Financial event atlas construction method, device and equipment

Abstract

The invention discloses a financial event atlas construction method, device and equipment. Specifically, when the construction method of the financial event atlas provided by the embodiment of the invention is executed, firstly, an abstract association relationship of financial news public opinions can be obtained; the known events and the existing events in the abstract incidence relation are combined through correlation calculation to obtain a combined link, the causal relation between the known events and the existing events in the combined link is extracted, and the existing events are the events existing in the retrieval database. And then extracting elements from financial news public opinions by the new events in the abstract association relationship to obtain element information, and integrating the element information to obtain specific event nodes. And then obtaining an event association object of the search data. And finally, inputting the causal relationship, the specific event node and the event association object into a graph database to generate a knowledge graph. The financial event atlas provided by the invention improves the precision of financial event analysis in the current market.

Description

Technical Field

[0001] The present application relates to the technical field of relationship graph construction, and in particular to a method, device and equipment for constructing a financial event graph. Background Art

[0002] Changes in financial markets are mainly affected by changes in participants' emotions, financial policies, etc. Financial practitioners use statistical inference methods or function fitting methods to describe and infer financial markets from financial-related information. For financial markets, financial data is usually the performance after a financial event occurs, and financial data is a way to describe financial markets.

[0003] However, in addition to event descriptions obtained from certain structured data (for example, a surge in an index or a sharp drop in futures trading volume), financial events in the financial market are mostly described in text form (for example, a policy issued by the state or a sudden resignation of a corporate executive). These text descriptions cannot be directly processed using conventional methods. At present, the most common practice in the industry is to classify financial news, public opinion, announcements and other data into events, and then use the statistical values ​​of the classification results as analysis indicators. This approach is relatively coarse-grained and it is difficult to make reasonable judgments on the evolution of events.

[0004] Therefore, how to provide a more fine-grained way to describe financial events is a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] Based on the above problems, the present application provides a method, device and equipment for constructing a financial event graph to improve the usability of icons in the terminal device interface for different users. The present application embodiment discloses the following technical solutions:

[0006] A method for constructing a financial event graph, the method further comprising:

[0007] Obtain abstract correlations of financial news and public opinion;

[0008] Using correlation calculation, the known events in the abstract association relationship are merged with the existing events to obtain a merged link, and the causal relationship in the merged link is extracted, wherein the causal relationship is an abstract relationship between the known event and the existing event; the existing event is an event existing in the search database;

[0009] Extracting elements from the financial news public opinion according to the new event in the abstract association relationship to obtain element information, and integrating the element information to obtain a specific event node;

[0010] Get the event-related target of the search data;

[0011] The causal relationship, the specific event nodes and the event-related targets are input into a graph database to generate a knowledge graph.

[0012] In some possible implementations, the step of obtaining the abstract association relationship of financial news public opinion includes:

[0013] Segment processing is performed on the financial news public opinion to obtain a first target sentence;

[0014] Inputting the first target sentence into a causal relationship processing model to obtain a plurality of causal triples and a plurality of probability scores corresponding to the plurality of causal triples;

[0015] The causal triples with probability scores greater than the score threshold are regarded as candidate triples;

[0016] The candidate triples whose event descriptions meet the length rule are used as the abstract association relationships.

[0017] In some possible implementations, the method further includes:

[0018] Calculating the correlation between each causal event in the abstract association relationship and the existing event;

[0019] The causal event whose correlation is greater than a correlation threshold is regarded as the known event; and the causal event whose correlation is less than or equal to the correlation threshold is regarded as the new event.

[0020] In some possible implementations, the method further includes:

[0021] The known events that cannot be merged with the existing events are added with argument definitions by generating new event modules.

[0022] In some possible implementations, extracting elements from the financial news public opinion according to the new event in the abstract association relationship to obtain element information includes:

[0023] Sentence processing is performed on the new event to obtain a second target sentence;

[0024] Inputting the second target sentence into an argument extraction model to obtain a plurality of candidate argument groups;

[0025] Acquire multiple argument confidences corresponding to the multiple candidate argument groups, and use the candidate argument group whose argument confidence is greater than a confidence threshold as the target argument group;

[0026] The element information is obtained by combining each target argument tuple using a greedy algorithm.

[0027] In some possible implementations, obtaining the event-related target of the search data includes:

[0028] Identify the event name and resource of the search data;

[0029] Grouping the event name and the resource according to time and user unique identifier to obtain multiple candidate tuples;

[0030] Count the occurrence frequency of each candidate bigram;

[0031] Taking the candidate bigrams whose occurrence frequency is higher than the frequency threshold as target bigrams;

[0032] The target binary group is filtered by cumulative information entropy to obtain the event-related target.

[0033] A device for constructing a financial event graph, the device comprising:

[0034] The first acquisition unit is used to acquire the abstract correlation relationship of financial news and public opinion;

[0035] A calculation extraction unit is used to merge the known events in the abstract association relationship with the existing events to obtain a merged link by using correlation calculation, and to extract the causal relationship in the merged link, wherein the causal relationship is an abstract relationship between the known event and the existing event; the existing event is an event existing in the search database;

[0036] An element extraction unit, configured to extract elements from the financial news public opinion according to the new event in the abstract association relationship to obtain element information, and integrate the element information to obtain a specific event node;

[0037] A second acquisition unit, used to acquire event-related targets of the search data;

[0038] A generating unit is used to input the causal relationship, the specific event node and the event-related target into a graph database to generate a knowledge graph.

[0039] In some possible implementations, the device further includes:

[0040] A first sentence processing unit, configured to perform sentence processing on the financial news public opinion to obtain a first target sentence;

[0041] A first input unit, configured to input the first target sentence into a causal relationship processing model to obtain a plurality of causal triples and a plurality of probability scores corresponding to the plurality of causal triples;

[0042] A first setting unit is used to take the causal triples whose probability scores are greater than the score threshold as candidate triples;

[0043] The second setting unit is used to take the candidate triples whose event descriptions meet the length rule as the abstract association relationship.

[0044] An electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for constructing a financial event map as described above is implemented.

[0045] A vehicle comprises a control module, wherein the control module is used to execute the method for constructing a financial event graph as described above.

