protected] 16 Computing Consensus Partition ws .nju.edu.cn

Approximation algorithm clustering-based Compute a partition on the union of individual partitions d omains first initialize a similarity matrix Mtrx=( ij ) begin with each URI forming an equivalence class separate ly for each class pair (i, j) , where ij > 0, merge together class es i,j , and update Mtrx ij [email protected] 17

Computing Consensus Partition [email protected] ws .nju.edu.cn 18 Evaluation ws .nju.edu.cn Build link between NYT and Dbpedia of OAEI benc hmark 10 fold cross validation

[email protected] 19 Evaluation ws .nju.edu.cn F-Measure [email protected] 20 Evaluation

ws .nju.edu.cn Examination Choose 50 popular URIs from falcons Invite 10 people to resolve URIcoreference on the 50 URIs using SView In average, 290.1 times verification, 32.0 accepted as positive 53.9 pair of URIs in individual partitions [email protected] 21

Evaluation ws .nju.edu.cn User study SUS Vs sigma 72 vs 68 [email protected] 22 Conclusion ws .nju.edu.cn Averaged Perceptron is feasible

User involvement is reduced [email protected] 23 Reference ws .nju.edu.cn A. Ferrara, A. Nikolov, J. Noessner, and F. Schare. Evaluation of instance matching tools: the experience of OAEI. Journal of Web Semantics, 21:49-60, 2013. M. Collins. Discriminative training methods for hidden markov models: theory and experiments with perceptron algorithms. In Proc. of EMNLP, pages 1-8, 2002. [email protected]

24

Recently Viewed Presentations

  • Informal document GRVA-02-09 2nd GRVA, 28 January  1

    Informal document GRVA-02-09 2nd GRVA, 28 January 1

    Furthermore, classical certification elements (track testing) are an essential part of the three-pillar approach (see from slide 14). Additions are needed to appropriately cover the software related aspects - they will augment and not replace the classical certification approach.
  • Welcome to the Simons Workshop on Learning, Algorithm Design ...

    Welcome to the Simons Workshop on Learning, Algorithm Design ...

    Welcome to the Simons Workshop onLearning, Algorithm Design and Beyond Worst-Case Analysis. November 14-18, 2016. Organizers: NirAilon, Nina Balcan, Avrim Blum, Ravi Kumar, Kevin Leyton-Brown, and Tim Roughgarden
  •  ENDODONTICS Assist prof. Dr.Emad Farhan Alkhalidi B.D.S., M.Sc.,

    ENDODONTICS Assist prof. Dr.Emad Farhan Alkhalidi B.D.S., M.Sc.,

    The buccal cusp tip lies over the central axis of the tooth so the oval - shaped cavity should lie as far buccally as possible encroaching towards the tip of the buccal cusp . If the tooth is to be...
  • OP U.S USFOR USFOR Section 886 of NDAA

    OP U.S USFOR USFOR Section 886 of NDAA

    Authority under the National Defense Authorization Act of 2008 that allows . procurements using other than full and open competition for Acquisitions . in support of Operations in Iraq and Afghanistan. Applies to products, services and construction. Justification & Approval...
  • Inorganic Chemistry Lecture PowerPoint - Science Prof Online

    Inorganic Chemistry Lecture PowerPoint - Science Prof Online

    "What Kind of Bonds Are These?" song and slide show by Mark Rosengarten. Chemical Bond Formation animated science tutorial. " Meet the Elements " music video by They Might Be Giants. Redox Reactions video lectureby Kahnacademy. Chem4Kids website by Rader....
  • Deep Brain stimulation for the Treatment of Chronic Pain

    Deep Brain stimulation for the Treatment of Chronic Pain

    Frontal Systems Behaviour Scale. The FrSBeis a 46-item behavior rating scale that was developed as a measure of behavior associated with damage to the frontal systems of the brain. Assesses behaviours associated with apathy, executive dysfunction and disinhibition
  • Colorado 2008

    Colorado 2008

    For 4 Gy radiation, the inhibition of cell proliferation was 26% and 47% at 3 and 6 days post radiation with respect to control non-radiated cells. For example, for a 8 Gy radiation dose, the inhibition of cell proliferation was...
  • NEW ZEALAND | Investment Briefing 2016 IS THE

    NEW ZEALAND | Investment Briefing 2016 IS THE

    but only the hedge funds were able to take advantage of that. A simplified exampleso how did it all work out? The active long-only managers were 0.4% overweight ABC on average. So as a group they lost 0.2% on the...