Data prediction method and device, large model training method and device and electronic equipment

Abstract

The invention provides a data prediction method and device, a large model training method and device and electronic equipment, and relates to the technical field of artificial intelligence, in particular to the technical fields of deep learning, data mining, data analysis, big data processing and the like. According to the specific implementation scheme, a plurality of information abstracts related to a target object are obtained; wherein the plurality of information abstracts have different information categories so as to be used for representing performance conditions and / or influenced degrees of the target object in different dimensions; and obtaining a feature data prediction result of the target object and interpretation information for the feature data prediction result by using the target large model based on the plurality of information abstracts.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, in particular to technical fields such as deep learning, data mining, data analysis, and big data processing, and specifically to a data prediction method, a large model training method, a device, and an electronic device. Background Art

[0002] With the continued steady growth of the economy, the financial market has become the core of the modern economic system and has attracted the attention of many investors. However, given the ever-changing market environment, the financial market may experience drastic fluctuations, which makes market participants expect to obtain an efficient method that can accurately predict some characteristic data related to the financial market in order to optimize investment decisions. Summary of the invention

[0003] The present disclosure provides a data prediction method, a large model training method, a device and an electronic device.

[0004] According to a first aspect of the present disclosure, a data prediction method is provided, comprising:

[0005] Acquire multiple information summaries related to the target object; wherein the multiple information summaries have different information categories, so as to characterize the performance and / or impact degree of the target object in different dimensions;

[0006] By using the target large model and based on multiple information summaries, the feature data prediction results of the target object and the explanatory information for the feature data prediction results are obtained.

[0007] According to a second aspect of the present disclosure, a large model training method is provided, comprising:

[0008] Acquire multiple information summary samples related to the target object sample; wherein the multiple information summary samples have different information categories, so as to characterize the performance and / or impact degree of the target object sample in different dimensions;

[0009] Using the initial large model, based on multiple information summary samples, a feature data prediction result sample for the target object sample and an explanation information sample for the feature data prediction result sample are obtained;

[0010] Based on the feature data prediction result samples and the explanation information samples, the initial large model is trained to obtain the target large model.

[0011] According to a third aspect of the present disclosure, there is provided a data prediction device, comprising:

[0012] A summary information acquisition unit, used to acquire a plurality of information summaries related to the target object; wherein the plurality of information summaries have different information categories, so as to characterize the performance and / or degree of influence of the target object in different dimensions;

[0013] The feature data prediction unit is used to obtain the feature data prediction result of the target object and the explanation information for the feature data prediction result based on multiple information summaries by using the target large model.

[0014] According to a fourth aspect of the present disclosure, a large model training device is provided, comprising:

[0015] A first sample acquisition unit is used to acquire a plurality of information summary samples related to the target object sample; wherein the plurality of information summary samples have different information categories, so as to characterize the performance and / or the degree of influence of the target object sample in different dimensions;

[0016] A second sample acquisition unit is used to obtain a feature data prediction result sample for the target object sample and an explanation information sample for the feature data prediction result sample based on the multiple information summary samples using the initial large model;

[0017] The large model training unit is used to train the initial large model based on the feature data prediction result samples and the explanation information samples to obtain the target large model.

[0018] According to a fifth aspect of the present disclosure, there is provided an electronic device, including:

[0019] at least one processor;

[0020] a memory communicatively coupled to the at least one processor;

[0021] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method provided by the embodiment of the present disclosure.

[0022] According to a sixth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the method provided according to an embodiment of the present disclosure.

[0023] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, which implements the method provided according to the embodiment of the present disclosure when executed by a processor.

[0024] The use of the present disclosure can improve the accuracy and availability of feature data prediction results.

[0025] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0027] Figure 1 A flowchart of a data prediction method provided by an embodiment of the present disclosure;

[0028] Figure 2 A flowchart for obtaining a news summary provided by an embodiment of the present disclosure;

[0029] Figure 3 A flowchart for obtaining a dynamic summary of feature data provided by an embodiment of the present disclosure;

[0030] Figure 4 A flowchart for obtaining a macro information summary provided by an embodiment of the present disclosure;

[0031] Figure 5 A flowchart illustrating a basic surface summary acquisition process provided by an embodiment of the present disclosure;

[0032] Figure 6 A complete flow diagram of a data prediction method provided by an embodiment of the present disclosure;

[0033] Figure 7 An auxiliary illustration of the integrity process of a data prediction method provided by an embodiment of the present disclosure;

[0034] Figure 8 A schematic diagram of a scenario of a data prediction method provided by an embodiment of the present disclosure;

[0035] Fig. 9 A flowchart of a large model training method provided in an embodiment of the present disclosure;

[0036] Fig.10 A complete flow chart of a large model training method provided in an embodiment of the present disclosure;

[0037] Fig.11 An auxiliary illustration of the integrity process of a large model training method provided in an embodiment of the present disclosure;

[0038] Fig.12 A schematic diagram of a scenario of a large model training method provided in an embodiment of the present disclosure;

[0039] Fig.13A schematic structural block diagram of a data prediction device provided in an embodiment of the present disclosure;

[0040] Fig.14 A schematic structural block diagram of a large model training device provided in an embodiment of the present disclosure;

[0041] Fig.15 A schematic structural block diagram of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0042] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope of the present disclosure. Similarly, for the sake of clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0043] As described in the background technology, with the continuous and steady growth of the economy, the financial market has become the core of the modern economic system and has attracted the attention of many investors. However, given the ever-changing market environment, the financial market may experience drastic fluctuations, which makes market participants expect to obtain an efficient method that can accurately predict some characteristic data related to the financial market in order to optimize investment decisions.

[0044] Taking the stock market as an example, at present, some market participants understand and analyze market information to obtain the target stock's stock price trend forecast results and make investment decisions accordingly. However, for retail investors, due to their limited ability to understand and analyze market information, prone to behavioral biases, and lack of risk management skills, they are not only likely to miss promising investment opportunities, but also increase the incidence of improper risks. Similarly, small and medium-sized asset or wealth management companies will find it difficult to conduct in-depth analysis of market information due to limited resources and a small range of choices. In addition, even large professional asset or wealth management companies will have problems such as poor communication and incentive diversity due to their large organizational size, which will affect their ability to understand and analyze market information, and ultimately affect the accuracy of the target stock's stock price trend forecast results.

[0045] In seeking solutions to the above problems, researchers have explored the use of deep learning technology to predict stock price trends, which mainly include:

[0046] (1) Stock price trend prediction based on text analysis

[0047] In the early research on stock price trend prediction based on text analysis, researchers mainly focused on how to use support vector machines (SVM) to evaluate the effectiveness of different text representations (e.g., bags of words, noun phrases, and named entities). Since then, these "shallow" features have been gradually replaced by structured data, forming events in the form of (Actor, Action, Object) tuples, which are used as inputs to neural network models to predict stock price trends. Later, with the development and progress of artificial intelligence technology, and considering the chaotic and diverse characteristics of text information, researchers began to use neural network models based on attention mechanisms to directly capture key information from pre-trained text embeddings to enhance the performance of neural network models.

[0048] (2) Using existing large language models to predict stock price trends

[0049] Zaremba and Demir's research revealed the great potential of ChatGPT to be applied in the financial field, especially in processing related tasks based on natural language processing technology, such as sentiment analysis of public opinion news. These tasks show a significant correlation between natural language processing objects and stock market dynamics. In addition, Lopez-Lira and Tang verified the accuracy of ChatGPT in sentiment analysis of public opinion news, and emphasized the positive correlation between the stock price trend prediction results (score values) generated by ChatGPT and the subsequent stock returns.

[0050] (3) Using large language models in vertical fields to predict stock price trends

[0051] Among the related studies on using large language models to process financial tasks, BloombergGPT is more prominent. It uses the existing financial corpus to train a 50 billion-parameter financial large language model (that is, a large language model in a vertical field) to achieve named entity recognition and sentiment analysis of public opinion news, thereby realizing stock price trend prediction.

