"Knowledge is power" is familiar to us, but there is actually a second half of Bacon's sentence "the more important thing is the skill to use knowledge". For artificial intelligence, the knowledge map is the embodiment of how it uses knowledge. In the financial field, how can this skill be used to better understand customer needs, improve business efficiency and customer satisfaction, and manage risk at the same time? China Merchants Bank gave their answer.
Author | Li Jinlong, He Yaohan, Zheng Guidong
Produced | New Programmer
A knowledge graph is a structured semantic network for describing entities, attributes and their relationships, usually presented in the form of a graphical model. Knowledge graphs can help machines understand information and support developments in areas such as natural language processing and search engine optimization. Applied in the business scenario of China Merchants Bank, we divide the knowledge map into three concepts from the bottom up: the bottom layer is a complex network analysis algorithm based on graph database; the middle layer is a data semantic network algorithm; the upper layer forms expert knowledge representation, and through recognition Intelligent computing is comprehensively applied in various scenarios in the industry.
Three Connotations of China Merchants Bank's Knowledge Graph
By building a knowledge graph in the domain (see Figure 1), we express the business scenarios in the industry in a semantic form to form new knowledge and empower each scenario.
Figure 1 Knowledge Graph Platform
complex graph analysis
Based on the symbolic representation, the knowledge map can also learn the characteristics of the graph through the graph analysis algorithm, and obtain a corresponding vector representation for each entity and relationship in the map. At the same time, efficient knowledge reasoning calculations are realized by using calculations between vectors, matrices or tensors. The rapid development of graph databases provides technical support for large-scale graph query and graph computing, thereby carrying out complex network analysis tasks, and is widely used in marketing, risk control and other scenarios in the financial field.
Semantic Web
Semantic Web (Semantic Web) was first proposed by Tim Berners-Lee in Science Magazine in 2001. Knowledge graph can also be regarded as a kind of data semantic network. A node in a semantic network can represent a concept (concept), an attribute (attribute), an event (event) or an entity (entity), while an arc is used to represent the relationship between nodes, and the label of the arc indicates the relationship type. The knowledge map represents knowledge in the form of a graph. Based on the federated knowledge map, the interconnection and intercommunication between each map knowledge is realized, so as to realize the global reasoning and prediction of the map.
expert knowledge representation
Knowledge Graph is a graph-related structured knowledge base, which is used to describe concepts and their relationships in the physical world in symbolic form. Its basic unit is the triplet form of "entity-attribute-entity". Entity They are interconnected through relationships to form a networked knowledge structure. The knowledge map stores expert knowledge through the map structure, which can serve the field of cognitive computing. It provides interpretable judgment criteria in scenarios involving text information acquisition and processing, and realizes the systematization and intelligence of information acquisition (see Figure 2). .
Figure 2 Distributed representation of graph relationships
Practice in the financial field
Knowledge graphs are currently widely used in the financial industry, and their structured knowledge can help banks better process and understand complex information. The construction of knowledge graphs in China Merchants Bank is divided according to "3+1 layers": the first layer is the knowledge layer, and the tools and knowledge at this layer are mainly used in the field of cognitive computing, such as knowledge centers, intelligent audits, AI quality inspections, etc. scenario; the second layer is the data semantic network, such as the construction of a federated knowledge map; the third layer is a graph database, which is used to improve the efficiency of decision-making models such as marketing, risk control, and anti-money laundering; the fourth layer is comprehensive applications, such as investment In the research field, it can be applied to customer-oriented intelligent online investment advisory scenarios.
Construction of a unified knowledge center and related application support - intelligent audit, quality inspection
As a knowledge-intensive field, the banking industry can generate a large amount of unstructured data every day in various business scenarios, and form a set of unified specifications and labels for this knowledge, so as to facilitate the tasks of production work and knowledge sharing among various institutions in the whole bank. Ultimately achieving the value generated by knowledge has always been a pain point that needs continuous improvement. In order to form a set of knowledge specifications in line with the development of AI, China Merchants Bank has spent many years sorting out, analyzing and organizing the various knowledge accumulated in the bank over the years to form a unified knowledge management center, which uses AI technology to integrate knowledge management and language representation. , semantic algorithms, knowledge utilization, and finally effectively support the intelligent application of knowledge.
In the application of underlying data, all kinds of business documents, rules and regulations, professional knowledge, question and answer knowledge, information, internal forums and other data are unified in the China Merchants Bank knowledge center and stored in the form of databases, graphs, etc. In terms of upper-level capabilities, businesses can pass In the form of knowledge dismantling, knowledge editing, knowledge authorization, knowledge search, and personalized knowledge recommendation of the scenario department, scenario knowledge sharing and intelligent services are carried out through various channels of China Merchants Bank, and rule reasoning and discovery are carried out in combination with pre-trained language models, so as to achieve Auxiliary business conducts intelligent reasoning and application.
