What is a knowledge graph? Easy to understand

Author: Liu Zhao Hongyu Liu Shubin Sun Pearl
Source: "Intelligent search and recommendation systems: Principles, Algorithms and Applications"

Guide: The concept of knowledge graph was born in 2012 and was first proposed by Google. The knowledge graph is proposed to accurately explain the relationship between people, things, and things, and it was first applied to search engines. The knowledge graph is to describe the semantics of the text and establish a knowledge database of entity relationships in the natural world. Under normal circumstances, we can use the relationship graph to represent the knowledge graph.

01 What is a knowledge graph

We can examine the concept of knowledge graph from different perspectives.

• From the perspective of the Web , knowledge graphs are like hyperlinks between simple texts, supporting semantic search by establishing semantic links between data.
• From the perspective of natural language processing , knowledge graph is to extract semantic and structured data from text.
• From the perspective of knowledge representation , knowledge graph is a method of using computer symbols to represent and process knowledge.
• From the perspective of artificial intelligence , the knowledge graph is a tool that uses the knowledge base to assist in the understanding of human language.
• From the perspective of the database , the knowledge graph is a way to store knowledge using graphs.

At present, academia has not yet given a unified definition of knowledge graphs. There is a clear description in the documents released by Google. The knowledge graph is a technical method that uses a graph model to describe the relationship between knowledge and modeling the world.

The knowledge graph is also a more general framework for formal description of semantic knowledge. It uses nodes to represent semantic symbols and edges to represent semantic relationships , as shown in Figure 3-1. In the knowledge graph, people, things, and things are usually called entities or ontology.

▲Figure 3-1 Example of knowledge graph

The three elements of knowledge graph include: entity, relationship and attribute.

• Entities : Also known as Ontology, it refers to things that exist objectively and can be distinguished from each other. They can be concrete people, things, things, or abstract concepts or connections. Entities are the most basic elements in the knowledge graph.
• Relationship : In the knowledge graph, the edge represents the relationship in the knowledge graph, which is used to indicate a certain connection between different entities. As shown in Figure 3-1, the relationship between Turing and artificial intelligence, the relationship between the knowledge graph and Google, and the relationship between Google and deep learning.
• Attributes : Entities and relationships in the knowledge graph can have their own attributes, as shown in Figure 3-2.
  
  ▲Figure 3-2 Attributes in the knowledge graph

The construction of knowledge graph involves knowledge modeling, relation extraction, graph storage, relation reasoning, entity fusion and many other technologies. The application of knowledge graphs is embodied in many fields such as semantic search, intelligent question and answer, language understanding, and decision analysis.

02 The value of knowledge graph

The knowledge graph was first applied to search engines. On the one hand, it realizes concept retrieval through reasoning, and on the other hand, it graphically shows users the structured knowledge after classification, so that people can free themselves from the mode of manually filtering web pages to find answers. Apply to intelligent question answering, natural language understanding, recommendation, etc.

The development of the knowledge graph benefits from the development of Web technology, and is affected by KR, NLP, Web, and AI. In the final analysis, the value of the knowledge graph is to make AI smarter .

1. Help search

The purpose of search is to enable people to find something conveniently and quickly in the Internet of Everything. At present, our search habits and search behaviors are still based on keywords. The emergence of knowledge graphs can completely change this search behavior pattern.

When the knowledge graph has not been applied to search engines, the search process is: find the URL with the highest matching degree from a large number of URLs, and return the results with the highest ranking scores to the user according to the query results. In the whole process, the search engine may not need to know what the user entered, because the system does not have the ability to reason, and it is slightly inadequate in precise search.

The search based on the knowledge graph can not only directly answer the user's question, but also has a certain semantic reasoning ability, which greatly improves the accuracy of the search . Figure 3-3 shows a schematic diagram of knowledge map assisted search.