[0052] However, the inventors have found that the above solution of using deep learning technology to predict stock price trends still has the following shortcomings:

[0053] (1) Single information source

[0054] When using existing large language models to predict stock price trends, usually only a single type of text data (for example, public opinion news) is considered as the basis for prediction, while other types of text data (for example, stock price dynamic data, macroeconomic research reports, fundamental information, etc.) are not considered, let alone other modal data (for example, image data, knowledge graphs, etc.). This makes it impossible for large language models to obtain relatively accurate stock price trend prediction results.

[0055] (2) There are difficulties in processing social texts and related public opinions

[0056] The existence of social texts and related public opinions increases the difficulty of predicting stock price trends using existing large language models. Specifically, breaking news, crisis events, government reports, etc. may have a significant impact on stock prices, while unconfirmed opinions, gossip, and vague comments usually do not have a significant impact on stock prices. However, it is difficult for existing large language models to accurately weigh these mixed social texts and related public opinions and derive maximum likelihood estimates to ensure the accuracy of stock price trend prediction results.

[0057] (3) Cumulative error affects the accuracy of stock price trend prediction results

[0058] When using existing large language models to predict stock price trends, multiple prediction stages are usually involved. Among them, if low-quality results are generated in the previous stage, the low-quality results generated in the previous stage will affect the prediction results in the next stage, and this influence will gradually accumulate errors, and ultimately affect the accuracy of the stock price trend prediction results.

[0059] (4) Insufficient explanatory ability

[0060] When using existing large-scale language models to predict stock price trends, usually only the stock price trend prediction results are obtained without corresponding explanation information. This will reduce the user's trust in the stock price trend prediction results, thereby reducing the usability of the stock price trend prediction results.

[0061] In summary, at present, the solution of using deep learning technology to predict stock price trends can neither obtain more accurate stock price trend prediction results nor ensure the availability of stock price trend prediction results.

[0062] In view of the above problems, the embodiments of the present disclosure provide a data prediction method, which can be applied to electronic devices. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a mainframe computer, a conventional computer (for example, a desktop computer, a laptop computer, a car computer, etc.), a personal digital processing unit or other similar computing devices. Figure 1The flowchart diagram shown illustrates a data prediction method provided by an embodiment of the present disclosure. It should be noted that although a logical sequence is shown in the flowchart diagram, in some cases, the steps shown or described may also be performed in other sequences.

[0063] Step S101, obtaining multiple information summaries related to the target object.

[0064] The target object may be any financial investment object (eg, stocks, funds, futures, etc.) for which feature data prediction is required. Here, the feature data may be the price trend of the target object.

[0065] In addition, in the embodiment of the present disclosure, the multiple information summaries have different information categories to characterize the performance and / or degree of influence of the target object in different dimensions. For example, the multiple information summaries may include at least one of a news summary, a feature data dynamic summary, a macro information summary, and a fundamental summary.

[0066] Among them, the news summary can be the summary information obtained by processing the public opinion news related to the target object, so as to characterize the degree of influence of the target object in the dimension of public opinion news; the feature data dynamic summary can be the summary information obtained by processing the price dynamic data of the target object, so as to characterize the market performance of the target object; the macro information summary can be the summary information obtained by processing the macro market economic data, so as to characterize the degree of influence of the target object in the dimension of macro market economy; the fundamental summary can be the summary information obtained by processing the fundamental data related to the target object, so as to characterize the degree of influence of the target object in the dimension of fundamental data.

[0067] Step S102, using the target macro model, based on multiple information summaries, obtains the feature data prediction results of the target object and the explanation information for the feature data prediction results.

[0068] The feature data prediction result may be a price trend prediction result of the target object.

[0069] In one example, answer guidance information can be obtained, and the answer guidance information and multiple information summaries can be input into the target large model, so that the target large model processes the multiple information summaries according to the answer guidance information to obtain the feature data prediction result of the target object and the explanation information for the feature data prediction result. The answer guidance information is the first prompt information, which is used to instruct the target large model how to obtain the feature data prediction result of the target object based on the multiple information summaries and the explanation information for the feature data prediction result.

[0070] In addition, in the disclosed embodiment, the target large model is the target signal generator, and the target signal generator can be a large language model. Here, the large language model can be a neural network model that has been pre-trained and parameter-tuned, and has general language knowledge, world knowledge, professional knowledge in various fields, etc., and is stored in the model in the form of parameters. Specifically, the large language model can be an autoregressive generative model of the Transformer architecture.

[0071] By adopting the data prediction method provided by the embodiment of the present disclosure, multiple information summaries related to the target object and having different information categories can be obtained for the performance and / or degree of influence of the target object in different dimensions, and then the target large model is used to obtain the characteristic data prediction results of the target object based on multiple information summaries, as well as explanatory information for the characteristic data prediction results. In this way, on the one hand, the target large model can comprehensively understand the performance and degree of influence of the target object from multiple dimensions based on multiple information summaries related to the target object and having different information categories, and provide a relatively comprehensive prediction basis to reduce the bias of a single information source, thereby improving the accuracy of the characteristic data prediction results; on the other hand, the target large model will not only obtain the characteristic data prediction results of the target object, but also obtain explanatory information for the characteristic data prediction results, so as to enhance the user's trust in the characteristic data prediction results by improving the interpretability of the characteristic data prediction results, thereby improving the availability of the characteristic data prediction results.

[0072] Since the public opinion news (including announcements, reports, analyst opinions and research results, etc.) of the target entity (that is, the economic organization that issues the target object) may have an immeasurable short-term or long-term impact on market sentiment and stock prices. Therefore, in the embodiment of the present disclosure, a news summarizer can be used to process (for example, collect, condense, etc.) the public opinion news related to the target object to obtain a daily news summary and / or a current time period news summary (for example, a weekly news summary, a monthly news summary, or a summary of a time period of other lengths) related to the target object. Based on this, it can be understood that in the embodiment of the present disclosure, multiple information summaries can include news summaries, and the news summary can include a daily news summary and / or a current time period news summary.

[0073] In some optional implementations, when the news summary includes a daily news summary, step S101, that is, "obtaining multiple information summaries related to the target object" may include:

[0074] Obtain N groups of daily public opinion news related to the target object; wherein the N groups of daily public opinion news correspond to N independent dates in the current period one by one; N ≥ 2 and is an integer;

[0075] Each group of daily public opinion news in the N groups of daily public opinion news is used as the public opinion news to be processed, and the public opinion news to be processed is processed by using the first news summarizer to obtain the daily news summary corresponding to the public opinion news to be processed.

[0076] The daily public opinion news may be in text form or in picture form, and this disclosure embodiment does not limit this.

[0077] In one example, an application programming interface (API) providing information about the stock market and financial news can be used to obtain N groups of daily public opinion news related to a target object, and each group of the N groups of daily public opinion news can be preprocessed to exclude news texts in the N groups of daily public opinion news that are not related to the target object (for example, news texts with title bait properties), and to ensure that each group of the N groups of daily public opinion news has an appropriate text format to achieve the purpose of supporting the first news summarizer.

[0078] After obtaining N groups of daily public opinion news, each group of daily public opinion news in the N groups of daily public opinion news can be used as public opinion news to be processed, and then the public opinion news to be processed can be input into the first news summarizer, so as to use the first news summarizer to identify and extract the organization name, entity name (for example, person name, product name, etc.) in the public opinion news to be processed, and the key events involved in the public opinion news (for example, performance report, merger and acquisition, etc.), and understand the contextual information behind the public opinion news (for example, the cause of the event, the organization's response, etc.), and finally, according to the preset output format, output the daily news summary corresponding to the public opinion news to be processed and store it in the centralized repository.

[0079] The first news summarizer may be a large language model. Here, the large language model may be a pre-trained neural network model, which has general language knowledge, world knowledge, professional knowledge in various fields, etc., and is stored in the model in the form of parameters. Specifically, the large language model may be an autoregressive generative model of the Transformer architecture.

[0080] Furthermore, in the case where the news summary also includes the news summary of the current period, step S101, that is, "obtaining multiple information summaries related to the target object" may also include:

[0081] Based on N daily news summaries corresponding to N groups of daily public opinion news, a daily news summary set is obtained;

[0082] Get the historical period news summary corresponding to the historical period;

[0083] The second news summarizer is used to process the daily news summary set based on the historical time period news summary to obtain the current time period news summary corresponding to the current time period.