Relying on the product data in the bank, China Merchants Bank organizes and designs the entities in the knowledge map, designs the entity relationship through product and service standard logic, and builds a knowledge map with a huge amount of data. The knowledge graph we built is widely used in various businesses in the bank. Not only is the knowledge graph used in various scenarios of outlet services to improve service quality and efficiency, but the knowledge graph is also used as an important basis for China Merchants Bank’s service quality inspection in customer consultation. are playing an increasingly important role in such scenarios.
Construction and Application of Federal Knowledge Graph
The federated knowledge map is a combination of knowledge map and open ecology. In general scenarios, the federated knowledge graph connects the business data of various departments within the bank, and introduces some external industrial and commercial data, etc., which provides support for open collaboration and is also an important part of the financial brain. Under this concept, China Merchants Bank has built a federated knowledge graph, which can support common graph analysis algorithms such as community discovery, tag propagation, and PageRank. It has imported a full amount of industrial and commercial data, built a massive entity relationship map, integrated retail data in the industry, and corporate boutique asset data internally, empowering multiple business scenarios such as corporate and retail, and output high-quality services (see Figure 3).
Figure 3 Data source of federated knowledge map
Complex Graph Analysis and Related Applications
Traditional graph analysis is mainly based on the technical solution of feature engineering, through statistical graph structural features, combined with downstream machine learning models, so as to complete the overall modeling. As the graph neural network algorithms of the GCN and GAT series become more mature, the industry can realize the vectorized representation of the knowledge graph in the actual business field, predict and mine the relationship that did not exist originally, and then apply it to subsequent marketing and risk control scenarios.
In terms of marketing activities, the node vectorization representation of the knowledge graph is used to seek the effect of spreading from point to surface. Based on the idea of LookLike, Fan Amplifier takes converted customer groups as seed customers, and selects target audiences who are very similar to seed customers as marketing objects through a certain evaluation algorithm, so as to achieve conversion amplification effect and significantly increase the average success rate of marketing activities.
In the application of the risk control field, relying on the complex relationship, comprehensively enrich the retail and public portraits, change the original view of the problem from the perspective of the individual to solve the problem from the perspective of the customer group, and analyze the equity between retail customers and public companies , transactions, events, etc., establish a risk control model, tap potential risk related groups, and explore risk transmission paths, thereby effectively assisting banks to avoid risks.
Comprehensive applications in the financial field
Wealth management and dialogue customer service are important applications of knowledge graph capabilities in the upper two scenarios of the financial industry. China Merchants Bank integrates the business capabilities of different customers, and effectively covers customers in different stages, industries and characteristics with differentiated and targeted technical services.
Wealth management is an important part of customer service. It is necessary to understand the actual demands of customers, find suitable asset management supply products, and maximize value through asset allocation and position adjustment. Among them, AI investment research ability is an important foundation of wealth management. The bottom layer of China Merchants Bank's AI investment research ability is based on a large amount of map data, integrating various AI technical capabilities such as public opinion analysis, research report analysis, and opinion generation, and calculating personality through the federal knowledge map. Make a reasonable allocation of wealth, form a user's personalized portrait label, and finally provide a reasonable user asset allocation.
For the dialogue customer service engine, China Merchants Bank will launch a smart investment assistant in 2021-AI small trick assistant. Technically, through field classification and slot identification, enter the preset service of intelligent financial advisors, so as to reasonably guide users to answer financial investment advisory tasks. Based on the knowledge of the knowledge center question and answer library, the financial customer service semantic understanding engine is trained to form semantic classification and semantic matching knowledge question and answer capabilities to solve user consultation problems. Combined with the characteristics of the wealth management scene, a large number of product knowledge graphs such as funds, wealth management, insurance, and gold have been built, and capabilities such as entity recognition, entity disambiguation, and semantic classification have been built. Finally, an investment advisory robot with knowledge graph + knowledge base Q&A + task-based Q&A was formed to assist account managers in assisting customers in financial management.
The Impact of Large Models on the Working Paradigm of Knowledge Mapping
From the construction of the traditional knowledge map to the application of the upper layer, it is necessary to divide the text task into various sub-task scenarios to solve, such as knowledge discovery, knowledge mining, knowledge representation, knowledge reasoning, knowledge application and other tasks, which involve unstructured data cleaning and Tasks such as extraction, word segmentation, semantic role labeling, entity extraction, relationship classification, entity disambiguation, semantic matching, graph query, and graph reasoning are different from the human complete end-to-end knowledge network construction process. This traditional method is destined to consume a lot of manpower and time to fine-tune various subtasks, and each task process needs to label a large number of high-quality fine-tuning data sets to form the fine-tuning task of the pre-trained small model of the scene class, but the subtask Errors in the interval will eventually pass on and affect the accuracy of the final application.
