Semantic Data Understanding with Character Level Learning

Transcript

1. Semantic Data Understanding
   with Character Level Learning
   Michael Mior, Rochester Institute of Technology
   Ken Q. Pu, Ontario Tech University
   

   View Slide

2. ● Motivation
   ● Neural Network Architecture
   ● Similarity Graph Generation
   ● Semantic Analysis of DBpedia
   ● Conclusion
   

   View Slide

3. Motivation
   ● Assign semantic types to values in data lakes
   ● Types should capture semantic similarity
   ● Find anomalies in any database-assigned types
   ● Character-level reasoning without prior knowledge
   

   View Slide

4. Example
   Person
   Dorothy Vaughn
   Alan Turing
   Josef Burg
   Settlement
   Pittsburgh
   Edinburgh
   Johannesburg
   Salzburg
   

   View Slide

5. Example
   Person
   Dorothy Vaughn
   Alan Turing
   Josef Burg
   Settlement
   Pittsburgh
   Edinburgh
   Johannesburg
   Salzburg
   

   View Slide

6. DBpedia
   Pattuelli 2013
   

   View Slide

7. ● Motivation
   ● Neural Network Architecture
   ● Similarity Graph Generation
   ● Semantic Analysis of DBpedia
   ● Conclusion
   

   View Slide

8. Network Architecture
   ● Use a convolution neural network to identify salient substrings
   in values that can be used for classification
   ● Start with an embedding layer for dimensionality reduction
   ● Use convolutional layers to capture substring patterns
   ● End with fully connected layers
   

   View Slide

9. Network Architecture
   Pittsburgh 7155683942
   {
   {
   CNN
   Max pool
   Flatten
   Dense
   Softmax
   Settlement
   ⋮ ⋮ ⋮ ⋮ ⋮
   Embedding
   

   View Slide

10. Applications of Semantic Typing
    ● Semantic similarity between database types
    ● Identify a hierarchy among database types
    ● Detect outliers and dirty data
    

    View Slide

11. ● Motivation
    ● Neural Network Architecture
    ● Similarity Graph Generation
    ● Semantic Analysis of DBpedia
    ● Conclusion
    

    View Slide

12. Semantic Similarity Between Database Types
    Many database types are semantically similar.
    ChemicalSubstance ChemicalCompound
    Biomolecule Protein
    NaturalPlace BodyOfWater
    Semantic types can be used to detect and measure the
    semantic similarities between database types.
    

    View Slide

13. Confusion Based Similarity
    We hypothesize that semantic similarity is the main source of confusion of the
    neural network. Thus, we can infer the degree of semantic similarity between two
    database types t and t’ based on their mutual confusion.
    

    View Slide

14. Semantic Graph
    ● The pairwise semantic similarity measure allows us to organize all the
    database types into a similarity graph.
    ● A logistic function is used to rescale the similarity measure so we can control
    the spatial layout of the graph with the parameters: a and b.
    

    View Slide

15. Semantic Clustering
    ● Spectral clustering is a time-proven technique to analyze graphs.
    ● We apply spectral clustering to organize the database types into clusters
    which represent semantic topics in the database.
    ● Spectral clustering can be applied recursively so that a hierarchical
    organization can be obtained.
    

    View Slide

16. Anomaly Detection
    ● Type anomaly is a form of dirty data where data values that are assigned
    incorrect types.
    Example: “Beethoven’s 9th Symphony” being classified as Artist.
    ● We can utilize the neural network to detect likely candidates of such dirty
    data.
    ● Using substring reasoning, we can also generate visual explanation of the
    abnormalities of the identified data values.
    

    View Slide

17. Anomaly Detection
    Substring in the
    sliding window
    Semantic
    typing
    Assigned
    type
    Inferred
    type
    Substring location
    Certainty of inference
    “...phony...”
    “...hoven...”
    

    View Slide

18. ● Motivation
    ● Neural Network Architecture
    ● Similarity Graph Generation
    ● Semantic Analysis of DBpedia
    ● Conclusion
    

    View Slide

19. Semantic Clusters
    

    View Slide

20. Semantic Graph
    Biology
    Chemicals
    Events
    Agent
    

    View Slide

21. Semantic Clusters
    Subtype structure
    within a single cluster
    

    View Slide

22. Semantic Clusters
    

    View Slide

23. Semantic Clusters
    

    View Slide

24. Anomalies
    model response
    

    View Slide

25. Anomalies
    model response
    

    View Slide

26. Conclusion
    ● Semantic information is difficult to infer from data lakes
    ● Similarity information derived from neural networks can
    be used to construct a hierarchy of semantic types
    ● Semantic types are useful for analysis and data cleaning
    

    View Slide

27. Questions?
    

    View Slide