[0084] Among them, when the length of the historical period is one week, the length of the current period can also be one week, and the current period and the historical period can be two consecutive weeks; when the length of the historical period is one month, the length of the current period can also be one month, and the current period and the historical period can be two consecutive months; the second news summarizer can be a large language model, and the second news summarizer and the first news summarizer can be the same large language model.

[0085] In one example, after obtaining N daily news summaries, the N daily news summaries may be spliced ​​to obtain a daily news summary set, and a historical period news summary corresponding to the historical period may be obtained, and then the second news summarizer may be used to process the daily news summary set based on the historical period news summary to obtain a current period news summary corresponding to the current period. The current period news summary may include the target entity's monthly market positioning, sales performance, legal and strategic initiatives, innovation and sustainable development, and supervisory challenges.

[0086] To further clarify the process of obtaining daily news summaries and current period news summaries, the following will use mathematical formulas and combine Figure 2 , describes the process of obtaining daily news summaries and current period news summaries:

[0087]

[0088] PN S,t =Summarize2(PN S,t-1 , N S,τ )

[0089] Among them, N S,i Used to represent the daily public opinion news corresponding to the i-th independent date in the current period, and 1≤i≤N; It is used to represent the processing result obtained by the first news summarizer after processing the public opinion news to be processed, that is, the daily news summary corresponding to the i-th independent date in the current period; N S,τ Used to represent a collection of daily news summaries; PN S,t-1 Used to represent news summaries of historical periods; It is used to represent the processing result obtained by the second news summarizer processing the daily news summary set, that is, PN S,t , which is specifically used to represent the news summary of the current period.

[0090] Through the above method, in the disclosed embodiment, after obtaining N groups of daily public opinion news related to the target object and taking each group of daily public opinion news in the N groups of daily public opinion news as the public opinion news to be processed, the first news summarizer can be used to process the public opinion news to be processed, and obtain the daily news summary corresponding to the public opinion news to be processed, thereby reducing the time cost required for manual understanding and analysis of the N groups of daily public opinion news, and at the same time, reducing the error of manual processing to improve the reliability of the daily news summary. In the disclosed embodiment, it is also possible to obtain a daily news summary set based on the N daily news summaries corresponding to the N groups of daily public opinion news, and obtain the historical period news summary corresponding to the historical period, and use the second news summarizer to process the daily news summary set based on the historical period news summary to obtain the current period news summary corresponding to the current period. In this way, it is also possible to reduce the time cost required for manual understanding and analysis of the daily news summary set, and at the same time, reduce the error of manual processing to improve the reliability of the current period news summary.

[0091] In some optional implementations, the multiple information summaries include a dynamic summary of feature data. Based on this, step S101, that is, "obtaining multiple information summaries related to the target object" may include:

[0092] Get the first market data of the target object;

[0093] Acquire second market data of a first target number of reference objects similar to the target object;

[0094] The feature data dynamic summarizer is used to obtain the feature data dynamic summary based on the first market data of the target object, the second market data of the first target number of reference objects, and the specific reference index.

[0095] Among them, taking the target object as a stock that needs to be predicted for feature data as an example, the first market data may include the target object's rate of return, volatility, Sharpe ratio, maximum drawdown, similarity-difference moving average, relative strength index, balance volume index, capital flow index, Bollinger band index, etc. within a preset period. Here, the preset period may be a historical period of one month, one year or other lengths.

[0096] In addition, in the disclosed embodiment, the first target number can be set according to actual application requirements, for example, it can be set to 3; the first target number of reference objects can be the first target number of financial investment objects issued by similar economic organizations that are highly similar to the target entity, selected by domain experts in the industry to which the target entity belongs based on industry expertise (Know-How) and professional experience; taking the reference object as stocks as an example, the second market data may include the rate of return, volatility, Sharpe ratio, maximum drawdown, moving average difference, relative strength index, balance volume indicator, capital flow indicator, Bollinger Band indicator, etc. of the stock reference object within a preset time period.

[0097] After obtaining the first market data of the target object and the second market data of a first target number of reference objects, a feature data dynamic summarizer can be used to obtain a dynamic summary of the feature data according to the first extraction instruction, based on the first market data of the target object, the second market data of the first target number of reference objects, and a specific reference index (for example, when the target object and the reference object are both stocks, the specific reference index may be the CSI 1000 Index).

[0098] Among them, the feature data dynamic summarizer can be a large language model; the first extraction instruction is the second Prompt information, which is used to instruct the feature data dynamic summarizer how to obtain the feature data dynamic summary based on the first market data of the target object, the second market data of the first target number of reference objects, and the specific reference index. For example, the first extraction instruction may include a summary extraction requirement for each sub-category feature data indicator in the multiple sub-category feature data indicators included in the feature data dynamic summary; similarly, when the target object and the reference object are both stocks, the multiple sub-category feature data indicators may include yield, volatility, Sharpe ratio, maximum drawdown, current and future market sentiment, conclusions, profit drivers, future financial prospects, etc.

[0099] To further clarify the process of obtaining dynamic summaries of feature data, the following mathematical formulas will be used in combination with Figure 3 , describes the process of obtaining the dynamic summary of feature data:

[0100]

[0101] Among them, Metrics S,t First market data for characterizing the target audience; The second market data for representing the j1th reference object among the first target number (ie, n1) stock reference objects, and 1≤j1≤n1; the CSI is used to represent a specific reference index; The processing result obtained by the dynamic summarizer of the characteristic data processing the first market data of the target object, the second market data of the first target number of reference objects, and the specific reference index, that is, P S,t , which is specifically used to characterize the dynamic summary of feature data.

[0102] Through the above method, in the embodiment of the present disclosure, the first market data of the target object can be obtained, and the second market data of the first target number of reference objects similar to the target object can be obtained, and the dynamic summarizer of the characteristic data can be used to obtain the dynamic summary of the characteristic data based on the first market data of the target object, the second market data of the first target number of reference objects, and the specific reference index. In this way, not only can the time cost required for manually understanding and analyzing the first market data of the target object and understanding and analyzing the second market data of the first target number of reference objects be reduced, but also, because the first market data, the second market data of the first target number of reference objects, and the specific reference index can relatively comprehensively characterize the market performance of the target object, the reliability of the dynamic summary of the characteristic data can be improved.

[0103] Furthermore, in-depth macroeconomic analysis is essential for making wise investment decisions and effective capital allocation. Specifically, it can provide key insights into the overall economic health and performance to help investors obtain key information in a timely manner, better understand the current economic environment and make wise investment decisions. Therefore, in the disclosed embodiment, the multiple information summaries may also include macro information summaries. Based on this, in some optional implementations, step S101, that is, "obtaining multiple information summaries related to the target object" may include:

[0104] Using the first macro summarizer, processing the second target number of macro research reports to obtain the second target number of macro research summaries corresponding to the second target number of macro research reports;

[0105] Using a second macro summarizer, the macro indicator data is processed to obtain a macro indicator summary;

[0106] A third macro summarizer is used to obtain a macro information summary based on a second target number of macro research summaries and macro indicator summaries.

[0107] Regarding macro research summaries, in one example, after obtaining a second target number of macro research reports (for example, macroeconomic research reports), each macro research report in the second target number of macro research reports can be used as a research report to be processed, and the first macro summarizer can be used to process the research report to be processed according to the second extraction instructions to obtain a macro research summary corresponding to the research report to be processed.

[0108] Among them, the second target number can be set according to actual application requirements, and the embodiments of the present disclosure do not impose any restrictions on this; the first macro summarizer can be a large language model; the second extraction instruction is the third Prompt information, which is used to instruct the first macro summarizer how to process the research report to be processed, and obtain a macro research summary corresponding to the research report to be processed; the macro research summary may include analysis of economic growth, inflation, employment market, monetary policy, fiscal policy, international trade and exchange rates, real estate market, overseas macro environment, macro environment and outlook, and assets and industry categories that are expected to rise or fall, etc., to obtain summary information.

