week1 - What_Is_A_Knowledge_Graphs_S.pdf
- 1. CS520: KNOWLEDGE GRAPHS Data Models, Knowledge Acquisition, Inference, Applications Learn about the basic concepts, latest research & applications Lectures and Invited Guests Spring 2021, Tu/Thu 4:30-5:50, cs520.Stanford.edu
- 3. Motivation for the Seminar • Knowledge Graphs are being used in • Web search • Answering questions • Data integration • Knowledge Graphs are also target of output for • NLP and computer vision algorithms • ML algorithms more generally • Knowledge Graphs are a topic of a major program from NSF • https://www.nsf.gov/od/oia/convergence- accelerator/Award%20Listings/track-a.jsp
- 4. Seminar Outline Knowledge Graph • What is it? • How do create it? • How do we reason with it? • How do we use it with modern AI algorithms? • Where is the research?
- 5. Course Design • Two 80-minute sessions each week (Tue/Thu) • Tuesday sessions based on the synthesis of key points from the 2020 series • The synthesis points are also available as written notes on the course website • Some Tuesday sessions will also have invited guests • Thursday sessions will feature invited guests • (Generally) two 30-minute presentations • Followed by Q & A • Recordings will be available on the course web site
- 6. For Stanford Students • Complete a quiz for all 10 of the Tuesday sessions • Submit a written summary for any 8 of the 10 Thursday sessions
- 7. What is a Knowledge Graph?
- 8. Outline • Knowledge Graph • Resurgence of interest in Knowledge Graphs • Search engines • Data integration • Artificial Intelligence • What is new and different?
- 9. What is a Knowledge Graph? B C A Directed Labeled Graph Nodes and edges have well-defined meanings
- 10. What is a Knowledge Graph? Predicate Object Subject Directed Labeled Graph Nodes and edges have well-defined meanings
- 11. What is a Knowledge Graph? Relation Entity Entity Directed Labeled Graph Nodes and edges have well-defined meanings
- 12. What is a Knowledge Graph? subclass of Class Class Directed Labeled Graph Nodes and edges have well-defined meanings
- 13. What is a Knowledge Graph? friends bob art Directed Labeled Graph Nodes and edges have well-defined meanings
- 14. What is a Knowledge Graph? subclass of Human Person Directed Labeled Graph Nodes and edges have well-defined meanings
- 15. Different ways to define meaning • Based on a user’s actions • friend relationship • Explanation in a human understandable language • E.g., linguistic resource Wordnet • Logical Specification • Using a set of axioms • Associating examples • Defining a cat using a set of images • Embeddings • Statistics on a corpus of text
- 16. Rich History of work on Knowledge Graphs • Knowledge Representation • Semantic networks • Description logics • Conceptual graphs • Database systems • Network databases • Triple stores
- 17. Outline • Graphs in Computer Science • Resurgence of interest in Knowledge Graphs • Search engines • Data integration • Artificial Intelligence • What is new and different?
- 18. Knowledge Graphs in Search • The Winterthur example • This example was introduced by Denny Vrandečić • For more details • Visit his Spring 2020 presentation • A story linked to the course website
- 19. Knowledge Graphs in Web Search
- 21. Knowledge Graphs in Web Search
- 22. Knowledge Graphs in Web Search
- 26. Problem • Twin Towns and Sister Cities are identical concepts • The reference to Winterthur in the Ontario Page appears in text description • There is no easy way to resolve the differences
- 27. Solution • Wikidata: Publicly curated Knowledge Graph
- 31. Graph Underlying Wikidata Winterthur Ontario Twinned administrative body
- 32. Graph Underlying Wikidata Winterthur Ontario Twinned administrative body Zurich Metropolitan Area Switzerland part of part of country United States part of North America
- 33. Graph Underlying Wikidata Winterthur Ontario Twinned administrative body Zurich Metropolitan Area Switzerland part of part of country United States part of North America Winterthur same as LIBRARY OF CONGRESS
- 35. We can also query the data Display on a map the birth cities of people who died in Winterthour? • Requires querying multiple data sources on the web • Requires understanding their schemas • Schemas published using Schema.Org vocabulary • Structured results can then be included in the search results on the web pages
- 36. Scale of Wikidata 80+ Million Objects
- 37. Scale of Wikidata 1+ Billion Relationships
- 38. Scale of Wikidata 4872+ Catalogs
- 39. Wikidata Knowledge Graph • A graph of unprecedented scale • Collaboratively created • Data may be curated manually or automatically • Semantic definitions in Schema.Org • Compelling use case: Web Search
- 40. Outline • Graphs in Computer Science • Resurgence of interest in Knowledge Graphs • Search engines • Data integration • Artificial Intelligence • What is new and different?