[0046] Compared with the prior art, the present application has the following beneficial effects: the present application discloses a method, device and apparatus for constructing a financial event map. Specifically, when executing the method for constructing a financial event map provided in the embodiment of the present application, the disease level based on the medical code division can be obtained first. Then, a medical map is constructed according to the disease name and the drug name, and the medical map includes a first entity layer, and the first entity layer includes the disease name and the drug name. Then, according to the correspondence between the disease name and the disease level in the first entity layer, the disease level and the medical map are integrated to obtain a first integrated map. Then, according to the insurance name and the disease name, an underwriting and claims map is constructed, and the underwriting and claims map includes a second entity layer, and the second entity layer includes the disease name and the insurance name. Finally, according to the correspondence between the disease name in the second entity layer and the first integrated map, the first integrated map and the underwriting and claims map are integrated to obtain a second integrated map. The financial event map of the present application can describe financial events in a more fine-grained manner. At the same time, the financial event map of the present application can organically combine the association relationship of events in the form of a graph network. The financial event map of the present application records the general rules of event changes and the specific content of financial events, so that the financial event map can effectively support the attribution and expected inference of financial events. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0048] Figure 1 A flowchart of a method for constructing a financial event graph provided in an embodiment of the present application;

[0049] Figure 2 A schematic diagram of the structure of a device for constructing a financial event graph provided in an embodiment of the present application. DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0051] To facilitate understanding of the technical solutions provided by the embodiments of the present application, the background technology involved in the embodiments of the present application will be described below.

[0052] Changes in financial markets are mainly affected by changes in participants' emotions, financial policies, etc. Financial practitioners use statistical inference methods or function fitting methods to describe and infer financial markets from financial-related information. For financial markets, financial data is usually the performance after a financial event occurs, and financial data is a way to describe financial markets.

[0053] Among them, financial markets refer to places and platforms for buying and selling financial assets, including stock markets, bond markets, foreign exchange markets, futures markets, etc. The main function of financial markets is to provide investors with financing and investment opportunities and promote the flow and allocation of funds. In financial markets, individuals and institutions can realize investment appreciation or financing needs by buying and selling financial assets.

[0054] The characteristics of financial markets include high liquidity, diversified risks, transparent prices, and rich information. Different types of financial markets have different sizes and trading methods. Investors can choose the market that suits them to trade according to their investment goals and risk tolerance.

[0055] The financial market plays an important role in economic development. It provides services such as fund raising, resource allocation, and risk management, and promotes the effective operation of the economy. At the same time, the financial market also faces some challenges, such as market volatility, information asymmetry, and regulatory risks. Supervision and risk management measures are needed to ensure the normal operation of the market.

[0056] However, in addition to event descriptions obtained from certain structured data (for example, a surge in an index or a sharp drop in futures trading volume), financial events in the financial market are mostly described in text form (for example, a policy issued by the state or a sudden resignation of a corporate executive). These text descriptions cannot be directly processed using conventional methods. At present, the most common practice in the industry is to classify financial news, public opinion, announcements and other data into events, and then use the statistical values ​​of the classification results as analysis indicators. This approach is relatively coarse-grained and it is difficult to make reasonable judgments on the evolution of events.

[0057] In order to solve this problem, a method, device and equipment for constructing a financial event graph are provided in an embodiment of the present application. In the present application, an abstract correlation relationship of financial news public opinion is first obtained. Then, the known events in the abstract correlation relationship are merged with the existing events to obtain a merged link by correlation calculation, and the causal relationship in the merged link is extracted, wherein the causal relationship is an abstract relationship between the known event and the existing event; the existing event is an event existing in the retrieval database. Next, according to the new event in the abstract correlation relationship, the element information is extracted from the financial news public opinion, and the element information is integrated to obtain a specific event node. After obtaining the specific event node, it is also necessary to obtain the event-related target of the search data. Finally, the causal relationship, the specific event node and the event-related target are input into the graph database to generate a knowledge graph. The financial event graph of the present application can describe financial events in a more fine-grained manner. At the same time, the financial event graph of the present application can organically combine the correlation relationship of events in the form of a graph network. The financial event map of the present application records the general rules of event changes and the specific content of financial events, so that the financial event map can effectively support the attribution and expected inference of financial events, so that a lot of manpower and time can be saved when determining whether the customer meets the conditions for underwriting and claims settlement.

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0059] See also Figure 1 , which is a flow chart of a method for constructing a financial event graph provided in an embodiment of the present application, such as Figure 1 As shown, the method for constructing the financial event graph may include steps S101-S105:

[0060] S101: Obtain the abstract correlation relationship of financial news public opinion.

[0061] In order to construct a financial event graph, the financial event graph construction system can first obtain the abstract correlation relationship of financial news and public opinion.

[0062] Among them, news and public opinion refers to the collection of attitudes, opinions and emotions in society towards a certain event or issue. It reflects the public's attention to a specific event, the public opinion trend and the possible impact on the relevant parties. News and public opinion can be obtained through the analysis and monitoring of media reports, social media content, online discussions and other information. In the financial field, news and public opinion have an important influence. Financial markets are often affected by the emotions and expectations of market participants, and news and public opinion plays an important role in shaping the emotions and expectations of participants. When an event or news is widely reported or discussed, market participants are often affected by it, leading to market fluctuations.

[0063] The development of modern technology has made the analysis and monitoring of news and public opinion more sophisticated and comprehensive. With the help of technologies such as natural language processing, machine learning and data mining, large amounts of news text and social media data can be automatically processed and analyzed to obtain key information and insights. These technical tools can help financial institutions and practitioners better understand and respond to the impact of news and public opinion on the market.

[0064] In some possible implementations, the abstract association relationship is displayed in the form of a triple.

[0065] In some possible implementations, the abstract association relationship of obtaining financial news public opinion includes A1-A4:

[0066] A1: Segment the financial news public opinion to obtain a first target sentence.

[0067] When obtaining the abstract correlation relationship of financial news public opinion, it is first necessary to segment the financial news public opinion into sentences to obtain the first target sentence.

[0068] A2: Inputting the first target sentence into a causal relationship processing model to obtain a plurality of causal triples and a plurality of probability scores corresponding to the plurality of causal triples.

[0069] After obtaining the first target sentence, it is also necessary to input the first target sentence into the causal relationship processing model to obtain multiple causal triples and multiple probability scores corresponding to the multiple causal triples.

[0070] Among them, the causal relationship processing model is a tool or method for analyzing and understanding causal relationships. It can help users determine the causal relationship between events, behaviors or variables, as well as the mechanism of action between them. The causal relationship processing model can identify causal relationships in text and generate corresponding causal triples (such as causal events, associated words, and result events). At the same time, these models can usually provide probability scores or confidence levels for the generated causal triples.

[0071] When training a causal relationship processing model, you can use labeled causal relationship data for supervised learning. The model learns the mapping relationship from input text to output causal triples to predict causal relationships as accurately as possible and assign them probability scores.