▲Figure 3-3 Knowledge graph helps search

2. Help recommend

Recommendation technology and search technology are very similar, but slightly different. The search technology uses the information pull method, and the recommendation technology uses the information push method, so there are some problems in the recommendation technology, such as cold start and data sparse problems.

Take e-commerce recommendation as an example to introduce the application of knowledge graph in recommendation. Suppose I bought a mobile phone, and the strong subordinate relationship of the mobile phone is the mobile phone case, so that the system can recommend the mobile phone case to me, as well as similar or complementary entities. Figure 3-4 is a schematic diagram of knowledge graph assisting recommendation.

▲Figure 3-4 Knowledge graph helps recommendation

3. Help Q&A

Question answering and dialogue systems have always been one of the key symbols of NLP in the field of artificial intelligence realization. The knowledge graph is equivalent to attaching a background knowledge base to the question answering and dialogue system .

For question answering and dialogue systems or chat robots, in addition to sparse large graphs in open fields such as entity knowledge graphs and interest knowledge graphs, they also need dense small graphs personalized for robots and users. At the same time, the knowledge graph needs to be dynamically updated. Figure 3-5 is a schematic diagram of the knowledge map assisted question and answer.

▲Figure 3-5 Knowledge graph helps Q&A

03 Architecture of Knowledge Graph

The architecture of the knowledge graph involves many aspects such as knowledge representation, knowledge acquisition, knowledge processing and knowledge utilization.

Under normal circumstances, the knowledge graph construction process is as follows : first determine the knowledge representation model, then select different knowledge acquisition methods according to different data sources and import relevant knowledge, and then use knowledge reasoning, knowledge fusion, knowledge mining and other technologies to construct the corresponding knowledge Graph, and finally design the representation of knowledge graph according to different application scenarios, such as: semantic search, intelligent recommendation, intelligent question and answer, etc.

Logically, we can divide the knowledge graph into two levels: the data layer and the model layer . The data layer can be a database stored in fact as a unit. The graph databases that can be selected include RDF4j, Virtuoso, Neo4j and other triples.

<entity, relationship, entity> or <entity, attribute, attribute value> can be used as a basic expression and stored in the graph database. The model layer is built on the data layer and is the core of the knowledge graph. Usually, the data layer is managed through the ontology library, and the concept of the ontology library is equivalent to the concept of "class" in the object. With the help of the ontology library, we can manage axioms, rules, and constraints, and regulate the relationships between entities, relationships, and attributes.

The knowledge graph can be constructed in two ways, top-down and bottom-up . Top-down construction refers to extracting ontology and pattern information with the help of encyclopedia data sources, and adding them to the knowledge base. Bottom-up construction refers to the use of certain technical means to extract resources from public data, select information with higher confidence, and add it to the knowledge base after manual review.

In the early stage of the development of knowledge graphs, most companies and institutions adopted a top-down approach to construct knowledge graphs. At present, most companies adopt a bottom-up approach to construct knowledge graphs.

The structure of the knowledge graph is shown in Figure 3-6.

▲Figure 3-6 The structure of the knowledge graph

• Knowledge sources : including structured data, unstructured data and semi-structured data.
• Information extraction : It is to extract entities, attributes, and relationships between entities from various types of data sources, and form ontology's knowledge representation on this basis. There is a large amount of unstructured or semi-structured data in the process of constructing the knowledge graph. In the process of constructing the knowledge graph, these data need to be extracted by natural language processing methods. From these data, we can extract entities, relationships and attributes.
• Knowledge fusion : The main work is to extract structured data and entity information extracted from information, and even third-party knowledge bases to perform entity alignment and entity disambiguation. The output of this stage should be various ontology information fused from various data sources.
• Knowledge processing : The stage of knowledge processing is shown in Figure 3-6. The important work in knowledge reasoning is the completion of the knowledge graph. Commonly used methods of complementing knowledge graphs include: complementing methods based on ontology reasoning, implementation of related reasoning mechanisms, and complementing methods based on graph structure and relationship path characteristics.