Category: General Catalog of "Knowledge Graph"
Knowledge Graph is learned artificial intelligence
There are many schools of thought in early artificial intelligence, two of which are often referred to as connectionism and symbolism . Connectionism advocates that the realization of intelligence should simulate the physiological structure of the human brain, that is, use a computer to simulate the neural network connections of the human brain. This genre has developed to this day, the so-called widely developed deep neural network. Another school is called Symbolism, which advocates that the realization of intelligence should simulate the human mind, that is, use computer symbols to record the memory of the human brain to represent the knowledge in the human brain, that is, the so-called knowledge engineering and expert systems. Deep learning first succeeded in perceptual tasks such as vision and hearing. It essentially solves the problem of pattern recognition, which can be compared to the realization of a smart AI. But perception is still a low-level intelligence, and the human brain relies on the knowledge learned to think, reason, and understand language. Therefore, there is another kind of AI that can be called learned and knowledgeable AI. This is closely related to the knowledge graph.
What is cognitive intelligence? Cognitive intelligence has two core research propositions, one is language understanding , and the other is knowledge representation and processing. Humans accumulate knowledge about the world by knowing the world, and solve the problems they encounter through the knowledge they have learned. For example, a doctor uses his medical knowledge to see a patient. Language is the most direct carrier of knowledge. So far, most of human knowledge has been described, recorded and inherited through natural language. At the same time, the correct understanding of language requires the help of knowledge. As shown in the picture below, here is an interesting example: "The G20 meeting was held in the morning, and Ma Yun said to his secretary: 'Buy me KFC at noon.' After 30 minutes, the secretary came back and said: 'Buy it, a total of 460 million US dollars, Are we Alipay or cash?'". This was of course just a joke. The news at the time was that Primavera Capital and Ant Financial had jointly invested $460 million in Yum! Brands, the parent company of KFC. The focus here is on the importance of background knowledge for proper understanding of language. If Jack Ma's secretary is an artificial intelligence, in the first context, it should recognize KFC as a food, and in the second context, it should recognize KFC as a company, and it also needs to know KFC The parent company of Yum! Brands is Yum! Ant Financial has invested in Yum! Jack Ma is the founder of Alibaba. Alibaba and Ant Financial have an associated relationship to properly establish the relationship between Jack Ma and KFC. The relational network behind this is actually the knowledge graph. In fact, everyone has a large amount of this type of knowledge graph about the relationship between everything in everyone's brain, and we rely heavily on this background knowledge to accurately understand language and make correct judgments.
The way of carrying and expressing knowledge
Human natural language, as well as created paintings and music, mathematical language, physical models and chemical formulas, are the representation and inheritance of human knowledge. The ability to acquire, represent and process knowledge is one of the most essential characteristics of the human mind that distinguishes it from the minds of other species. The traditional field of artificial intelligence has a classic research direction - knowledge engineering and expert systems. The basic idea of this often referred to as GOFAI (Good Old Fashioned AI) is to build a system that can acquire knowledge from the brain of experts, and then provide services for non-expert users through an inference engine, such as auxiliary diagnosis, case judgment, etc. This process of acquiring knowledge from the human brain is called knowledge engineering. Knowledge has many expression carriers and forms of existence. For example, natural language is the most important expression carrier of human knowledge. Since the human brain can acquire and learn knowledge from text through reading, the machine brain should also have the ability to extract knowledge from text. But text strings don't seem very friendly to machines, which are still struggling to understand human language. For example, products like Microsoft Xiaoice, Apple Siri, and Xiaomi Xiaoai speakers still face huge challenges in the experience of human-machine dialogue.
Currently, there are two main technical routes for understanding knowledge in text through machines. The first is extraction technology, such as identifying entities, relationships and logical structures from text; the second is language pre-training, that is, a large neural network model is trained through a large amount of text corpus, and the knowledge in the text is implicit in the parameters vectorized models, while vectorized representations and neural networks are machine-friendly. Therefore, the text itself can also be used as a knowledge base (Knowledge Base).
The essence of knowledge graph is a structured knowledge representation. Simply put, knowledge graphs are designed to use graph structures to model, identify and infer complex relationships between things and precipitate domain knowledge, and have been widely used in semantic search, intelligent question answering, language understanding, media understanding, reasoning engines and Decision engine and many other areas, as shown in the figure below. Compared with text, structured data is easier for machines to process, such as query and question answering. At the same time, the graph structure can express richer semantics and knowledge than the string sequence.
For machines, graph structures are certainly more friendly than text. The rise of deep learning, or more precisely representation learning, suggests that parameterized vectors and neural networks are more suitable information carriers for machines to perform fast computations. For example, in natural language, a vector representation can be learned for each word; in image processing, a vector representation can be learned for each object in a visual scene; in knowledge graphs, a vector representation can be learned for each entity and relationship a vector representation. We usually refer to these vectorized representations as Embedding or Distributed Vector Representation. As shown in the figure below, if we project the vector representations of all digital objects into a vector space, we will find that the vectors of different images of the same digital object are closer in space. Further, by projecting words, entities, objects and relationships into the vector space, it is more convenient to operate on these language, visual and entity objects in the vector space, and even use neural networks to achieve logical reasoning.
Knowledge graph is a world model knowledge
The graph can essentially be regarded as a world model - the World Model. Throughout the history of the development of artificial intelligence-related directions, there has always been a core proposition to find a suitable machine representation of everything to record knowledge about the world. In the era of traditional expert systems, people invented symbolic knowledge representation methods such as description logic to describe everything. Human natural language is also a symbolic representation of the objective world. In the Internet era, people imagine using ontology and semantics to link various data and knowledge published on the Internet, which is also one of the origins of knowledge graphs.
With the rise of representation learning and neural networks, it has been found that numerical vector representations are easier to capture those hidden knowledge that cannot be easily represented explicitly, and are easier for machines to process than symbolic representations. Knowledge graph embraces both the symbolic representation and vector representation of machines, and can organically combine the two to solve problems in search, question answering, reasoning, and analysis . Combined with new artificial intelligence technology means such as knowledge graph and neural network, the expert system can be reconstructed. The original knowledge base can use the method of knowledge graph to make the means of knowledge acquisition easier. In addition to traditional symbolic representation, it is also necessary to consider how to use vectors to represent knowledge such as entities and relationships. In terms of knowledge acquisition, experience at the expert level provides an important input for the construction of modern knowledge. In addition, there is now a large amount of machine data such as device sensor data, automatically collected log data, and data from multiple modalities. To a certain extent, the knowledge graph can act as a bridge, combining the empirical knowledge of experts with machine data through a more effective representation.
In terms of reasoning engines, there are many traditional symbolic reasoning engines. Due to many bottlenecks, such as very high requirements for knowledge quality, they have not been able to achieve large-scale commercial applications. The current reasoning engine can achieve better reasoning under the blessing of deep learning technologies such as neural networks and representation learning. In terms of user-oriented interaction, it is no longer just a simple query, but also can realize search, intelligent question answering and decision analysis based on graph analysis. And through some visual means to support the overall analysis of all data in all dimensions, and provide interpretability for all inference results.
References:
[1] Chen Huajun. Introduction to Knowledge Graph [M]. Electronics Industry Press