[0072] It should be noted that different causal relationship processing models may adopt different architectures and algorithms, so the way of providing probability scores or confidences may be different. Depending on the selected model and task requirements, an appropriate model can be selected and the probability scores of causal triples can be evaluated based on its output.

[0073] The following are several common causal relationship processing models: ① Traditional statistical models: Traditional statistical models (such as linear regression, logistic regression, etc.) can be used to analyze causal relationships. By controlling other variables, we can derive the degree of influence of one variable on another variable. ② Experimental design model: The experimental design model determines the causal relationship by comparing the experimental group and the control group, controlling and excluding other interfering factors. For example, a randomized controlled experiment can be used to evaluate the therapeutic effect of a drug on disease symptoms. ③ Causal graph model: The causal graph model is a method used to describe and visualize causal relationships. It uses arrows to represent the causal relationship between variables and can show direct and indirect causal paths. The causal graph model helps to better understand complex causal networks. ④ Structural equation model: The structural equation model (SEM) combines measurement models and structural models to evaluate the causal relationship between multiple variables. It can quantify the direct and indirect effects between variables and provide a comprehensive model to explain the data. These models are used in different contexts and research fields. The selection of an appropriate causal relationship processing model depends on the nature of the specific problem, the type of available data, and the purpose of the research.

[0074] A3: The causal triples whose probability scores are greater than the score threshold are taken as candidate triples.

[0075] After obtaining the causal triples and the probability scores corresponding to the causal triples, the financial event graph construction system can take the causal triples with probability scores greater than the score threshold among multiple causal triples as candidate triples.

[0076] In some possible implementations, the score threshold may be, but is not limited to, 0.75. The score threshold in this application may be adjusted according to specific circumstances, and this application does not impose any specific restrictions on the score threshold.

[0077] A4: The candidate triples whose event descriptions meet the length rule are used as the abstract association relations.

[0078] After obtaining the candidate triples, the financial event graph construction system can take the candidate triples whose event description lengths conform to the length rule as abstract association relationships.

[0079] The length of the event description can be determined according to specific needs and application scenarios. Generally, the length of the event description can range from a few words to hundreds of words.

[0080] For a brief event description, it may only contain the main key information and elements, such as the subject, time, location and basic background of the event. Such a description is usually used to quickly understand the overview of the event and provide a concise and clear summary.

[0081] For more detailed event descriptions, more details and contextual information may be included to fully understand the occurrence and impact of the event. These descriptions may involve the cause, process, participants, motivation, consequences, etc. Such descriptions are usually used for in-depth analysis and research of events, providing a more comprehensive perspective and insight.

[0082] In some possible implementations, compliance with the length rule may include, but is not limited to, event descriptions having a length greater than 3 characters and less than 20 characters complying with the length rule.

[0083] In some possible implementations, the candidate triples whose event descriptions do not comply with the length rule (ie, event descriptions with a length less than 3 words and event descriptions with a length greater than 20 words do not comply with the length rule) need to be deleted.

[0084] S102: Using correlation calculation, merge the known events in the abstract association relationship with the existing events to obtain a merged link, and extract the causal relationship in the merged link, wherein the causal relationship is an abstract relationship between the known event and the existing event; the existing event is an event existing in the retrieval database.

[0085] After obtaining the abstract correlation of financial news public opinion, the known events in the abstract correlation can be merged with the existing events to obtain the merged link by using correlation calculation, and the causal relationship between the known events and the existing events in the merged link can be extracted, where the existing events are events existing in the search database. For example, if there is a relationship in the library that is "Project Winning---(Causal Relationship)--> Enterprise Revenue Increase", if a new relationship "Revenue Growth--(Causal Relationship)--→Stock Rise" is found from financial news public opinion. By calculating the correlation, it is found that the correlation between the two events "Enterprise Revenue Increase" and "Revenue Growth" is very high. Therefore, these two events can be merged into the same one. Finally, the causal link "Project Winning---(Causal Relationship)--> Enterprise Revenue Increase--(Causal Relationship)--→Stock Rise" is constructed in the library.

[0086] Among them, correlation calculation is a method used to measure the similarity or correlation between documents or data. In database retrieval and search engines,

[0087] In some possible implementations, the relevance calculation method includes but is not limited to the following methods: ① Cosine Similarity: Cosine similarity measures the similarity between vectors by calculating the angle between them. Documents are represented as vectors and the cosine similarity between them is calculated. The value is between -1 and 1, and the closer to 1, the more similar. ② TF-IDF (Term Frequency-Inverse Document Frequency): TF-IDF is an indicator that measures the importance of a word in a document collection. It is based on two factors: term frequency (the number of times a word appears in a document) and inverse document frequency (the rarity of a word in the entire document collection). By calculating the TF-IDF value of each word in the document, relevance sorting and matching can be performed. ③ BM25 (Best Matching 25): BM25 is a popular information retrieval algorithm used to calculate the relevance between documents and queries. It takes into account factors such as term frequency, document length, and inverse document frequency of query terms, and obtains the final relevance score by weighted calculation of these factors. ④ Jaccard Similarity: Jaccard Similarity is used to compare the similarity between two collections. It measures similarity by calculating the size of the intersection of two sets divided by the size of the union. The value is between 0 and 1, and the closer it is to 1, the more similar it is. These correlation calculation methods can be used to select appropriate methods for calculation according to specific application scenarios and data types. In database retrieval and search engines, multiple methods are often used in combination to comprehensively evaluate and rank the relevance of documents to provide more accurate search results.

[0088] In some possible implementations, the retrieval database includes a vector database and Elasticsearch. The vector database is a database specifically used to store, manage, and retrieve vector data. It treats vectors as the main data type and can effectively perform similarity search and matching. The vector database uses metrics such as cosine similarity to calculate the similarity between vectors and optimizes search performance through indexing and efficient query algorithms. This makes the vector database very suitable for application scenarios that require large-scale similarity search, such as face recognition, image search, recommendation systems, etc.

[0089] Elasticsearch is a full-text search and analysis engine with powerful search and query capabilities. It can be used to store and index various types of data, including text, structured data, and vector data. Although Elasticsearch is not specifically designed to process vector data, vector indexing and similarity search can be implemented through extension plug-ins (such as the Elasticsearch VectorScoring Plugin) or custom scripts. Elasticsearch's distributed architecture and high scalability make it an excellent choice for processing large amounts of text data and complex queries.

[0090] In some possible implementations, the method further includes B1-B2:

[0091] B1: Calculate the correlation between each causal event in the abstract association relationship and the existing event.

[0092] In order to obtain known events and new events, the financial event graph construction system first needs to calculate the correlation between each causal event and existing events in the abstract correlation relationship of financial news public opinion.

