Overview of knowledge map technology system
Article directory
1. Architecture of Knowledge Graph
The knowledge graph includes many technologies such as knowledge production, knowledge representation storage, and knowledge application. The above figure involves construction technology, storage technology, application technology, and many other related technologies. It can also be seen that the knowledge map includes the research of most subdivisions in the field of artificial intelligence, and also includes a large number of big data technologies, such as construction technology, there are also big data-related mapping construction technologies, and extraction-based construction Aspects of natural language processing and computer vision and sound processing techniques.
In terms of storage of knowledge graphs, graph databases are now the mainstream in the industry. Graph databases involve a large number of big data technologies. For example, distributed graph databases have strong technical associations with distributed computing and distributed storage such as Hadoop and Spark. And some relatively new technologies, such as vector databases, are also a relatively cutting-edge research field.
In terms of the application of knowledge graphs, it also includes a lot of content, such as knowledge retrieval, knowledge exploration, and involves graph database query languages Gremlin, Cypher, and SparQL. In terms of knowledge search, some other methods will be used to obtain the required knowledge from the graph database or other storage methods. In more applications, such as knowledge computing, knowledge reasoning, etc. Knowledge computing is more about calculations related to graphs, and knowledge reasoning is more about deep learning, such as graph neural networks. There are also many specific applications, such as question and answer, recommendation, data analysis, knowledge traceability, and decision-making assistance. These are applications of artificial intelligence or knowledge graphs that are directly related to specific businesses.
2. Knowledge map construction technology
1. Knowledge graph mode
There are two methods of knowledge graph in application: schema-free and schema-restricted. The knowledge graphs we usually apply in the industry are limited in schema. This mode refers to the knowledge map mode, also known as ontology, class map or concept map.
Knowledge map mode: defines what content the knowledge map needs to contain, or it is a standardized expression of the conceptualization of knowledge. How to design a knowledge graph model is the first problem to be considered in the industry, because it is often combined with our business, and requires knowledge graph experts and business-oriented experts to work together to sort out the design. Some methods are needed to tell business experts how to understand the knowledge graph pattern and how to design the pattern.
2. Entity extraction
After designing the knowledge map model, one of the problems is how to construct the knowledge map from the existing data. It mainly involves two aspects: the first aspect, if our data is structured, we can write some rules and add some simple mapping to build a knowledge graph. This method is often called mapping construction.
On the other hand, there are a large number of documents, such as Word documents, PDF files and other semi-structured content, for which we need to perform entity extraction, relationship extraction, and reference resolution and other aspects of knowledge extraction. The most basic of these is entity extraction, which is also called named entity recognition in natural language processing. Many of the technologies here have a history of 40 to 50 years from the 1980s to the present. There are various methods. The relatively simple one is to write rules, and the more mature one is the machine learning method, such as CRF conditions. Airports, hidden Markov models, and deep learning, the current BERT-based pre-training model, and weakly supervised learning and entity extraction methods based on deep reinforcement learning, etc.
3. Relationship extraction
The biggest feature of the knowledge graph is the relationship. After entity extraction, relation extraction is performed.
Relationship extraction is a very important content. How do we establish various relationships between entities, which is also a very important point that distinguishes knowledge graphs from labels or other structured data. Relation extraction is also a very difficult part of knowledge graph construction. There are many different methods, such as rule-based methods, deep learning-based methods, and weakly supervised learning methods can be used for relation extraction. Of course, there is also a very popular method after the rise of deep learning: the entity-relationship joint extraction method, which realizes the extraction of entities and relationships in one model at the same time. In a specific application, you can choose an appropriate method for relation extraction according to your needs.
4. Event extraction
The other part is event extraction, which is more complex and similar to entity relationship extraction.
Event extraction needs to extract more content, including various elements such as subject, object, place and time. And there are various ways to associate it with relational extraction. In general, it is divided into pipeline mode and joint extraction mode. The pipeline mode is to extract the trigger words first, and then classify the trigger words for element identification. The joint extraction method is to identify all elements and subject and object trigger words together at one time, and give the final prediction result. Both methods are used a lot, and you can choose different methods for different scenarios.
3. Knowledge storage
Whether it is entity relationship or event extraction, there must always be a place to store it. Now the industry basically uses graph databases, that is, attribute graph databases, which are completely compatible with knowledge graphs.
4. Application technology of knowledge map
1. Knowledge Computing
For the knowledge map that has been built and stored in the graph database or other places, we have to apply it.
Common algorithms such as path analysis, community classification, and centrality. The most common path analysis is the shortest path algorithm. In community classification, Louvain and GN algorithms are widely used. There are also many well-known algorithms in centrality, such as betweenness centrality and PageRank algorithm, which can help us select some key nodes, key paths, and key edges in the knowledge base under certain conditions. Underneath it are some corresponding libraries, including single versions such as NetworkX and igraph, and distributed ones such as GraphX. The query languages Gremlin, Cypher, and SparQL carried by graph databases can also support knowledge computing and graph algorithms to a certain extent. Some related applications.
2. Knowledge reasoning
Next is knowledge reasoning. Knowledge reasoning is a key technology for the knowledge base to realize decision-making reasoning and human-like reasoning.
Reasoning itself includes probabilistic reasoning, inductive reasoning, deductive reasoning, causal reasoning, etc.
3. Intelligent Q&A
Intelligent question and answer based on knowledge graph is a very popular content. With knowledge graph, we can do many corresponding algorithms or knowledge acquisition content.
KBQA has a lot of papers and algorithms to directly realize the question and answer of knowledge graphs. There are many effective algorithms. You can refer to some of the most advanced SOTA algorithms. Of course, these algorithms are aimed at a specific problem and cannot be generalized. They are related to knowledge. Graph bound. The question and answer we are introducing now is actually a framework, through decomposition, and then through retrieval or the combination of previous search engines and knowledge graphs to provide such a complex framework structure. Through this architecture to achieve universal intelligent question answering, many adaptations can be made in the industry, and it can be applied to different scenarios and some applications in different directions. The versatility will be stronger and the adaptability will be stronger. Of course, in terms of complexity It is also much more complex than a single algorithm.
4. Cognitive recommendation
We distinguish knowledge graph-based recommendations from ordinary recommendations. The former is called cognitive recommendation, which refers to doing many things with knowledge graphs as the core.
This includes diversity recommendation, multimodal recommendation, giving reasons for recommendation and persuasive recommendation, etc. It actually simulates some situations when we humans make recommendations. For example, when you ask me how to learn about knowledge graphs, I may tell you that you can start by simulating a paper and reading the Everest book, and then help you consolidate by recommending corresponding projects on Github learned knowledge. Through this kind of learning, you may be able to learn the content of the knowledge graph well.
Summarize
This article mainly introduces the architecture of the knowledge graph, and analyzes the scenarios of the knowledge graph in practical applications from a macro perspective. If you are interested, you can communicate with each other. For the text content, please refer to Wang Wenguang's "Shuzhi", if you are interested, you can read the original text