- 41. Example Use Case • 360 Degree View of a Customer INTERNAL COMPANY DATA Who is funding who? Who supplies to who? Who are my customers? Risk Analysis for Lending Decisions Business Intelligence for Marketing
- 42. Data Integration • Data reside in multiple sources • Company directory, product catalog, government database, weather report, … • Answering queries requires combining data from multiple sources • We need to provide translations of data between multiple sources • Direct mappings • Shared schema
- 43. Data Integration • Schema-free approach to data integration • Convert the relational data from multiple sources into triples • Stored in a graph database • Referred to as a knowledge graph • Deal with schema mappings/translations on “pay as you go” basis • Visualization • Optimized for graph traversals
- 44. Outline • Graphs in Computer Science • Resurgence of interest in Knowledge Graphs • Search engines • Data integration • Artificial Intelligence • What is new and different?
- 45. Artificial Intelligence • Output representation for • Natural Language Processing • Computer Vision • Input representation for machine learning • Language Models • Graph Models
- 46. Natural Language Processing Albert Einstein was a German- born theoretical physicist who developed the theory of relativity.
- 47. Natural Language Processing • Entity Extraction Albert Einstein was a German- born theoretical physicist who developed the theory of relativity.
- 48. Natural Language Processing • Entity Extraction Albert Einstein was a German- born theoretical physicist who developed the theory of relativity. • Relation Extraction
- 49. Natural Language Processing • Entity Extraction Albert Einstein was a German- born theoretical physicist who developed the theory of relativity. • Relation Extraction Question Answering Common Sense Reasoning
- 50. Computer Vision
- 52. Computer Vision Object Detection • Edge Detection Man glasses bucket horse wearing feeding holding eating from
- 53. Computer Vision Object Detection • Edge Detection Man glasses bucket horse wearing feeding holding eating from Visual Question Answering
- 54. Input to Machine Learning • Machine learning requires numerical input • Symbolic inputs must be converted to numerical input • A process known as embedding • Word Embeddings • Graph Embeddings
- 55. Word Embedding • Primary use case is to calculate similarity between words • “like” is similar to “enjoy” • But, generally useful for a variety of language understanding tasks • Key idea: capture the meaning of a word by counting how often it occurs next to other words
- 56. Word Embedding I like knowledge graphs. I like databases. I enjoy running.
- 57. Word Embedding I like knowledge graphs. I like databases. I enjoy running.
- 58. Word Embedding Meaning of a word is captured by the vector corresponding to each row of co-occurrence counts Word similarity can be calculated using the distance between the vectors
- 59. Word Embedding • A large-scale text corpus can have billion plus words • The storage requirement for the vectors blows up • Dimensionality reduction (typically in the range of 200) • Linear algebra techniques (e.g., Singular Value Decomposition) • Automatic learning of the necessary parameters
- 60. Word Embedding • A popular application is to predict the next word
- 61. Graph Embedding • Application areas • Recommendation engines • Generalize what we did for word embeddings • Goal is still to reduce the nodes to vectors so that we can calculate the node similarity as a difference between the vectors
- 62. Word Embedding to Graph Embedding • Word embeddings view the text as a linear graph • Word prediction is the instance of more general problem of link prediction I like knowledge graphs .
- 63. Graph Embedding • Example encoding function • Randomly walk the graph • Compute the cooccurrence counts between the nodes • Once nodes have been converted into vectors • calculate node similarity • Optimize the encoding function
- 64. Knowledge Graphs and AI • Output representation for • Natural language processing • Computer vision • Input representation for machine learning • Language models • Graph models
- 65. Summary • Graphs are a fundamental construct in discrete mathematics • Defining meaning is the crux of the problem for knowledge graphs • Rich history in knowledge representation and databases • Recent surge of interest driven by • Use of structured data in web search results • Progress in NLP and vision • Progress in ML to perform predictive tasks • What’s new? • Scale • Bottom-up development • Multiple modes of construction
- 66. What are Knowledge Graphs and why do we need them? Prof. Chaitanya Baru National Science Foundation Thursday, April 1, 2021