[0093] In some possible implementations, a similarity or correlation measurement method (such as cosine similarity, Pearson correlation coefficient, etc.) can be used to calculate the similarity or correlation score between each causal event and the existing event. The specific method can be selected according to the data type and requirements.

[0094] B2: The causal event whose correlation is greater than the correlation threshold is regarded as the known event; and the causal event whose correlation is less than or equal to the correlation threshold is regarded as the new event.

[0095] After calculating the correlation between each causal event and the existing events in the abstract association relationship, the financial event map construction system can regard the causal events whose correlation is greater than the relevant threshold as known events; and regard the causal events whose correlation is less than or equal to the relevant threshold as new events.

[0096] In some possible implementations, the relevant threshold can be determined according to the actual situation. If a greater recall is expected, the threshold can be appropriately lowered, for example, to 0.4. If a better precision is desired, the threshold can be increased, for example, to 0.8.

[0097] In some possible implementations, the method further includes:

[0098] The known events that cannot be merged with the existing events are added with argument definitions by generating new event modules.

[0099] If you encounter an event that cannot be merged, you can add argument definitions by generating new events. This can be used to expand the model labels in the next iteration of the event feature extraction model.

[0100] In some possible implementations, the following steps can be used to generate a new event module to add argument definitions: (1) Define a new event: Based on the existing event elements and related information, choose a clear and descriptive name to define the new event. (2) Determine arguments: Analyze the relationship between existing events and elements, and determine the arguments required for the new event. Arguments are key elements or attributes of an event, such as participants, time, and location. (3) Add argument definitions: Add definitions for each argument in the new event. The definition should clearly and accurately describe the meaning and function of the argument. (4) Data annotation: Use manual annotation methods to add the arguments of the new event as labels to the training data. You can match the existing event text data with the definition of the new event, and then annotate the arguments of the new event. (5) Model training: Retrain the event element extraction model using the expanded annotated data. During the model training process, the model will learn the argument definition and recognition method of the new event. (6) Iterative testing and improvement: Use the newly trained model for testing and evaluation, and adjust and improve it based on actual results. This process is repeated until the model can accurately extract event elements, including the newly added arguments.

[0101] In this way, the ability of the event element extraction model can be gradually improved so that it can recognize and extract more types of event elements.

[0102] S103: Extracting elements from the financial news public opinion according to the new event in the abstract association relationship to obtain element information, and integrating the element information to obtain a specific event node.

[0103] After obtaining a merged link by merging known events with existing events in the abstract association relationship using correlation calculation, and extracting the causal relationship in the merged link, the financial event graph construction system can extract element information from financial news and public opinion based on new events in the abstract association relationship, and integrate the element information to obtain specific event nodes.

[0104] In some possible implementations, extracting elements from the financial news public opinion according to the new event in the abstract association relationship to obtain element information includes C1-C4:

[0105] C1: Segment the new event to obtain a second target sentence.

[0106] In order to extract elements from financial news and public opinion and obtain element information, the financial event graph construction system first needs to sentence the new events in the abstract association relationship of financial news and public opinion to obtain the second target sentence.

[0107] C2: Input the second target sentence into the argument extraction model to obtain multiple candidate argument groups.

[0108] After the new event is processed into sentences to obtain the second target sentence, the financial event graph construction system can input the second target sentence into the argument extraction model to obtain multiple candidate argument groups.

[0109] Among them, argument extraction is a task in natural language processing, which aims to identify arguments related to a specific predicate from a given text. The following is a common argument extraction model framework: ① Feature extraction: First, useful features need to be extracted from the text to represent each word or phrase. These features can include part of speech, syntactic relations, dependencies, contextual information, etc. A common method is to use pre-trained word vectors (such as Word2Vec, GloVe) to represent words. ② Sequence labeling model: The argument extraction problem is modeled as a sequence labeling problem, which can be modeled using models such as conditional random fields (CRF) or recurrent neural networks (such as long short-term memory networks LSTM). The input of the model is a text sequence after feature extraction, and the output is a label for each word, indicating whether the word is an argument and the argument type it belongs to. ③ Annotation data preparation: In order to train the argument extraction model, corpus data with associated predicates needs to be annotated. Each sentence in the corpus needs to be annotated with arguments related to the target predicate and specify its argument type. This usually requires manual annotation or the use of existing annotated datasets. ④Model training and optimization: Use labeled data to train the argument extraction model. Adjust the model's parameters and weights to improve the model's performance by maximizing predefined performance indicators (such as accuracy, recall, F1 value, etc.). The model can be trained using the backpropagation algorithm or other optimization algorithms. ⑤Inference and evaluation: After training, the model can be used to infer new texts and identify arguments related to the predicate. The performance of the model can be evaluated by comparing it with manually labeled or test data sets and calculating indicators such as the model's accuracy, recall, and F1 value.

[0110] It should be noted that implementing a complete argument extraction model requires comprehensive consideration of multiple aspects such as feature selection, model architecture, labeled data preparation, and model training, and proper debugging and optimization. At the same time, the performance of the model is also limited by the quality and scale of the available labeled data.

[0111] C3: Obtain multiple argument confidences corresponding to the multiple candidate argument groups, and use the candidate argument group whose argument confidence is greater than the confidence threshold as the target argument group.

[0112] After inputting the second target sentence into the argument extraction model to obtain multiple candidate argument groups, the financial event graph construction system can obtain multiple argument confidences corresponding to the multiple candidate argument groups, and take the candidate argument group whose argument confidence is greater than the confidence threshold as the target argument group.

[0113] In some possible implementations, the confidence threshold may be, but is not limited to, 0.5 or 0.6, which is determined according to the specific scenario. The confidence threshold in this application may be adjusted according to the specific situation, and this application does not impose any specific restrictions on the confidence threshold.

[0114] In some possible implementations, the following methods can be used to estimate the confidence of arguments: ① Predicted probability scores: Some machine learning models (such as logistic regression and support vector machines) provide the ability to predict probability scores. These probabilities represent the model's confidence in each argument category. You can use these probabilities as an approximation of the confidence. ② Post-processing threshold setting: By applying a threshold to the score output by the model, you can filter out arguments with high confidence. A higher score may mean that the model is more confident in predicting the argument. You can set the threshold according to specific needs and select arguments with higher confidence. ③ Model-based uncertainty estimation: Some models can estimate the uncertainty of their predictions through ensemble methods (such as Bootstrap, Dropout) or Bayesian inference. By considering the uncertainty information of the model, the confidence of the argument can be better evaluated.

[0115] C4: Use a greedy algorithm to combine each target tuple to obtain the element information.