[0109] Regarding the macro-indicator summary, in one example, after obtaining the macro-indicator data (eg, macroeconomic indicator data), a second macro-summarizer may be used to process the macro-indicator data according to a third extraction instruction to obtain a macro-indicator summary.

[0110] Among them, the macro-indicator data may include the year-on-year growth rate of the total amount of money in circulation (M0) / the total amount of money in circulation plus the demand deposits of individuals and enterprises (M1) / the total amount of money in circulation plus the time deposits of individuals and enterprises (M2), expected liquidity indicators, the US dollar-RMB exchange rate, etc.; the second macro-summarizer may be a large language model, and the second macro-summarizer and the first macro-summarizer may be the same large language model; the third extraction instruction is the fourth Prompt information, which is used to instruct the second macro-summarizer how to process the macro-indicator data to obtain a macro-indicator summary. For example, the third extraction instruction may include macro-indicator data of a reference historical period (for example, last month) and a current period (for example, this month) to generate a macro-indicator summary; the macro-indicator summary may include summary information obtained by analyzing economic growth, inflation, the job market, monetary policy, fiscal policy, international trade and exchange rates, etc.

[0111] Regarding the macro information summary, in one example, after obtaining the second target number of macro research summaries and macro indicator summaries, a third macro summarizer can be used to obtain the macro information summary based on the second target number of macro research summaries and macro indicator summaries in accordance with the fourth extraction instructions.

[0112] Among them, the third macro summarizer can be a large language model, and the third macro summarizer, the first macro summarizer and the second macro summarizer can be the same large language model; the fourth extraction instruction is the fifth Prompt information, which is used to instruct the third macro summarizer how to obtain the macro information summary based on the second target number of macro research summaries and macro indicator summaries; the macro information summary can include summary information obtained by analyzing economic growth, inflation, job market, monetary policy, fiscal policy, international trade and exchange rates, assets expected to rise and fall, etc.

[0113] To further clarify the process of obtaining macro information summaries, the following will use mathematical formulas and combine Figure 4 , describes the process of obtaining the macro information summary:

[0114]

[0115] in, The j2th macroeconomic report is used to represent the second target number (i.e., n2) of macroeconomic research reports, and 1≤j2≤n2; Report is used to represent the processing result obtained by the first macro summarizer processing the j2th macroeconomic report in the second target number of macroeconomic research reports, that is, the macroeconomic research summary corresponding to the j2th macroeconomic report; t-1 Macroeconomic indicators used to characterize historical periods; Report t Macroeconomic indicators used to characterize the current period; used to represent the processing result obtained by the second macro summarizer processing the macro indicator data, that is, the macro indicator summary; It is used to represent the processing result obtained by the third macro summarizer processing the second target number of macro research summaries and macro indicator summaries, that is, M t , which is specifically used to characterize macro information summary.

[0116] Through the above method, in the embodiment of the present disclosure, the first macro summarizer can be used to process the second target number of macro research reports to obtain the second target number of macro research summaries corresponding to the second target number of macro research reports, and the second macro summarizer can be used to process the macro indicator data to obtain the macro indicator summary, and the third macro summarizer can be used to obtain the macro information summary based on the second target number of macro research summaries and macro indicator summaries. In this way, not only can the time cost required for manual understanding and analysis of the second target number of macro research reports and understanding and analysis of macro indicator data be reduced, but also, because the second target number of macro research reports and macro indicator data can relatively comprehensively characterize the degree of influence of the target object in the dimension of macro market economy, the reliability of the macro information summary can be improved.

[0117] In some optional implementations, the multiple information summaries include a basic surface summary. Based on this, step S101, that is, "obtaining multiple information summaries related to the target object" may include:

[0118] Using a first fundamental summarizer, the management level analysis results of the target entity are processed to obtain a management level fundamental summary; wherein the target entity is an economic organization that issues the target object;

[0119] Using the second fundamental summarizer, the analyst research reports related to the target entity are processed to obtain the analyst fundamental summary;

[0120] A third fundamental summarizer is used to obtain a fundamental summary based on the management fundamental summary, the analyst fundamental summary, and specific data of the target entity.

[0121] Regarding the management fundamentals summary, in one example, the latest and second-latest management analysis results of the target entity can be obtained, and after obtaining the management analysis results, the first fundamentals summarizer is used to process the management analysis results according to the fifth extraction instruction to obtain the management fundamentals summary.

[0122] Among them, the management analysis results can be the analysis results obtained by the management of the target entity when discussing and analyzing the current health status and future development trajectory of the target entity; the first fundamental summarizer can be a large language model; the fifth extraction instruction is the sixth Prompt information, which is used to instruct the first fundamental summarizer how to process the management analysis results to obtain the management fundamental summary; the management fundamental summary can include the emotional tone of the management, the potential risks and opportunities expressed by the management for the target entity, the profit drivers expressed by the management for the target entity, etc., and the summary information obtained.

[0123] Regarding the analyst fundamental summary, in one example, after obtaining the first analyst research report on the target entity, the first analyst research report on the target entity can be separately used as an analyst research report related to the target entity, and a second fundamental summarizer can be used, which can be a large language model; according to the sixth extraction instruction, the analyst research report related to the target entity is processed to obtain the analyst fundamental summary.

[0124] Among them, the first analyst research report can be the analysis result obtained by the financial analyst when discussing and analyzing the current health status and future development trajectory of the target entity; the second fundamental summarizer can be a large language model, and the second fundamental summarizer and the first fundamental summarizer can be the same large language model; the sixth extraction instruction is the seventh Prompt information, which is used to instruct the second fundamental summarizer how to process the analyst research report related to the target entity to obtain the analyst fundamental summary; the analyst fundamental summary can include the summary information obtained by analyzing the emotional tone of the financial analyst, the potential risks and opportunities expressed by the financial analyst for the target entity, and the profit drivers expressed by the financial analyst for the target entity.

[0125] For the analyst fundamental summary, in another example, a first analyst research report on the target entity can be obtained, as well as a second analyst research report on a third target number reference entity similar to the target entity (for example, a reference economic organization, which can be a listed company), which can be used together as the analyst research report related to the target entity. The second fundamental summarizer can be used to process the analyst research report related to the target entity in accordance with the seventh extraction instruction to obtain the analyst fundamental summary.

[0126] Among them, the first analyst research report can be the analysis result obtained by the financial analyst when discussing and analyzing the current health status and future development trajectory of the target entity; the third target number can be set according to actual application requirements, for example, it can be set to 5; the second analyst research report can include the analysis results obtained by the financial analyst when discussing and analyzing the health status and development trajectory of the target entity in the historical period (for example, last month) and the current period (for example, this month), that is, the second analyst research report can include the analyst research report of the historical period and the analyst research report of the current period; the second fundamental summarizer can be a large language model, and the second fundamental summarizer and the first fundamental summarizer can be the same large language model; the seventh extraction instruction is the eighth prompt information, which is used to instruct the second fundamental summarizer how to process the analyst research report related to the target entity to obtain the analyst fundamental summary; the analyst fundamental summary can include the summary information obtained by analyzing the emotional tone of the financial analyst, the potential risks and opportunities expressed by the financial analyst for the target entity, and the profit drivers expressed by the financial analyst for the target entity.

[0127] Regarding the fundamental summary, in one example, after obtaining the management fundamental summary, the analyst fundamental summary, and the specific data of the target entity (for example, the structured financial data of the target entity, which may be in Excel format), the third fundamental summarizer may be used to process the management fundamental summary, the analyst fundamental summary, and the specific data of the target entity in accordance with the eighth extraction instruction to obtain the fundamental summary.

[0128] Among them, the third fundamental summarizer can be a large language model, and the third fundamental summarizer, the first fundamental summarizer and the second fundamental summarizer can be the same large language model; the eighth extraction instruction is the ninth Prompt information, which is used to instruct the third fundamental summarizer how to process the management fundamental summary, the analyst fundamental summary, and the specific data of the target entity to obtain the fundamental summary; the fundamental summary may include the valuation level, profitability, profit quality, revenue growth rate, debt repayment ability, corporate governance ability, emotional tone of management, potential risks and opportunities for the target entity expressed by management, profit drivers for the target entity expressed by management, future financial prospects for the target entity expressed by management, emotional tone of financial analysts, potential risks and opportunities for the target entity expressed by financial analysts, profit drivers for the target entity expressed by financial analysts, future financial prospects for the target entity expressed by financial analysts, etc. of both the target entity and the reference entity to obtain summary information.

