Natural language processing is more dependent on the acquisition of information sources, the team does not want to focus on the early corpus crawling, although this is very important for the results, but the main direction of the early development is to successfully run the whole framework out of the results.

The difficulty of obtaining information sources determines the cost and effort of the team in setting the product framework at the beginning of the project, and we do not want to spend too much manpower and material resources here. For the early stage, we can directly obtain the second information from the enterprise search and other secondary information acquisition websites for the purpose of convenience and rapidness, even if there are errors in the incomplete crawling data. After the product is ready for commercial use, we need to expand the scale and sign a commercial agreement to obtain it from the original news channel. In the way of drawing a picture of a tiger according to a cat, we crawl what data sources are used in existing enterprise search, sky-eye search and other secondary credit retrieval websites, and if they can’t, we certainly can’t either.

We must first make something, then apply for intellectual property protection as soon as possible, and outsource it to Wang Wen to apply for funds. Follow the trend to do data, already some credit information products are doing, must be able to crawl, allow crawling data source! If you can’t, neither can they. The wall falls apart.

The product is expected to make a recommendation system, recommend the enterprise, and retrieve the real contact object, understand its business and investment, key news, quick investigation

For example, industrial and commercial information network, qualification network, bidding network, financing network, tax network, legal information network and so on.

For example, corporate official websites, vertical websites, industry websites, recruitment websites, portals, map sites and other data that are more closely related to the company’s business.

The functions of empty number detection and intelligent recommendation help users reach the relevant persons in charge of the enterprise in all directions. It is not just a legal person hanging under the company. The pain points and problems to be solved-direct retrieval is not sure what it is like. After clicking in, you can learn more about how this tool can be implemented.

All of the above data platforms can be given Ming number, and all of them can support data export except trace.

The output of the result is product-oriented and incorporated into other tracks to participate in the competition, and the most direct results are software copyright, papers (low-level) and competitions (non-financial tracks), laying the foundation for future resumes.

From the perspective of curriculum vitae, we should strengthen our own application in machine learning, deep learning, neural network direction and learning experience, and reverse the project experience we need according to the purpose of our own acquisition. and several experiences had better be related, as their own characteristics, it is best to be associated with the project they applied for, and at present these experiences are still too data (text data). It doesn’t involve too much business data analysis (more participants in data science competitions) or financial data strategy construction. The next step is to consider changing the direction of the project and how to guide the project to be more commercialized (the formation of user flow). Look more at what other people’s internships and project experiences do)

Simple knowledge graph and neural network are different fields, so we must consider how to empower graph neural network to knowledge graph. Retrieval methods: 1. We have relevant reports on knowledge graph, but the technical route of application and interpretation is relatively limited. 2, search Wechat, the application of Zhihu Zhongtu neural network and general ideas (the general literature review in the opening stage needs to be completed)

Note: considering the influence of the project quantity, it is best to select the technical route is very mature, there are more ready-made codes and previous explanations in the network, but the differences caused by changing different application scenarios, objects or fields, such as transforming the application of graph neural network in social network into studying the “social” relationship network of graph neural network between enterprises to form a recommendation system.

It is related to whether the output of the project can maximize its effectiveness before you apply, taking into account the size of the personnel required by the project (such as the construction of front-end pages, front-end links, access and cleaning of crawlers, construction of internal models). Time required (expected, planned, output within half a year)

Using the recommendation system in social media for reference, it uses social network (constructing semantic network according to users’ click, purchase, access and other information data) to form intelligent recommendation and relationship retrieval. We replace the objects and study the investment, supply, competition, cooperation and other relationships between different enterprises. after clicking to enter, we can be transformed into the interface of emotional analysis to get the top 10 key news of the enterprise.

The establishment of laboratory opening, the establishment of a formal organization to open up the track, a clear feedback mechanism, including the use of both short-term and long-term, short-term competition as the medium and carrier.

The fire at the city gate affected the fish in the pond. The deterioration of the credit risk of related parties may be transmitted to the subjects with reasonable qualifications. Therefore, the research on the credit risk transmission of the related parties of the bond issuer becomes important and necessary.

The discovery and confirmation of related parties depends on the relationship information of the enterprise. Which relational information can reveal the related parties, whether the disclosed relational information can refer to the real related parties, and whether the deep relational links can be mined efficiently through database query, it is a problem that should be considered by all parties of data analysis, model construction and product development.

There are many kinds of relationships among enterprises, but not all of them will lead to the transmission effect of enterprise credit risk. Factors such as the strength of the same kind of relationship and the risk resistance of the enterprise itself need to be taken into consideration, otherwise it will lead to the mistransmission of a large range of normal enterprises. In addition, the lack of transmission case data, coupled with the high degree of abstraction of risk transmission, brings great challenges to model training.

Knowledge graph is a kind of graph-based data structure, which is a semantic network composed of nodes (entities) and edges (relationships). It is often used to analyze complex relationship problems. Therefore, we try to build an enterprise-based knowledge graph (hereinafter referred to as enterprise graph), that is, based on the graph database to display enterprise relationship information in the form of graph, in order to mine enterprise related parties.

At the level of covert related party mining, because the covert means of related parties are becoming more and more clever, mining is becoming more and more difficult. Through a large number of research and analysis, we find that problem enterprises generally have business or capital exchanges with hidden related parties for the purpose of benefit transportation and concealment or financial statement beautification. In this regard, we focus on the main supply chain and capital counterpart of the target enterprise, based on the enterprise graph, by considering the association pattern, the correlation clustering distance of all kinds of information of the associated entity, and so on, mining the graph relationship path that may have hidden association relationship.

However, not all relational data are helpful to the mining of related parties. If the simple calculation is carried out based on the above enterprise graph, the subsequent modeling is bound to be disturbed by useless relational information. Our approach is to draw lessons from the experience of experts and conduct double verification according to the historical data of the back test to screen out the important types of relationships that are helpful to the study of risk transmission.

Although the assets and operation of the real estate company and its travel company are independent and market-oriented, in fact, the controlling shareholders are often inextricably linked with listed companies, such as business transactions, financing credit endorsement and so on. Problems on one side are likely to affect the other. And in reality, there is also a view that the cultural travel real estate with large investment and long return constitutes a great test for the overall cash flow of the parent holding company.