[0116] After obtaining the target theory tuple, the financial event graph construction system can use a greedy algorithm to combine each target theory tuple to obtain element information.

[0117] An effective approach is to use a greedy algorithm to combine the complete argument information. A greedy algorithm is a heuristic algorithm that gradually builds a solution by selecting the best decision at each step.

[0118] S104: Acquire event-related targets of search data.

[0119] After obtaining the specific event node, the financial event graph construction system can obtain the event-related targets of the search data.

[0120] Among them, event-related targets refer to specific objects, entities or assets related to a specific event or activity. These targets usually have a certain correlation with the event, and may be participants, influencing factors, objects or targets of the event. For example, in the financial field, an event-related target can be a company's stock, commodity futures, foreign exchange and other financial products. If there is an important announcement or financial statement data release about the company, the prices of these financial products may change, so these financial products are associated with this event. Event-related targets are specific objects, entities or assets related to an event in the context of a certain event, which are used to describe the impact, dependence or relationship of the event on the related objects. This correlation can be used for analysis, prediction and decision-making.

[0121] In some possible implementations, the event-related targets of the acquired search data include D1-D5:

[0122] D1: Identify the event name and resource of the search data.

[0123] In order to obtain the event-related targets of the search data, the financial event graph construction system first needs to identify the event name and resource (asset) of each search data.

[0124] In some possible implementations, natural language processing (NLP) techniques and models can be used to identify event names and resources of search data. That is, deep learning-based NLP models, such as BERT or GPT, can be used to perform semantic understanding and information extraction on queries.

[0125] D2: Group the event name and the resource according to time and user unique identifier to obtain multiple candidate tuples.

[0126] In order to obtain event-related targets of search data, the financial event graph construction system also needs to group event names and resources according to time and user unique identifiers to obtain multiple candidate tuples.

[0127] Among them, the user ID is a unique identifier assigned to each user by the system or platform, which is used to identify the user and manage user information.

[0128] D3: Count the occurrence frequency of each candidate bigram.

[0129] After obtaining the candidate binary groups, the financial event graph construction system can count the occurrence frequency of each candidate binary group.

[0130] D4: taking the candidate bigrams whose occurrence frequency is higher than the frequency threshold as target bigrams.

[0131] After obtaining the candidate bigrams, the financial event graph construction system can use the candidate bigrams whose occurrence frequency is higher than the frequency threshold as target bigrams.

[0132] In some possible implementations, the frequency threshold may be, but is not limited to, a time slice determined by day, and the frequency is generally greater than 50. The frequency threshold in the present application may be adjusted according to specific circumstances, and the present application does not impose specific restrictions on the frequency threshold.

[0133] D5: Perform cumulative information entropy filtering on the target binary group to obtain the event-related target.

[0134] After obtaining the target binary tuple, the financial event graph construction system can filter the target binary tuple based on the accumulated information entropy to obtain the event-related targets.

[0135] S105: Input the causal relationship, the specific event node, and the event-related target into a graph database to generate a knowledge graph.

[0136] After obtaining the causal relationship, specific event nodes and event-related targets, the financial event graph construction system can input the causal relationship, specific event nodes and event-related targets into the graph database to generate a knowledge graph.

[0137] Among them, the graph database is a database system designed for storing and processing graph-structured data. It is different from traditional relational databases or document databases in that it organizes and represents data in the form of a graph, where nodes represent entities and edges represent the relationships between nodes. The core of the graph database is the graph model, which consists of nodes and edges. Nodes represent entities or objects, such as people, places, items, etc., while edges represent the relationships between nodes. Both nodes and edges can contain attributes for storing more detailed information. Common graph database products include Neo4j, Amazon Neptune, JanusGraph, etc. These databases provide rich functions and APIs, allowing developers to easily process and manipulate graph data.

[0138] Based on the contents of S101-S105, it can be known that the abstract correlation relationship of financial news public opinion is obtained, and the known events in the abstract correlation relationship are merged with the existing events to obtain a merged link by using correlation calculation, and then the causal relationship in the merged link is extracted, wherein the causal relationship is the abstract relationship between the known event and the existing event; the existing event is the event existing in the retrieval database. Then, according to the new event in the abstract correlation relationship, the element information is extracted from the financial news public opinion, and the element information is integrated to obtain a specific event node. Then, the event correlation target of the search data is obtained, and finally the causal relationship, the specific event node and the event correlation target are input into the graph database to generate a knowledge graph. The financial event graph of this application can describe financial events in a more fine-grained manner. At the same time, the financial event graph of this application can organically combine the correlation relationship of events in the form of a graph network. The financial event graph of this application records the general laws of event changes and the specific content of financial events, so that the financial event graph can effectively support the attribution and expected inference of financial events.

[0139] See also Figure 2 , which is a schematic diagram of the structure of a device for constructing a financial event graph provided in an embodiment of the present application. Figure 2 As shown, the construction device of the financial event map includes:

[0140] The first acquisition unit 201 is used to acquire the abstract correlation relationship of financial news public opinion.

[0141] Among them, news and public opinion refers to the collection of attitudes, opinions and emotions in society towards a certain event or issue. It reflects the public's attention to a specific event, the public opinion trend and the possible impact on the relevant parties. News and public opinion can be obtained through the analysis and monitoring of media reports, social media content, online discussions and other information. In the financial field, news and public opinion have an important influence. Financial markets are often affected by the emotions and expectations of market participants, and news and public opinion plays an important role in shaping the emotions and expectations of participants. When an event or news is widely reported or discussed, market participants are often affected by it, leading to market fluctuations.

[0142] The development of modern technology has made the analysis and monitoring of news and public opinion more sophisticated and comprehensive. With the help of technologies such as natural language processing, machine learning and data mining, large amounts of news text and social media data can be automatically processed and analyzed to obtain key information and insights. These technical tools can help financial institutions and practitioners better understand and respond to the impact of news and public opinion on the market.

[0143] In some possible implementations, the abstract association relationship is displayed in the form of a triple.

[0144] The calculation extraction unit 202 is used to merge the known events in the abstract association relationship with the existing events to obtain a merged link by using correlation calculation, and to extract the causal relationship in the merged link, wherein the causal relationship is an abstract relationship between the known events and the existing events; and the existing events are events existing in the retrieval database.

[0145] For example, if there is a relationship in the database that is "Project winning --- (causal relationship) --> corporate revenue increase", if a new relationship "Revenue growth --- (causal relationship) ---> stock price increase" is discovered from financial news and public opinion. By calculating the correlation, it is found that the correlation between the two events "Corporate revenue increase" and "Revenue growth" is very high. Therefore, these two events can be merged into one. Finally, the causal link "Project winning --- (causal relationship) --> corporate revenue increase --- (causal relationship) ---> stock price increase" is constructed in the database.