[0129] In addition, in the embodiment of the present disclosure, the fundamental summary may be a plain text format document, for example, it may be a text document written in the plain text format Markdown.

[0130] To further clarify the process of obtaining the fundamental summary, the following will use mathematical formulas and combine Figure 5 , the process of obtaining the fundamental summary is described as follows:

[0131]

[0132] in, Management analysis results used to characterize the target entity; used to represent a processing result obtained by the first fundamental summarizer processing the management level analysis result of the target entity, that is, a management level fundamental summary; Used to characterize analyst research related to the target entity; used to represent the processing result obtained by the second fundamental summarizer processing the analyst research report related to the target entity, that is, the analyst fundamental summary; Specific data used to characterize the target entity; It is used to represent the processing result obtained by the third fundamental summarizer on the management fundamental summary, the analyst fundamental summary, and the specific data of the target entity, that is, F S , which is specifically used to characterize the fundamental summary.

[0133] Through the above method, in the embodiment of the present disclosure, the first fundamental summarizer can be used to process the management analysis results of the target entity to obtain the management fundamental summary, and the second fundamental summarizer can be used to process the analyst research report related to the target entity to obtain the analyst fundamental summary, and then the third fundamental summarizer can be used to obtain the fundamental summary based on the management fundamental summary, the analyst fundamental summary, and the specific data of the target entity. In this way, not only can the time cost required for manually understanding and analyzing the management analysis results of the target entity and understanding and analyzing the analyst research report related to the target entity be reduced, but also, because the management analysis results of the target entity and the analyst research report related to the target entity can relatively comprehensively characterize the degree of influence of the target object in the dimension of fundamental data, the reliability of the fundamental summary can be improved.

[0134] In some optional implementations, step S102, i.e., “obtaining a feature data prediction result of the target object based on multiple information summaries, and explanation information for the feature data prediction result” may include:

[0135] Get answer guidance information;

[0136] Based on the answer guidance information, the chain logic reasoning steps are obtained;

[0137] Based on multiple information summaries and following the chain logic reasoning steps, the feature data prediction results of the target object and the explanatory information for the feature data prediction results are obtained.

[0138] Among them, the answer guidance information is the tenth Prompt information, which is used to instruct the target large model how to obtain the feature data prediction results of the target object based on multiple information summaries, as well as the explanatory information for the feature data prediction results.

[0139] For example, the answer guidance information is used to instruct the target large model to define its own role as an expert financial analyst, and adopt the Chain-Of-Thought (COT) reasoning method, based on multiple information summaries, to obtain the characteristic data prediction results of the target object, and the explanatory information for the characteristic data prediction results, that is, based on multiple information summaries, according to the chain logic reasoning steps, to obtain the characteristic data prediction results of the target object, and the explanatory information for the characteristic data prediction results. Among them, the chain logic reasoning steps may include multiple reasoning steps, and the multiple reasoning steps may have a fixed execution order. Based on this, it can be understood that in the embodiment of the present disclosure, the target large model can be based on multiple information summaries, and execute multiple reasoning steps in a fixed execution order to obtain the characteristic data prediction results of the target object, and the explanatory information for the characteristic data prediction results.

[0140] In addition, in the disclosed embodiments, in addition to providing answer guidance information for the target large model, answer format suggestions, answer examples, and answer restrictions, etc. can also be provided for the target large model to provide data structure, arrangement order, and content restriction provisions for the feature data prediction results of the target object and / or the explanatory information for the feature data prediction results, so that the feature data prediction results of the target object and / or the explanatory information for the feature data prediction results are more standardized. Based on this, in one example, after obtaining the answer guidance information and obtaining the chain logic reasoning steps based on the answer guidance information, based on multiple information summaries, according to the chain logic reasoning steps, answer format suggestions, answer examples, and answer restrictions, the feature data prediction results of the target object and the explanatory information for the feature data prediction results can be obtained.

[0141] Further, taking the target object as any financial investment object that needs to be predicted by feature data, and the feature data prediction result as the price trend prediction result of the target object as an example, in the embodiment of the present disclosure, the feature data prediction result of the target object can be represented by a prediction score located in the first numerical interval [0, 10]. Specifically, when the prediction score is lower than 5, the feature data prediction result representing the target object is bearish, and the lower the prediction score, the greater the decline; when the prediction score is equal to 5, the feature data prediction result representing the target object is neutral, that is, neither falling nor rising; when the prediction score is higher than 5, the feature data prediction result representing the target object is bullish, and the higher the prediction score, the greater the rise. In addition, in the embodiment of the present disclosure, the explanatory information for the feature data prediction result can be a clear explanation of the reasoning basis behind the feature data prediction result.

[0142] Through the above method, in the embodiment of the present disclosure, after obtaining the answer guidance information and obtaining the chain logic reasoning steps based on the answer guidance information, the target big model can obtain the feature data prediction results of the target object and the explanatory information for the feature data prediction results based on multiple information summaries and according to the chain logic reasoning steps. The chain logic reasoning steps can provide the target big model with clearer reasoning guidance, help reduce the occurrence rate of reasoning errors, thereby further improving the accuracy of the feature data prediction results. At the same time, based on the explanatory information for the feature data prediction results, it is also possible to enhance the user's trust in the feature data prediction results by improving the interpretability of the feature data prediction results, thereby improving the availability of the feature data prediction results.

[0143] In some optional implementations, the multiple information summaries include news summaries and feature data dynamic summaries, and the chain logic reasoning steps include a first step and a second step. Based on this, the "based on the multiple information summaries, according to the chain logic reasoning steps, obtain the feature data prediction results of the target object, and the explanation information for the feature data prediction results" in step S102 may also include:

[0144] Performing a first step to obtain a first inference result based on the news summary;

[0145] Performing a second step to obtain a second reasoning result based on the dynamic summary of the feature data;

[0146] Based on the first reasoning result and the second reasoning result, a first feature data prediction result of the target object in a short-term future period and explanation information for the first feature data prediction result are obtained.

[0147] Among them, the first step can be: evaluating the degree of influence of the daily news summary and the current time period news summary related to the target object on the target object in the dimension of public opinion news, as the first reasoning result; the second step can be: comparing the dynamic summary of the target object's characteristic data with the price dynamics of the reference object (that is, the financial investment object similar to the target object) and the price dynamics of each financial investment object in the overall financial market, and evaluating the market performance of the target object as the second reasoning result.

[0148] After obtaining the first reasoning result and the second reasoning result, the target large model can be used to weigh the first reasoning result and the second reasoning result, and to reason and summarize the prediction result of the first feature data of the target object in the short-term future period, and to obtain explanatory information for the prediction result of the first feature data. The short-term future period can be the next week, the next month, the next few months, etc.

[0149] Further, taking the target object as any financial investment object that needs to be predicted by feature data, and the feature data prediction result as the price trend prediction result of the target object as an example, in the embodiment of the present disclosure, the first feature data prediction result of the target object in the short-term future period can be characterized by a prediction score located in the second numerical interval [0, 10]. Specifically, when the prediction score is lower than 5, the first feature data prediction result is characterized as bearish, and the lower the prediction score, the greater the decline; when the prediction score is equal to 5, the first feature data prediction result is characterized as neutral, that is, neither falling nor rising; when the prediction score is higher than 5, the first feature data prediction result is characterized as bullish, and the higher the prediction score, the greater the rise. In addition, in the embodiment of the present disclosure, the explanatory information for the first feature data prediction result can be a clear explanation of the reasoning basis behind the first feature data prediction result. In addition, in the embodiment of the present disclosure, the explanatory information for the first feature data prediction result can be a clear explanation of the reasoning basis behind the first feature data prediction result.

