UCAS-AI Academy-Special Course of Knowledge Graph-Lesson 1-Course Notes

Knowledge Graph Overview

What is a knowledge graph

• Three stages of artificial intelligence
  • Computational intelligence (computing and storage)
  • Perceived intelligence (interact with nature)
  • Cognition can only (rational cognition formed in the process of perception and interaction)-inseparable from knowledge
• Data: Perceptual signals reflecting the motion state of objective things, the most primitive record of human brain perception
  • Raw explanation
  • No connection with other data
  • Does not have semantics and cannot answer any questions
• Information: processed and interpreted data that has meaning through some association
• Knowledge: Systematic information that can be used for decision-making after selection and transformation
• Knowledge Engineering: Data, Information-Knowledge
• Types of knowledge:
  • Declarative knowledge: Describes static information such as traits and relationships of transaction, which is divided into three levels: things (specific), concepts (a category), and propositions (relationships between things)
  • Process knowledge: describe dynamic information such as problems and solutions, including rules (cause and effect) and control results (solution steps)
• Knowledge base: on the basis of collecting and sorting out various knowledge, it is formally expressed, stored according to a certain method, and provides query means
  • The foundation of knowledge sharing and application
  • Knowledge graph is a form of knowledge base
• Big Data Knowledge Engineering: Yang Computer manages and uses information more effectively
  • Knowledge: structured and related
  • Features: large-scale, open domain, multi-dimensional, self-learning
• Knowledge Graph: Structure of the graph
  • Structured triples -entities and their relationships: $\mathcal G = (\mathcal E, \mathcal R, \mathcal S) ,\ \mathcal S \subseteq \mathcal R \times \mathcal E \times \mathcal E$
  • Triple element: head entity, tail entity, inter-entity relationship
  • Relationships are sometimes called attributes, and the tail entity is now the attribute value
  • Entities are nodes and relationships are directed edges of linked nodes
• Knowledge graph features:
  • Structured, use graphs to model the structure
  • Correlation, natural correlation of multi-source knowledge
  • Standardized, semantic web framework, easy to share and use
• Important infrastructure for artificial intelligence

The development course of knowledge graph

• Artificial intelligence: the digitization of knowledge-let computers represent, organize and store human knowledge
• Semantic Web: Knowledge of data-let data support only tasks such as inference
• Knowledge Graph: a knowledge engineering product based on the semantic network theory, based on the semantic web technical framework and engineering specifications, for the knowledge of Internet data

Types of knowledge graphs and representative knowledge graphs

Entities, relationships and words

• Entity: things that exist objectively and can be distinguished from each other, can be specific people, things, things, or can be abstract concepts
• Relation: Various connections between different entities
  • Taxonomic relationship (affiliation, hierarchy)
  • Non-taxonomic relationships-part of the whole, thematic decisions, attributes, ownership, causation
• Words and Phrases: Language units used to describe entities and relations of these cognitive units
  • The exact meaning of entities and relationships
  • Words are ambiguous

Ontology, knowledge base and database

• Ontology: A set of shared conceptual systems that describe the objective world
  • Describe the definition of data, not instance data describing specific things
  • Concepts, relationships and axioms
  • Formal ontology: extensive use of axioms
  • Lightweight ontology: no or little use of axioms
• Knowledge Base (Knowledge Base) subject to Ontology control of knowledge examples and its carriers
• Database: A product designed and developed for representing and storing data required by computer applications with a computer

Taxonomy, Ontology and Social Taxonomy

• Taxonomy: a professional-level category system compiled by experts
  • There are also some strict classification systems
• Ontology: Specification of shared concepts
  • Covers the classification system between concepts, and more importantly, there is a correlation between concepts, and inference rules based on this
  • With strict specifications, users are difficult to build
• Social Taxonomy (Folksonomy): a taxonomy that is automatically formed by users ’free tags
  • Labeling method, not necessarily classification system
  • The label is arbitrary (ambiguity)

Types of knowledge

• Language knowledge: knowledge at the language level
• Encyclopedia knowledge: general knowledge covering various industries and fields
• Domain knowledge: expertise in a specific field
• Scene knowledge: the knowledge required in a specific scene or the need to complete a task
• Common sense knowledge: knowledge that everyone recognizes
• Each knowledge type can be composed of the corresponding type of knowledge graph

Representative knowledge graph

• Cyc

  • Terms + Assertions
  • Supports deductive and inductive reasoning
  • Relying on experts, relying on first-order predicate logic is not flexible enough