However, large language models like ChatGPT rely on large-scale parameters and high-quality human feedback mechanism learning to simulate humans well, allowing the model to show the ability of AI. Let us see that the traditional paradigm in the field of natural language processing and the scene-based fine-tuning method represented by Bert are no longer applicable. Instead, the large model can rely on its emergent ability and powerful common sense, reasoning, and interaction capabilities to process most of the NLP downstream applications based on a unified paradigm, and the generation effect is gradually close to the real world. It is even difficult for non-domain professionals to identify the real content. fake.
Large-scale language models have the potential to revolutionize the way we approach knowledge graphs. Knowledge graphs are powerful tools for representing complex knowledge structures and relationships, but require considerable effort to build and maintain. Large-scale language models can automate many of the tasks required to build and maintain knowledge graphs, such as semantic understanding tasks such as entity recognition, relation extraction, and classification matching. Under the impact of the big model, we have to think about whether it is possible for the knowledge map to realize a new unified work paradigm, organically integrate the knowledge associations stored in the knowledge map into the big model, and teach the big model to master the knowledge and reasoning ability of the map, so that To achieve a unified end-to-end working paradigm for downstream applications?
To address these challenges, we need to develop new techniques and tools to integrate large language models with knowledge graphs. One approach is to use natural language processing techniques to extract structured data from unstructured text generated by large language models, which can ensure that the information generated by large language models is accurately represented in the knowledge graph. Another method is to detect and correct errors in the generation of large-scale language models by developing technologies and combining knowledge graphs, and integrate factual results into large-scale language models to make the generation of large-scale language models more credible.
In summary, addressing the impact of large language models on knowledge graphs requires a combination of technical expertise, domain knowledge, and creativity. By developing new algorithms and tools and combining the power of large language models to create more powerful and accurate knowledge graphs, knowledge graphs can also be used to improve the accuracy of large language models in knowledge application and reasoning.
epilogue
A knowledge graph is an artificial intelligence-based technology for building a knowledge base and representing it as a graphical model with entities and relationships. In banking business, knowledge graph plays an important role. By establishing a knowledge map in the banking business field, banks can better understand customer needs, product information, market trends, etc., thereby improving business efficiency and customer satisfaction.
Knowledge graphs can help banks build intelligent customer service systems. In terms of customer service, knowledge graphs can help banks build natural language processing systems to provide more accurate and timely solutions by understanding and analyzing customer problems.
In addition, knowledge graphs can assist banks in risk management. The risk management of banking business needs to collect, integrate and analyze various information, including market information, customer information, assets and liabilities information, etc. Knowledge graphs can help banks integrate this information into a comprehensive risk management knowledge base to better identify and manage risks.
Not only that, the knowledge graph can also assist banks in product recommendation and cross-selling. Banks can analyze customer needs and historical transaction data through knowledge graphs, and then recommend more suitable products and services to customers based on these data, and at the same time improve customer value through cross-selling.
In short, the application prospect of knowledge graph in banking business is very broad. It can help banks better understand customer needs, improve business efficiency and customer satisfaction, and also help banks in risk management and product recommendation. With the support of generative large language models such as ChatGPT, how to combine with traditional knowledge graphs to better serve customers and continuously improve their own competitiveness is also worthy of continuous exploration and attention.
about the author
Li Jinlong, Director of Artificial Intelligence Lab, China Merchants Bank
Lead the team engaged in the research and development of artificial intelligence technology and its application in the field of intelligent finance. The science and technology project presided over won the first prize of the China Banking and Insurance Regulatory Commission, the second prize of the People's Bank of China Science and Technology Development Award twice, participated in the compilation of each issue of the CF40 "China Smart Finance Development Report", participated in more than ten academic papers in the field of artificial intelligence, and won the number of national patents ten items.
He Yaohan , Intelligent Science R&D Engineer, Artificial Intelligence Laboratory, China Merchants Bank
Graduated from the Department of Automation, Tsinghua University. Since joining China Merchants Bank, his main research directions include natural language processing, construction and application of knowledge graphs, etc., to build a product China Merchants Bank Smart Network, serving the marketing and risk control applications of the whole bank. Published many English papers.
Graduated from Harbin Institute of Technology with a master's degree, he is mainly engaged in the research of algorithms such as natural language processing, speech recognition, and pre-trained language models. He has participated in the application of intelligent dialogue engine algorithms such as China Merchants Bank's knowledge graph question answering system, and published many papers at domestic and foreign conferences.