[0150] Finally, in the embodiment of the present disclosure, the output result of the target large model can be characterized as:

[0151]

[0152] Among them, N S,τ Used to characterize the daily news summary set in news summary; PN S,t Used to represent the current period news summary in the news summary; P S,t Used to represent dynamic summaries of feature data; The target large model is used to characterize the target object based on the news summary and the dynamic summary of the feature data, and obtain the first feature data prediction result of the target object in the short-term future period, as well as the explanation information for the first feature data prediction result. Here, the first feature data prediction result is obtained by Characterization; the explanation information of the prediction result of the first feature data is obtained by Representation.

[0153] Through the above method, in the embodiment of the present disclosure, the first step can be executed to obtain the first reasoning result based on the news summary, and the second step can be executed to obtain the second reasoning result based on the dynamic summary of the feature data, and based on the first reasoning result and the second reasoning result, the first feature data prediction result of the target object in the short-term future period and the explanatory information for the first feature data prediction result can be obtained. In this process, the chained first step and the second step can provide a clearer reasoning guidance for the target large model, help reduce the occurrence rate of reasoning errors, thereby improving the accuracy of the first feature data prediction result. At the same time, based on the explanatory information for the first feature data prediction result, the user's trust in the first feature data prediction result can be enhanced by improving the interpretability of the first feature data prediction result, thereby improving the availability of the first feature data prediction result.

[0154] In some optional implementations, the multiple information summaries include a macro information summary and a fundamental summary, and the chain logic reasoning step includes a third step and a fourth step. Based on this, the step S102 of "based on the multiple information summaries, according to the chain logic reasoning step, obtaining the characteristic data prediction results of the target object, and the explanation information for the characteristic data prediction results" may also include:

[0155] Performing a third step to obtain a third reasoning result based on the macro information summary;

[0156] Performing a fourth step to obtain a fourth reasoning result based on the fundamental summary;

[0157] Based on the third reasoning result and the fourth reasoning result, a second feature data prediction result of the target object in a long-term future period and explanatory information for the second feature data prediction result are obtained.

[0158] Among them, the third step can be: based on the macro information summary, evaluate the degree of influence of the target object in the dimension of macro market economy (for example, evaluate the impact of the broad macroeconomic pattern on the target object), as the third reasoning result; the fourth step can be: based on the fundamentals summary, analyze the degree of influence of the target object in the dimension of fundamental data (for example, analyze the financial status and future prospects of the target entity), as the fourth reasoning result.

[0159] After obtaining the third and fourth inference results, the target large model can be used to weigh the third and fourth inference results, and to infer and summarize the second feature data prediction results of the target object in the long-term future period, and to obtain explanatory information for the second feature data prediction results. The time length of the long-term future period is greater than the time length of the short-term future period. Specifically, the long-term future period can be one year or several years in the future.

[0160] Further, taking the target object as any financial investment object that needs to be predicted by feature data, and the feature data prediction result as the price trend prediction result of the target object as an example, in the embodiment of the present disclosure, the second feature data prediction result of the target object in the short-term future period can be represented by a prediction score located in the third numerical interval [0, 10]. Specifically, when the prediction score is lower than 5, the second feature data prediction result is characterized as bearish, and the lower the prediction score, the greater the decline; when the prediction score is equal to 5, the second feature data prediction result is characterized as neutral, that is, neither falling nor rising; when the prediction score is higher than 5, the first feature data prediction result is characterized as bullish, and the higher the prediction score, the greater the rise. In addition, in the embodiment of the present disclosure, the explanatory information for the second feature data prediction result can be a clear explanation of the reasoning basis behind the second feature data prediction result.

[0161] Finally, in the embodiment of the present disclosure, the output result of the target large model can be characterized as:

[0162]

[0163] Among them, M t Used to represent macro information summary; F S Used to characterize a summary of fundamentals; The target large model is used to characterize the target object based on the macro information summary and the fundamental summary, and obtain the second feature data prediction results of the target object in the long-term future period, as well as the explanation information for the second feature data prediction results. Here, the second feature data prediction results are obtained through Characterization; the explanation information of the prediction result of the first feature data is obtained by Representation.

[0164] Through the above method, in the embodiment of the present disclosure, the third step can be executed to obtain the third reasoning result based on the macro information summary, and the fourth step can be executed to obtain the fourth reasoning result based on the fundamental summary, and based on the third reasoning result and the fourth reasoning result, the second feature data prediction result of the target object in the long-term future period and the explanatory information for the second feature data prediction result can be obtained. In this process, the chained third step and the fourth step can provide a clearer reasoning guidance for the target large model, help reduce the occurrence rate of reasoning errors, thereby improving the accuracy of the second feature data prediction result. At the same time, based on the explanatory information for the second feature data prediction result, the user's trust in the feature data prediction result can be enhanced by improving the interpretability of the second feature data prediction result, thereby improving the availability of the second feature data prediction result.

[0165] In some optional implementations, the multiple information summaries include news summaries, feature data dynamic summaries, macro information summaries, and fundamentals summaries, and the chain logic reasoning steps include the first step, the second step, the third step, and the fourth step. Based on this, the "based on the multiple information summaries, according to the chain logic reasoning steps, obtain the feature data prediction results of the target object, and the explanation information for the feature data prediction results" in step S102 may also include:

[0166] Performing a first step to obtain a first inference result based on the news summary;

[0167] Performing a second step to obtain a second reasoning result based on the dynamic summary of the feature data;

[0168] Performing a third step to obtain a third reasoning result based on the macro information summary;

[0169] Performing a fourth step to obtain a fourth reasoning result based on the fundamental summary;

[0170] Based on the first reasoning result, the second reasoning result, the third reasoning result and the fourth reasoning result, the first feature data prediction result of the target object in the short-term future period, the second feature data prediction result of the target object in the long-term future period, and the explanatory information for the first feature data prediction result and the second feature data prediction result are obtained.

[0171] Among them, the first step can be: evaluating the degree of influence of the daily news summary and the current period news summary related to the target object on the target object in the dimension of public opinion news, as the first reasoning result; the second step can be: comparing the dynamic summary of the characteristic data of the target object with the price dynamics of the reference object (that is, the financial investment object similar to the target object) and the price dynamics of each financial investment object in the overall financial market, and evaluating the market performance of the target object as the second reasoning result; the third step can be: based on the macro information summary, evaluating the degree of influence of the target object in the dimension of macro market economy (for example, evaluating the broad macroeconomic pattern and its impact on the target object), as the third reasoning result; the fourth step can be: based on the fundamental summary, analyzing the degree of influence of the target object in the dimension of fundamental data (for example, analyzing the financial status and future prospects of the target entity), as the fourth reasoning result.

[0172] After obtaining the first reasoning step, the second reasoning step, the third reasoning step and the fourth reasoning step, the target large model can be used to weigh the first reasoning result, the second reasoning result, the third reasoning result and the fourth reasoning result, and the first feature data prediction result of the target object in the short-term future period can be reasoned and summarized (specifically, the first reasoning result and the second reasoning result can be combined to reason and summarize the first feature data prediction result of the target object in the short-term future period), the second feature data prediction result of the target object in the long-term future period can be reasoned and summarized (specifically, the third reasoning step and the fourth reasoning step can be combined to reason and summarize the second feature data prediction result of the target object in the short-term future period), and the explanation information for the first feature data prediction result and the second feature data prediction result can be obtained. Among them, the short-term future period can be the next week, the next month, the next few months, etc.; the time length of the long-term future period is greater than the time length of the short-term future period. Specifically, the long-term future period can be the next year, several years, etc.

[0173] In the embodiment of the present disclosure, the output result of the target large model can be characterized as follows:

[0174]

[0175] Among them, N S,τ Used to characterize the daily news summary set in news summary; PN S,t Used to represent the current period news summary in the news summary; P S,t Used to represent dynamic summary of feature data; M t Used to represent macro information summary; F S Used to characterize a summary of fundamentals; The target large model is used to characterize the first feature data prediction result of the target object in the short-term future period, the second feature data prediction result of the target object in the long-term future period, and the explanation information for the first feature data prediction result and the second feature data prediction result based on the first reasoning result, the second reasoning result, the third reasoning result and the fourth reasoning result. Here, the first feature data prediction result is obtained by Characterization; The second feature data prediction result is passed Characterization; Explanation information for the prediction result of the first feature data and the prediction result of the second feature data is obtained by Representation.

