Presentation is loading. Please wait.

Presentation is loading. Please wait.

Websoft Research Group

Similar presentations


Presentation on theme: "Websoft Research Group"— Presentation transcript:

1 Websoft Research Group
ESWC 2015论文阅读报告(两篇) 报告人:仇宏磊 Websoft Research Group

2 Requirements for and Evaluation of User Support for Large-scale Ontology Alignment
Valentina Ivanova1,2, Patrick Lambrix1,2(mail), and Johan Åberg1 1 Department of Computer and Information Science, Linköping University, Linköping, Sweden 2 The Swedish e-Science Research Centre, Linköping University, Linköping, Sweden Websoft Research Group

3 Websoft Research Group
Outline Motivation Related Work Requirements for User Support Literature Study User Interface Evaluations Websoft Research Group

4 Websoft Research Group
Motivation Short of user interface. Websoft Research Group

5 Websoft Research Group
Related Work Websoft Research Group

6 Websoft Research Group
Requirements for User Support Requirements for users managing larger ontologies Extracted from existing works and systems Extracted from authors’ personal experience Interaction & Representation Manipulation Inspection Explanatory #X.X – No. on the previous page and #5.X is listed on the next page Websoft Research Group

7 Websoft Research Group
Requirements for User Support #5.X #5.5 Environment #5.1 Interruption of the alignment process #5.6 Recommendations #5.7 A debugging step during the alignment process #5.2 Divide large-scale tasks into smaller tasks #5.8 Configuring the matching process #5.3 Reduce unnecessary user interventions #5.9.1 Trial execution of mappings #5.4 Social and collaborative matching #5.9.2 Support for temporary decisions Websoft Research Group

8 Websoft Research Group
Literature Study Well supported The first four systems Clustering of the content Clustering during computation Level: need a user study Defects Some cannot distinguish mapping and mapping suggestions Reasons assists users in accepting mapping suggestions Social and collaborative matching Trial execution of mappings Websoft Research Group

9 Websoft Research Group
User Interface Evaluation CogZ, COMA 3.0 and SAMBO A heuristic evaluation Usability issues Nielsen’s ten heuristics An observational study 3 master and 5 PhD (7 male, 1 female) Each perform 11 and 17 tasks. 2 sessions lasted for 2h and 1h. OAEI 2014-AMA (2737 concepts, 1807 asserted is-a relations) NCI-A (3298 concepts, 3761 asserted is-a relations) Websoft Research Group

10 Websoft Research Group
User Interface Evaluation Number of participants (max 8) successfully completed a task. / Average task time per system in seconds. Websoft Research Group

11 Semi-supervised Instance Matching Using Boosted Classifiers
Mayank Kejriwal(mail) and Daniel P. Miranker University of Texas at Austin, Austin, USA Websoft Research Group

12 Websoft Research Group
Outline Background Core idea Approach Experiment Websoft Research Group

13 Websoft Research Group
Background Blocking and Classification Blocking: place instances into clusters Classification: pair and classify instances Supervised systems Good performance Heavy labeling effort Minimally supervised Less training data Poor performance Parameter tuning Websoft Research Group

14 Websoft Research Group
Core idea Boost Adaboost, ensemble classifier Random Forest VS Multilayer Perception Probabilistic instance matching Iterative semi-supervised learning Classification quality is certainly poor Choose the most confidently labeled pairs Iteratively self-train Websoft Research Group

15 Websoft Research Group
Approach Websoft Research Group

16 Websoft Research Group
Approach Pre-classification Steps Blocking approach: trigrams-based Attribute Clustering Generating Restriction Sets: sets of class and property alignments (Q) Extract Features: FEBRL package (G) 2 numeric features 8 string and token-based features, e.g. Common Token 18 phonetic features Feature vector: |Q||G| Classification Step Parameters Iteration number = 7 factor D & factor N = 2 Websoft Research Group

17 Websoft Research Group
Experiment Data Persons1, Persons2, Restaurants from IAEI Real-world benchmark: ACM-DBLP, Amazon-GoogleProducts, Abt-Buy 1k-2k matching pairs Pre-classification Results Perfect for IAEI benchmarks Real-world: 95.44%, 97.43%, 83.54% F-measure, re-weight F-measure Training data 2% or 50 (whichever is less) Websoft Research Group

18 Websoft Research Group
Experiment Random Forest Websoft Research Group

19 Websoft Research Group
Experiment Multilayer Perceptron Websoft Research Group

20 Websoft Research Group
Experiment Comparison to Minimally Supervised Approaches Comparison to Supervised Approaches 10-fold cross: 79.25% 100 samples, FEBRL & Marlin Amazon-GoogleProducts:30-40%, 50%; Abt-Buy, 20%, 60% Websoft Research Group

21 Websoft Research Group
Reference Ivanova V, Lambrix P, Åberg J. Requirements for and Evaluation of User Support for Large-Scale Ontology Alignment[M]//The Semantic Web. Latest Advances and New Domains. Springer International Publishing, 2015: 3-20. Kejriwal M, Miranker D P. Semi-supervised Instance Matching Using Boosted Classifiers[M]//The Semantic Web. Latest Advances and New Domains. Springer International Publishing, 2015: Falconer S M, Storey M A. A cognitive support framework for ontology mapping[M]. Springer Berlin Heidelberg, 2007. Papadakis G, Ioannou E, Palpanas T, et al. A blocking framework for entity resolution in highly heterogeneous information spaces[J]. Knowledge and Data Engineering, IEEE Transactions on, 2013, 25(12): Websoft Research Group

22 Websoft Research Group
THANKS Websoft Research Group


Download ppt "Websoft Research Group"

Similar presentations


Ads by Google