[0146] Among them, correlation calculation is a method used to measure the similarity or correlation between documents or data. In database retrieval and search engines,

[0147] In some possible implementations, the relevance calculation method includes but is not limited to the following methods: ① Cosine Similarity: Cosine similarity measures the similarity between vectors by calculating the angle between them. Documents are represented as vectors and the cosine similarity between them is calculated. The value is between -1 and 1, and the closer to 1, the more similar. ② TF-IDF (Term Frequency-Inverse Document Frequency): TF-IDF is an indicator that measures the importance of a word in a document collection. It is based on two factors: term frequency (the number of times a word appears in a document) and inverse document frequency (the rarity of a word in the entire document collection). By calculating the TF-IDF value of each word in the document, relevance sorting and matching can be performed. ③ BM25 (Best Matching 25): BM25 is a popular information retrieval algorithm used to calculate the relevance between documents and queries. It takes into account factors such as term frequency, document length, and inverse document frequency of query terms, and obtains the final relevance score by weighted calculation of these factors. ④ Jaccard Similarity: Jaccard Similarity is used to compare the similarity between two collections. It measures similarity by calculating the size of the intersection of two sets divided by the size of the union. The value is between 0 and 1, and the closer it is to 1, the more similar it is. These correlation calculation methods can be used to select appropriate methods for calculation according to specific application scenarios and data types. In database retrieval and search engines, multiple methods are often used in combination to comprehensively evaluate and rank the relevance of documents to provide more accurate search results.

[0148] In some possible implementations, the retrieval database includes a vector database and Elasticsearch. The vector database is a database specifically used to store, manage, and retrieve vector data. It treats vectors as the main data type and can effectively perform similarity search and matching. The vector database uses metrics such as cosine similarity to calculate the similarity between vectors and optimizes search performance through indexing and efficient query algorithms. This makes the vector database very suitable for application scenarios that require large-scale similarity search, such as face recognition, image search, recommendation systems, etc.

[0149] Elasticsearch is a full-text search and analysis engine with powerful search and query capabilities. It can be used to store and index various types of data, including text, structured data, and vector data. Although Elasticsearch is not specifically designed to process vector data, vector indexing and similarity search can be implemented through extension plug-ins (such as the Elasticsearch VectorScoring Plugin) or custom scripts. Elasticsearch's distributed architecture and high scalability make it an excellent choice for processing large amounts of text data and complex queries.

[0150] The element extraction unit 203 is used to extract elements from the financial news public opinion according to the new event in the abstract association relationship to obtain element information, and integrate the element information to obtain a specific event node.

[0151] The second acquisition unit 204 is used to acquire the event-related target of the search data.

[0152] Among them, event-related targets refer to specific objects, entities or assets related to a specific event or activity. These targets usually have a certain correlation with the event, and may be participants, influencing factors, objects or targets of the event. For example, in the financial field, an event-related target can be a company's stock, commodity futures, foreign exchange and other financial products. If there is an important announcement or financial statement data release about the company, the prices of these financial products may change, so these financial products are associated with this event. Event-related targets are specific objects, entities or assets related to an event in the context of a certain event, which are used to describe the impact, dependence or relationship of the event on the related objects. This correlation can be used for analysis, prediction and decision-making.

[0153] The generation unit 205 is used to input the causal relationship, the specific event node and the event-related target into the graph database to generate a knowledge graph.

[0154] Among them, the graph database is a database system designed for storing and processing graph-structured data. It is different from traditional relational databases or document databases in that it organizes and represents data in the form of a graph, where nodes represent entities and edges represent the relationships between nodes. The core of the graph database is the graph model, which consists of nodes and edges. Nodes represent entities or objects, such as people, places, items, etc., while edges represent the relationships between nodes. Both nodes and edges can contain attributes for storing more detailed information. Common graph database products include Neo4j, Amazon Neptune, JanusGraph, etc. These databases provide rich functions and APIs, allowing developers to easily process and manipulate graph data.

[0155] In some possible implementations, the device further includes:

[0156] The first sentence processing unit is used to perform sentence processing on the financial news public opinion to obtain a first target sentence.

[0157] The first input unit is used to input the first target sentence into a causal relationship processing model to obtain a plurality of causal triples and a plurality of probability scores corresponding to the plurality of causal triples.

[0158] Among them, the causal relationship processing model is a tool or method for analyzing and understanding causal relationships. It can help users determine the causal relationship between events, behaviors or variables, as well as the mechanism of action between them. The causal relationship processing model can identify causal relationships in text and generate corresponding causal triples (such as causal events, associated words, and result events). At the same time, these models can usually provide probability scores or confidence levels for the generated causal triples.

[0159] When training a causal relationship processing model, you can use labeled causal relationship data for supervised learning. The model learns the mapping relationship from input text to output causal triples to predict causal relationships as accurately as possible and assign them probability scores.

[0160] It should be noted that different causal relationship processing models may adopt different architectures and algorithms, so the way of providing probability scores or confidences may be different. Depending on the selected model and task requirements, an appropriate model can be selected and the probability scores of causal triples can be evaluated based on its output.

[0161] The following are several common causal relationship processing models: ① Traditional statistical models: Traditional statistical models (such as linear regression, logistic regression, etc.) can be used to analyze causal relationships. By controlling other variables, we can derive the degree of influence of one variable on another variable. ② Experimental design model: The experimental design model determines the causal relationship by comparing the experimental group and the control group, controlling and excluding other interfering factors. For example, a randomized controlled experiment can be used to evaluate the therapeutic effect of a drug on disease symptoms. ③ Causal graph model: The causal graph model is a method used to describe and visualize causal relationships. It uses arrows to represent the causal relationship between variables and can show direct and indirect causal paths. The causal graph model helps to better understand complex causal networks. ④ Structural equation model: The structural equation model (SEM) combines measurement models and structural models to evaluate the causal relationship between multiple variables. It can quantify the direct and indirect effects between variables and provide a comprehensive model to explain the data. These models are used in different contexts and research fields. The selection of an appropriate causal relationship processing model depends on the nature of the specific problem, the type of available data, and the purpose of the research.

[0162] The first setting unit is used to take the causal triples whose probability scores are greater than the score threshold as candidate triples.

[0163] In some possible implementations, the score threshold may be, but is not limited to, 0.75. The score threshold in this application may be adjusted according to specific circumstances, and this application does not impose any specific restrictions on the score threshold.