[0176] Through the above manner, in the embodiment of the present disclosure, a clearer reasoning guidance can be provided for the target large model through the chained first step, second step, third step and fourth step, which helps to reduce the occurrence rate of reasoning errors, thereby improving the accuracy of the first feature data prediction result and the second feature data prediction result. At the same time, based on the explanatory information for the first feature data prediction result and the second feature data prediction result, the user's trust in the first feature data prediction result and the second feature data prediction result can be enhanced by improving the interpretability of the first feature data prediction result and the second feature data prediction result, thereby improving the availability of the first feature data prediction result and the second feature data prediction result.

[0177] The following will be combined Figure 6 and Figure 7 , the integrity process of a data prediction method provided by an embodiment of the present disclosure is described.

[0178] Step S601, obtain N groups of daily public opinion news related to the target object; take each group of daily public opinion news in the N groups of daily public opinion news as the public opinion news to be processed, use the first news summarizer to process the public opinion news to be processed, and obtain the daily news summary corresponding to the public opinion news to be processed; obtain the historical time period news summary corresponding to the historical time period; use the second news summarizer to process the daily news summary set based on the historical time period news summary, and obtain the current time period news summary corresponding to the current time period; wherein, the N groups of daily public opinion news correspond one-to-one to N independent dates in the current time period; N≥2 and is an integer.

[0179] The first news summarizer and the second news summarizer may be the same news summarizer, and the news summarizer may be a large language model.

[0180] In addition, it should be noted that, in the embodiment of the present disclosure, the specific description of step S601 can refer to the relevant description in the above-mentioned step S101, and will not be repeated here.

[0181] Step S602, obtaining the first market data of the target object; obtaining the second market data of a first target number of reference objects similar to the target object; using a feature data dynamic summarizer, based on the first market data of the target object, the second market data of the first target number of reference objects, and a specific reference index, obtaining a dynamic summary of the feature data.

[0182] Among them, the feature data dynamic summarizer can be a large language model.

[0183] In addition, it should be noted that, in the embodiment of the present disclosure, the specific description of step S602 can refer to the relevant description in the above-mentioned step S101, and will not be repeated here.

[0184] Step S603, using the first macro summarizer, processing the second target number of macro research reports to obtain a second target number of macro research summaries corresponding one to one with the second target number of macro research reports; using the second macro summarizer, processing the macro indicator data to obtain a macro indicator summary; using the third macro summarizer, obtaining a macro information summary based on the second target number of macro research summaries and the macro indicator summary.

[0185] The first macro-summarizer, the second macro-summarizer and the third macro-summarizer may be the same macro-summarizer, and the macro-summarizer may be a large language model.

[0186] In addition, it should be noted that, in the embodiment of the present disclosure, the specific description of step S603 can refer to the relevant description in the above-mentioned step S101, and will not be repeated here.

[0187] Step S604, using the first fundamental summarizer to process the management analysis results of the target entity to obtain a management fundamental summary; wherein the target entity is the economic organization that issues the target object; using the second fundamental summarizer to process the analyst research report related to the target entity to obtain the analyst fundamental summary; using the third fundamental summarizer to obtain the fundamental summary based on the management fundamental summary, the analyst fundamental summary, and the specific data of the target entity.

[0188] The first fundamental summarizer, the second fundamental summarizer and the third fundamental summarizer may be the same fundamental summarizer, and the fundamental summarizer may be a large language model.

[0189] In addition, it should be noted that, in the embodiment of the present disclosure, the specific description of step S604 can refer to the relevant description in the above-mentioned step S101, and will not be repeated here.

[0190] Step S605, using the target big model, obtain answer guidance information; based on the answer guidance information, obtain chain logic reasoning steps; based on multiple information summaries, according to the chain logic reasoning steps, obtain the feature data prediction results of the target object and the explanatory information for the feature data prediction results.

[0191] In some optional implementations, the multiple information summaries include a macro information summary and a fundamental summary, and the chain logic reasoning step includes a third step and a fourth step. Based on this, the step S605 of "based on the multiple information summaries, according to the chain logic reasoning step, obtaining the characteristic data prediction results of the target object, and the explanation information for the characteristic data prediction results" may also include:

[0192] Performing a first step to obtain a first inference result based on the news summary;

[0193] Performing a second step to obtain a second reasoning result based on the dynamic summary of the feature data;

[0194] Performing a third step to obtain a third reasoning result based on the macro information summary;

[0195] Performing a fourth step to obtain a fourth reasoning result based on the fundamental summary;

[0196] Based on the first reasoning result, the second reasoning result, the third reasoning result and the fourth reasoning result, the first feature data prediction result of the target object in the short-term future period, the second feature data prediction result of the target object in the long-term future period, and the explanatory information for the first feature data prediction result and the second feature data prediction result are obtained.

[0197] In addition, it should be noted that, in the embodiment of the present disclosure, the specific description of step S605 can refer to the relevant description in the above-mentioned step S102, and will not be repeated here.

[0198] It can be understood that in the embodiments of the present disclosure, an Exchange-of-Thought (EoT) framework is actually introduced, that is, four independent large language models, namely, news summarizer, feature data dynamic summarizer, macro summarizer and fundamental summarizer, are used to form a star network topology structure (with the target large model as the central node, collecting and summarizing the output results of the news summarizer, feature data dynamic summarizer, macro summarizer and fundamental summarizer), which solves the problem of difficult information exchange between the news summarizer, feature data dynamic summarizer, macro summarizer and fundamental summarizer, and improves the reasoning ability of the target large model, so that when the target large model is used to obtain the feature data prediction results of the target object and the explanatory information of the feature data prediction results, the accuracy of the feature data prediction results can be further improved, and the interpretability of the feature data prediction results can be further improved.

[0199] See also Figure 8 , which is a scenario diagram of a data prediction method provided in an embodiment of the present disclosure.

[0200] As mentioned above, the data prediction method provided by the embodiment of the present disclosure is applied to an electronic device. The electronic device may be a server or a terminal device. Here, the terminal device may be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a vehicle-mounted computer, etc.), a personal digital processing or other similar computing device.

[0201] Electronic devices are used for:

[0202] Acquire multiple information summaries related to the target object; wherein the multiple information summaries have different information categories, so as to characterize the performance and / or impact degree of the target object in different dimensions;

[0203] By using the target large model and based on multiple information summaries, the feature data prediction results of the target object and the explanatory information for the feature data prediction results are obtained.

[0204] It should be noted that in the embodiments of the present disclosure, Figure 8 The scene diagram shown is only illustrative and not restrictive. Those skilled in the art can Figure 8 Various obvious changes and / or substitutions are made to the examples, and the obtained technical solutions still fall within the scope of the disclosure of the embodiments of the present disclosure.

[0205] The present disclosure provides a large model training method, which can be applied to electronic devices. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a car computer, etc.), a personal digital processing unit or other similar computing device. Fig. 9 The flowchart diagram shown illustrates a large model training method provided by an embodiment of the present disclosure. It should be noted that although a logical order is shown in the flowchart diagram, in some cases, the steps shown or described may also be performed in other orders.

[0206] Step S901: Acquire multiple information summary samples related to the target object sample.

[0207] The target object sample may be any financial investment object (eg, stocks, funds, futures) that requires feature data trend prediction. Here, the feature data may be the price trend of the target object.

[0208] In addition, in the embodiment of the present disclosure, the multiple information summary samples have different information categories to characterize the performance and / or degree of influence of the target object sample in different dimensions. For example, the multiple information summary samples may include at least one of a news summary sample, a feature data dynamic summary sample, a macro information summary sample, and a fundamental summary sample.

[0209] Among them, the news summary sample can be a summary information sample obtained by processing the public opinion news related to the target object sample, so as to characterize the degree of influence of the target object sample in the dimension of public opinion news; the feature data dynamic summary sample can be a summary information sample obtained by processing the stock price dynamic data of the target object sample, so as to characterize the market performance of the target object sample; the macro information summary sample can be a summary information sample obtained by processing the macro market economic data, so as to characterize the degree of influence of the target object sample in the dimension of macro market economy; the fundamental summary sample can be a summary information sample obtained by processing the fundamental data related to the target object sample, so as to characterize the degree of influence of the target object sample in the dimension of fundamental data.

