Keyword Search over RDF Graphs. Elisa Menendez

Transcription

1 Elisa Menendez

2 Summary Motivation Keyword Search over RDF Process Challenges Example QUIOW System Next Steps

3 Motivation

4 Motivation Keyword search is an easy way to retrieve information Whether in Web pages or in databases

5 Motivation Keyword search is useful for databases It eliminates the overhead of writing structured queries The user does not need to know the database schema SELECT movie.title FROM movie, director WHERE movie.director = director.id AND director.name = 'Tim Burton'

6 Motivation From the Semantic Web and Linked Data RDF graphs emerged as a new kind of databases That models data as triples nflx:beetlejuice mo:directedby nflx:tim_burton subject predicate object PREFIX mo: < PREFIX nflx: <

7 Motivation A collection of RDF triples forms an RDF graph Beetlejuice mo:title mo:releaseyear nflx:beetlejuice rdf mo:movie 1988 mo:synopsis mo:directedby A couple of ghosts contract nflx:tim_burton rdf mo:director

8 Motivation RDF graphs are a type of databases LOD Cloud datasets

9 Motivation RDF graphs are interesting sources of knowledge Give more flexibility to describe resources Use W3C standardized graph formats Use popular ontologies Keep increasing

10 Motivation RDF graphs are usually queried with SPARQL queries SELECT?title WHERE{?movie rdf :Movie.?movie mo:directedby?director.?director "Tim Burton".?movie mo:title?title } => Keyword search is also useful for RDF Graphs

11 Motivation Advantages of RDF Graphs for Keyword Search Well-known graph theories Data and metadata all together And easy to access with SPARQL Integration of multiple sources And easy to access with SPARQL

12

13 Keyword Search Process Keywords Documents query formulation indexing Query Indexed Documents feedback Matching + Ranking Retrieved Documents

14 Keyword Search over RDF Process Keywords Graph query formulation indexing Query Indexed Triples feedback Matching + Ranking + Connecting Retrieved Sub Graphs

15 Main Challenges Find pieces of information in the graph Indexing and Matching problems Connect the pieces of information NP-Complete problem Ranking sub graphs

16 Problem definition Given an RDF graph G and a set of keywords K An answer for K is A minimal sub graph of G that covers all keywords in K Goal: Find all possible answers for K Rank them by relevance

17 Example: Graph G "Denzel Washington" R 3 "South Dakota" R 1 :Actor R 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

18 Index literals values "Denzel Washington" R 3 "South Dakota" R 1 :Actor R 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

19 Associating with the corresponding resources "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

20 Simple indexing example Words Dakota Denzel Fanning South Washington Resource (Property) R 2 (name), R 3 (place) R 1 (name) R 2 (name) R 3 (place) R 1 (name), R 5 (place)

21 Example: K = { Dakota, Washington } "Denzel Washington" R 3 "South Dakota" R 1 :Actor R 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

22 Dakota matches with R 2 (name) and R 3 (place) "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

23 Washington matches with R 1 (name) and R 5 (place) "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

24 Combinations = { (R 2, R 5 ), (R 1, R 3 ), (R 3, R 5 ), (R 1, R 2 ) } "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

25 Combinations = { (R 2, R 5 ), (R 1, R 3 ), (R 3, R 5 ), (R 1, R 2 ) } "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

26 A movie in Washington with Dakota Fanning "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

27 Combinations = { (R 2, R 5 ), (R 1, R 3 ), (R 3, R 5 ), (R 1, R 2 ) } "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

28 A movie in South Dakota with Denzel Washington "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

29 Combinations = { (R 2, R 5 ), (R 1, R 3 ), (R 3, R 5 ), (R 1, R 2 ) } "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

30 A movie in South Dakota and a movie in Washington "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

31 Combinations = { (R 2, R 5 ), (R 1, R 3 ), (R 3, R 5 ), (R 1, R 2 ) } "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

32 Actor Denzel Washington and actress Dakota Fanning "Denzel Washington" R 3 "South Dakota" R 1 :Actor r 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

33 A movie with Denzel Washington and Dakota Fanning "Denzel Washington" R 3 "South Dakota" R 1 :Actor R 4 :Movie R 2 "Dakota Fanning" R 5 "Washington"

34 Problem definition Given an RDF graph G and a set of keywords K An answer for K is A minimal sub graph of G that covers all keywords in K Goal: Find all possible answers for K <= Not reasonable! Rank them by relevance

35 QUIOW A Keyword Search over RDF System

36 QUIOW

37 R2RML Oracle 12C Auxiliary tables RDB RDF Join Class Prop Value QUIOW SPARQL query cores graph Steiner Tree cores Matching Executor Greedy cores Parser Presentation results keywords

38 QUIOW QUIOW approaches Find pieces of information in the graph Auxiliary tables for data and metadata Indexing and Matching with Oracle Text Scoring cores based on The keywords they match The type of resource: Class > Property > Value Some basic statics

39 QUIOW QUIOW approaches Find pieces of information in the graph Greedy algorithm to select minimal cores Connect the pieces of information Steiner Tree approximation for the nucleuses Synthesize a SPARQL query to represent the Steiner Tree Rank the results based on the resources score

40 QUIOW García, G. M., Izquierdo, Y. T., Menendez, E. S., Dartayre, F., & Casanova, M. A. RDF Keyword-based Query Technology Meets a Real-World Dataset. In 20th International Conference on Extending Database Technology (EDBT), Evaluation Petrobras Petroleum Dataset Mondial - 64% of the 50 queries IMDb - 72% of the 50 queries

41 Next Steps

42 Next Steps QUIOW problems The score of cores are mostly based on keyword terms It is basically a Boolean Model Not suitable for ambiguous data as in IMDb and Mondial How can we improve the quality of QUIOW results? More sophisticated Information Retrieval models

43 Next Steps Which IR models we can adapt to RDF graphs? Vector Space Model? Language Model? Page Rank? User profiling? Which of them improves the quality of QUIOW results?

44 Next Steps Which of them improves the quality of QUIOW results? IMDb and Mondial Benchmarks DBpedia Adapt other benchmarks to RDF Social Book Search Last.fm

45 Thank you! Elisa Menendez