• WordNet

  • English Vocabulary Semantic Knowledge Base
  • Semantic field theory
  • Manual annotation, component semantic concept network
  • Synonyms, corresponding to semantic concepts

• FrameNet

  • Lexical semantic knowledge base
  • Frame semantic theory
  • Semantic Framework: Knowledge Presupposition on Concepts Contained in Words—Knowledge Framework and Concept Tools
  • Hierarchical organization structure

• HowNet

  • A networked knowledge system that describes the relationship between concepts and the relationship between concept attributes
  • Yoshihara describes the concept, and the Yoshihara is related by the Yoshihara relationship to form a network system

• ConceptNet

  • An open, multilingual attempt knowledge graph composed of common sense describing concepts and their relationships
  • Help the computer understand the meaning of everyday words
  • Closer to WordNet, but more involved
  • Nodes are words, ambiguous through part-of-speech and category elimination

• Wikipedia

  • Free online encyclopedia
  • Document structure: Each page defines a deterministic entity, and hyperlinks form the mapping relationship of words
  • Wikipedia-based knowledge bases are used to mine knowledge, but different knowledge bases use different treatments for node labeling, ambiguity processing, taxonomic design, etc.

• Construction strategy based on Wikipedia knowledge base

  • Determine entity collection by page title and hyperlink vocabulary
  • Entities are divided into different categories
  • Categories are related to each other through relationships such as upper and lower positions
  • Entities and categories are described by attributes and relationships between them
  • Relations can be inferred through implication relations

• DBPedia

  • Community building

• YAGO

  • Knowledge system based on WordNet, linking Wikipedia entries with it
  • Language ontology and world knowledge fusion

• BabelNet

  • Multilingual vocabulary semantic network and ontology
  • Add multilingual support

• FreeBase

  • Completely structured knowledge resources built using group-only methods

• Vocabulary vs. entity

• Language vs. Encyclopedia vs. Common Sense

• Manual build vs. machine build

• Based on ontology vs. based on classification system

Life cycle of knowledge graph

Knowledge ontology construction

• Knowledge modeling
• How to express knowledge and describe the target knowledge
  • Quilt system
  • Entity, concept
  • Semantic relations
  • Inference rules
• Input: fields, application scenarios
• Output: domain knowledge ontology
• Key technology: ontology engineering
• Semantic web knowledge modeling
  • concept
  • relationship
  • Conceptual relationship
  • Use the resource description framework RDF for description
    • Resource (object)
    • Predicate (relationship between characteristics and resources)
    • Statement: RDP triplet <S, P, O>

Knowledge acquisition and verification

• Input: domain knowledge ontology, massive data
• Output: instance knowledge (entity collections, relationships, attributes)
• Main technologies: information mining, text extraction
• Estimate the credibility of knowledge

Knowledge fusion

• Knowledge integration
• Shards assembled into a network
• Input: extracted knowledge, knowledge ontology, existing knowledge base
• Output: unified knowledge base, knowledge confidence
• Key technologies: ontology matching, entity link
• Task: To integrate knowledge from different sources, different languages ​​or different structures, so as to continuously deduplicate and update existing knowledge
• Divided into knowledge ontology fusion (fusion of knowledge system) and knowledge instance fusion
• Can also be divided into vertical integration (different levels) and horizontal integration (same level)

Knowledge storage and query

• Input: large-scale knowledge graph
• Output: Knowledge base storage structure, query service
• Main technologies: knowledge representation, knowledge query language, storage, retrieval engine
• RDF graph model:
  • RDF triples, stored line by line in text
  • Query language SPARQL
• Attribute graph model:
  • Quintuple $G = (V, E, \ rho, \ lambda, \ sigma)$
  • The last three items: association, assignment of labels, and assignment of associated attributes
  • Query language: Cypher

Knowledge reasoning

• Task: Use inference to discover the hidden knowledge in the existing knowledge
• Input: large-scale knowledge graph
• Output: implicit knowledge
• Main technologies: reasoning based on logic rules (symbols), reasoning based on representation learning (numbers)
• Symbolic reasoning: infer directly on entities and relational symbols in the graph
  • The essence: learn and apply inference rules
  • Learning inference rules (induction)
  • Apply rules to reason about specific facts (deduction)
• Numerical reasoning: use numerical calculations (vector matrix calculations) to capture implicit associations on the knowledge graph
  • Essential: distributed knowledge representation
  • Core idea: Represent symbolic entities and relationships in a low-dimensional continuous limited space

Knowledge application

• Semantic search
• Q & A (with reasoning ability)
• recommend

Knowledge graph and deep learning

• Know its AI-"Know it's AI-interpretability