[0164] The second setting unit is used to take the candidate triples whose event descriptions meet the length rule as the abstract association relationship.

[0165] The length of the event description can be determined according to specific needs and application scenarios. Generally, the length of the event description can range from a few words to hundreds of words.

[0166] For a brief event description, it may only contain the main key information and elements, such as the subject, time, location and basic background of the event. Such a description is usually used to quickly understand the overview of the event and provide a concise and clear summary.

[0167] For more detailed event descriptions, more details and contextual information may be included to fully understand the occurrence and impact of the event. These descriptions may involve the cause, process, participants, motivation, consequences, etc. Such descriptions are usually used for in-depth analysis and research of events, providing a more comprehensive perspective and insight.

[0168] In some possible implementations, compliance with the length rule may include, but is not limited to, event descriptions having a length greater than 3 characters and less than 20 characters complying with the length rule.

[0169] In some possible implementations, the candidate triples whose event descriptions do not comply with the length rule (ie, event descriptions with a length less than 3 words and event descriptions with a length greater than 20 words do not comply with the length rule) need to be deleted.

[0170] In some possible implementations, the device further includes:

[0171] A calculation unit is used to calculate the correlation between each causal event in the abstract association relationship and the existing event.

[0172] In some possible implementations, a similarity or correlation measurement method (such as cosine similarity, Pearson correlation coefficient, etc.) can be used to calculate the similarity or correlation score between each causal event and the existing event. The specific method can be selected according to the data type and requirements.

[0173] The third setting unit is used to set the causal event whose correlation is greater than the correlation threshold as the known event; and set the causal event whose correlation is less than or equal to the correlation threshold as the new event.

[0174] In some possible implementations, the relevant threshold can be determined according to the actual situation. If a greater recall is expected, the threshold can be appropriately lowered, for example, to 0.4. If a better precision is desired, the threshold can be increased, for example, to 0.8.

[0175] In some possible implementations, the device further includes:

[0176] A generating and adding unit is used to add argument definitions to the known events that cannot be merged with the existing events by generating a new event module.

[0177] In some possible implementations, the following steps can be used to generate a new event module to add argument definitions: (1) Define a new event: Based on the existing event elements and related information, choose a clear and descriptive name to define the new event. (2) Determine arguments: Analyze the relationship between existing events and elements, and determine the arguments required for the new event. Arguments are key elements or attributes of an event, such as participants, time, and location. (3) Add argument definitions: Add definitions for each argument in the new event. The definition should clearly and accurately describe the meaning and function of the argument. (4) Data annotation: Use manual annotation methods to add the arguments of the new event as labels to the training data. You can match the existing event text data with the definition of the new event, and then annotate the arguments of the new event. (5) Model training: Retrain the event element extraction model using the expanded annotated data. During the model training process, the model will learn the argument definition and recognition method of the new event. (6) Iterative testing and improvement: Use the newly trained model for testing and evaluation, and adjust and improve it based on actual results. This process is repeated until the model can accurately extract event elements, including the newly added arguments.

[0178] In this way, the ability of the event element extraction model can be gradually improved so that it can recognize and extract more types of event elements.

[0179] In some possible implementations, the device further includes:

[0180] The second sentence processing unit is used to perform sentence processing on the new event to obtain a second target sentence.

[0181] The second input unit is used to input the second target sentence into the argument extraction model to obtain multiple candidate argument groups.

[0182] Among them, argument extraction is a task in natural language processing, which aims to identify arguments related to a specific predicate from a given text. The following is a common argument extraction model framework: ① Feature extraction: First, useful features need to be extracted from the text to represent each word or phrase. These features can include part of speech, syntactic relations, dependencies, contextual information, etc. A common method is to use pre-trained word vectors (such as Word2Vec, GloVe) to represent words. ② Sequence labeling model: The argument extraction problem is modeled as a sequence labeling problem, which can be modeled using models such as conditional random fields (CRF) or recurrent neural networks (such as long short-term memory networks LSTM). The input of the model is a text sequence after feature extraction, and the output is a label for each word, indicating whether the word is an argument and the argument type it belongs to. ③ Annotation data preparation: In order to train the argument extraction model, corpus data with associated predicates needs to be annotated. Each sentence in the corpus needs to be annotated with arguments related to the target predicate and specify its argument type. This usually requires manual annotation or the use of existing annotated datasets. ④Model training and optimization: Use labeled data to train the argument extraction model. Adjust the model's parameters and weights to improve the model's performance by maximizing predefined performance indicators (such as accuracy, recall, F1 value, etc.). The model can be trained using the backpropagation algorithm or other optimization algorithms. ⑤Inference and evaluation: After training, the model can be used to infer new texts and identify arguments related to the predicate. The performance of the model can be evaluated by comparing it with manually labeled or test data sets and calculating indicators such as the model's accuracy, recall, and F1 value.

[0183] It should be noted that implementing a complete argument extraction model requires comprehensive consideration of multiple aspects such as feature selection, model architecture, labeled data preparation, and model training, and proper debugging and optimization. At the same time, the performance of the model is also limited by the quality and scale of the available labeled data.

[0184] The third acquisition unit is used to acquire multiple argument confidences corresponding to the multiple candidate argument groups, and take the candidate argument group whose argument confidence is greater than the confidence threshold as the target argument group.

[0185] In some possible implementations, the confidence threshold may be, but is not limited to, 0.5 or 0.6, which is determined according to the specific scenario. The confidence threshold in this application may be adjusted according to the specific situation, and this application does not impose any specific restrictions on the confidence threshold.

[0186] In some possible implementations, the following methods can be used to estimate the confidence of arguments: ① Predicted probability scores: Some machine learning models (such as logistic regression and support vector machines) provide the ability to predict probability scores. These probabilities represent the model's confidence in each argument category. You can use these probabilities as an approximation of the confidence. ② Post-processing threshold setting: By applying a threshold to the score output by the model, you can filter out arguments with high confidence. A higher score may mean that the model is more confident in predicting the argument. You can set the threshold according to specific needs and select arguments with higher confidence. ③ Model-based uncertainty estimation: Some models can estimate the uncertainty of their predictions through ensemble methods (such as Bootstrap, Dropout) or Bayesian inference. By considering the uncertainty information of the model, the confidence of the argument can be better evaluated.

[0187] The combining unit is used to combine each target argument tuple using a greedy algorithm to obtain the element information.

[0188] An effective approach is to use a greedy algorithm to combine the complete argument information. A greedy algorithm is a heuristic algorithm that gradually builds a solution by selecting the best decision at each step.

[0189] In some possible implementations, the device further includes:

[0190] The identification unit is used to identify the event name and resource of the search data.

[0191] In some possible implementations, natural language processing (NLP) techniques and models can be used to identify event names and resources of search data. That is, deep learning-based NLP models, such as BERT or GPT, can be used to perform semantic understanding and information extraction on queries.

[0192] The grouping unit is used to group the event name and the resource according to time and user unique identifier to obtain multiple candidate tuples.

[0193] Among them, the user ID is a unique identifier assigned to each user by the system or platform, which is used to identify the user and manage user information.

[0194] The statistical unit is used to count the occurrence frequency of each candidate bigram.

[0195] The third setting unit is configured to take the candidate bigrams whose occurrence frequency is higher than the frequency threshold as target bigrams.

[0196] In some possible implementations, the frequency threshold may be, but is not limited to, a time slice determined by day, and the frequency is generally greater than 50. The frequency threshold in the present application may be adjusted according to specific circumstances, and the present application does not impose specific restrictions on the frequency threshold.

[0197] A filtering unit is used to filter the target binary group by cumulative information entropy to obtain the event-related target.

[0198] In addition, an embodiment of the present application also provides a device for constructing a financial event graph, the device including a memory and a processor, the memory being used to store programs or codes, and the processor being used to run the programs or codes stored in the memory to implement the above-mentioned method for constructing a financial event graph.

[0199] In addition, an embodiment of the present application further provides a computer-readable storage medium, characterized in that a code is stored in the computer-readable storage medium, and when the code is executed, the device executing the code implements the above-mentioned method for constructing a financial event map.

[0200] The embodiment of the present application provides a device for constructing a financial event graph. First, the first acquisition unit 201 can be used to acquire the abstract association relationship of financial news public opinion. Then, the calculation and extraction unit 202 uses correlation calculation to merge the known events in the abstract association relationship with the existing events to obtain a merged link, and extracts the causal relationship in the merged link, wherein the causal relationship is an abstract relationship between the known event and the existing event; the existing event is an event existing in the retrieval database. Then the element extraction unit 203 extracts the new event in the abstract association relationship from the financial news public opinion to obtain element information, and integrates the element information to obtain a specific event node. The second acquisition unit 204 acquires the event association target of the search data. Finally, the generation unit 205 inputs the causal relationship, the specific event node and the event association target into the graph database to generate a knowledge graph. The financial event graph of the present application can describe financial events in a more fine-grained manner. At the same time, the financial event graph of the present application can organically combine the association relationship of events in the form of a graph network. The financial event map of the present application records the general rules of event changes and the specific content of financial events, so that the financial event map can effectively support the attribution and expected inference of financial events.

[0201] The above is a detailed introduction to the construction method, device, equipment and storage medium of a financial event map provided by the present application. The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

[0202] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0203] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0204] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0205] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a financial event graph, characterized in that: The method further comprises: Obtain abstract correlations of financial news and public opinion; Using correlation calculation, the known events in the abstract association relationship are merged with the existing events to obtain a merged link, and the causal relationship in the merged link is extracted, wherein the causal relationship is an abstract relationship between the known event and the existing event; the existing event is an event existing in the search database; Extracting elements from the financial news public opinion according to the new event in the abstract association relationship to obtain element information, and integrating the element information to obtain a specific event node; Get the event-related target of the search data; The causal relationship, the specific event nodes and the event-related targets are input into a graph database to generate a knowledge graph.

2. The method according to claim 1, characterized in that: The abstract association relationship of obtaining financial news public opinion includes: Segment processing is performed on the financial news public opinion to obtain a first target sentence; Inputting the first target sentence into a causal relationship processing model to obtain a plurality of causal triples and a plurality of probability scores corresponding to the plurality of causal triples; The causal triples with probability scores greater than the score threshold are regarded as candidate triples; The candidate triples whose event descriptions meet the length rule are used as the abstract association relationships.

3. The method according to claim 1, characterized in that The method further comprises: Calculating the correlation between each causal event in the abstract association relationship and the existing event; The causal event whose correlation is greater than a correlation threshold is regarded as the known event; and the causal event whose correlation is less than or equal to the correlation threshold is regarded as the new event.

4. The method according to claim 1, characterized in that: The method further comprises: The known events that cannot be merged with the existing events are added with argument definitions by generating new event modules.

5. The method according to claim 1, characterized in that The extracting elements from the financial news public opinion according to the new event in the abstract association relationship to obtain element information includes: Sentence processing is performed on the new event to obtain a second target sentence; Inputting the second target sentence into an argument extraction model to obtain a plurality of candidate argument groups; Acquire multiple argument confidences corresponding to the multiple candidate argument groups, and use the candidate argument group whose argument confidence is greater than a confidence threshold as the target argument group; The element information is obtained by combining each target argument tuple using a greedy algorithm.

6. The method according to claim 1, characterized in that The event-related target of obtaining the search data includes: Identify the event name and resource of the search data; Grouping the event name and the resource according to time and user unique identifier to obtain multiple candidate tuples; Count the frequency of occurrence of each candidate bigram; Taking the candidate bigrams whose occurrence frequency is higher than the frequency threshold as target bigrams; The target binary group is filtered by cumulative information entropy to obtain the event-related target.

7. A device for constructing a financial event graph, characterized in that: The device comprises: The first acquisition unit is used to acquire the abstract correlation relationship of financial news and public opinion; A calculation extraction unit is used to merge the known events in the abstract association relationship with the existing events to obtain a merged link by using correlation calculation, and to extract the causal relationship in the merged link, wherein the causal relationship is an abstract relationship between the known event and the existing event; the existing event is an event existing in the search database; An element extraction unit, configured to extract elements from the financial news public opinion according to the new event in the abstract association relationship to obtain element information, and integrate the element information to obtain a specific event node; A second acquisition unit, used to acquire event-related targets of the search data; A generating unit is used to input the causal relationship, the specific event node and the event-related target into a graph database to generate a knowledge graph.

8. The device according to claim 7, characterized in that The device also includes: A first sentence processing unit, configured to perform sentence processing on the financial news public opinion to obtain a first target sentence; A first input unit, configured to input the first target sentence into a causal relationship processing model to obtain a plurality of causal triples and a plurality of probability scores corresponding to the plurality of causal triples; A first setting unit is used to take the causal triples whose probability scores are greater than the score threshold as candidate triples; The second setting unit is used to take the candidate triples whose event descriptions meet the length rule as the abstract association relationship.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for constructing a financial event graph as described in any one of claims 1 to 6 is implemented.

10. A vehicle, characterized in that: The vehicle includes a control module, and the control module is used to execute the method for constructing a financial event graph as described in any one of claims 1-6.

Images