[0210] Step S902: using the initial large model, based on multiple information summary samples, obtain a feature data prediction result sample for the target object sample and an explanation information sample for the feature data prediction result sample.

[0211] Among them, the feature data prediction result sample can be a price trend prediction result sample of the target object.

[0212] In one example, a sample of answer guidance information may be obtained, and the sample of answer guidance information and multiple information summary samples may be input into the initial large model, so that the initial large model processes multiple information summary samples according to the sample of answer guidance information to obtain a sample of feature data prediction result of the target object sample, and an explanation information sample for the sample of feature data prediction result. The sample of answer guidance information is the eleventh prompt information, which is used to instruct the initial large model how to obtain a sample of feature data prediction result of the target object sample, and an explanation information sample for the sample of feature data prediction result, based on multiple information summary samples.

[0213] In addition, in the disclosed embodiment, the initial large model is the initial signal generator, and the initial signal generator can be a large language model. Here, the large language model can be a pre-trained neural network model, which has general language knowledge, world knowledge, professional knowledge in various fields, etc., and is stored in the model in the form of parameters. Specifically, the initial large model can be an autoregressive generative model of the Transformer architecture.

[0214] Step S903: Based on the feature data prediction result samples and the explanation information samples, the initial large model is trained to obtain the target large model.

[0215] In one example, real feature data for a target object sample can be obtained, and the initial large model can be reflected on the feature data prediction result sample, the explanation information sample and the real feature data, thereby achieving the purpose of training the initial large model to obtain the target large model.

[0216] By adopting the large model training method provided by the embodiment of the present disclosure, multiple information summary samples related to the target object sample can be obtained, and the initial large model is used to obtain the feature data prediction result sample for the target object sample based on the multiple information summary samples, and the explanatory information sample for the feature data prediction result sample, and then the initial large model is trained based on the feature data prediction result sample and the explanatory information sample, that is, the parameters of the initial large model are fine-tuned to obtain the target large model. In this way, the performance of the target large model can be improved, so that when the target large model is used to obtain the feature data prediction result of the target object and the explanatory information for the feature data prediction result, the accuracy of the feature data prediction result can be further improved, and the interpretability of the feature data prediction result can be further improved.

[0217] In some optional implementations, step S903, that is, “training the initial large model based on the feature data prediction result sample and the explanation information sample to obtain the target large model” may include:

[0218] Obtaining real feature data for target object samples;

[0219] Generate reflection evidence including multiple information summary samples, feature data prediction result samples, explanation information samples and real feature data;

[0220] Using the self-reflection model, based on the reflection basis, the current reflection results for the feature data prediction result samples and the explanation information samples are obtained;

[0221] Based on the current reflection results, the initial large model is trained to obtain the target large model.

[0222] Among them, the multiple information summary samples may include at least one of a news summary sample, a feature data dynamic summary sample, a macro information summary sample, and a fundamental summary sample; the feature data prediction result sample may include a first feature data prediction result sample of the target object sample in a short-term future period, and a second feature data prediction result sample of the target object sample in a long-term future period; the explanation information sample may be a clear explanation of the reasoning basis behind the first feature data prediction result sample and the second feature data prediction result sample. Here, the short-term future period may be the next week, the next month, the next few months, etc.; the time length of the long-term future period is greater than the time length of the short-term future period, specifically, the long-term future period may be the next year, several years, etc.

[0223] In one example, “generating a reflection basis including a plurality of information summary samples, feature data prediction result samples, explanation information samples, and real feature data” may include:

[0224] Using the target scoring model, obtaining scoring results for feature data prediction result samples and explanation information samples, as well as scoring explanations for the scoring results;

[0225] Generates reflection evidence including multiple information summary samples, feature data prediction result samples, explanation information samples, scoring results, scoring explanations and real feature data.

[0226] The target scoring model may be a multi-layer perceptron model (Multi-Layer Perceptron, MLP).

[0227] In the disclosed embodiment, the target scoring model can be used to evaluate the consistency between the feature data prediction result sample and the actual feature data to obtain the scoring result, and combined with the explanatory information sample to obtain the scoring explanation for the scoring result, and then generate a reflection basis including multiple information summary samples, feature data prediction result samples, explanatory information samples, scoring results, scoring explanations and actual feature data.

[0228] Furthermore, in a specific example, a new model fine-tuning method (e.g., PromptTuning) can be used to train the initial scoring module to obtain a target scoring model, thereby improving the performance of the target scoring model, thereby improving the accuracy of the scoring results for the feature data prediction result samples and the explanation information samples, and improving the accuracy of the scoring explanation for the scoring results. The process may specifically include:

[0229] Acquire a first sample pair; wherein the first sample pair includes a first feature data prediction sample and a first explanation information sample, and first real feature data corresponding to the first feature data prediction sample; and the similarity between the first feature data prediction sample and the first real feature data is greater than or equal to a first preset similarity threshold;

[0230] Acquire a second sample pair; wherein the second sample pair includes a second feature data prediction sample and a second interpretation information sample, and second real feature data corresponding to the second feature data prediction sample; the similarity between the second feature data prediction sample and the second real feature data is less than a second similarity threshold, and the second similarity threshold is less than a first preset similarity threshold;

[0231] Based on the first sample pair and the second sample pair, the initial scoring module is trained to obtain a target scoring model.

[0232] The first sample pair is a positive sample pair; the first similarity threshold can be set according to actual application requirements, for example, it can be set to 0.95; the second sample pair is a negative sample pair; the second similarity threshold can be set according to actual application requirements, for example, it can be set to 0.05. In addition, in the embodiment of the present disclosure, the first sample pair and the second sample pair can be collected during the training of the initial large model and stored in the comparison data set.

[0233] After obtaining the first sample pair (including the first feature data prediction sample and the first explanation information sample), the initial scoring model can learn how to obtain the first scoring result sample for the first sample pair based on the similarity between the first feature data prediction sample and the first real feature data (the higher the first scoring result sample, the more accurate the first feature data prediction result sample is). Specifically, it can learn how to evaluate the consistency between the first feature data prediction sample and the first real feature data included in the first sample pair to obtain the first scoring result sample, and combine it with the first explanation information sample to obtain the first scoring explanation sample for the first scoring result sample.

[0234] Similarly, after obtaining the second sample pair (including the second feature data prediction sample and the second explanation information sample), the initial scoring model can learn how to obtain the second scoring result sample for the second sample pair based on the similarity between the second feature data prediction sample and the second real feature data (the higher the second scoring result sample, the more accurate the second feature data prediction result sample is). Specifically, it can learn how to evaluate the consistency between the second feature data prediction sample and the second real feature data included in the second sample pair to obtain the second scoring result sample, and combine with the second explanation information sample to obtain the second scoring explanation sample for the second scoring result sample.

[0235] It should be noted that in the embodiments of the present disclosure, the model parameters of the initial scoring model can be adjusted with the goal of maximizing the sum of the first scoring result sample and the second scoring result sample, so as to train the initial scoring module and obtain the target scoring model.

[0236] It should also be noted that in the embodiment of the present disclosure, when the reflection basis includes multiple information summary samples, feature data prediction result samples, explanation information samples, scoring results, scoring explanations and real feature data, "using the self-reflection model, based on the reflection basis, to obtain the current reflection results for the feature data prediction result samples and the explanation information samples" can be characterized as:

[0237]

[0238] Among them, N S,τ ′ is used to characterize the daily news summary set samples in the news summary samples; PNS,t ′ is used to characterize the current period news summary sample in the news summary sample; P S,t ′ is used to characterize the dynamic summary sample of feature data; M t ′ is used to characterize the macro information summary sample; F S ′ is used to characterize the fundamental summary sample; Used to characterize the first feature data prediction result sample; Used to characterize the prediction result sample of the second feature data; It is used to characterize the explanation information sample for the first feature data prediction result sample and the second feature data prediction result sample; SR is used to characterize the scoring result for the feature data prediction result sample and the explanation information sample; SE is used to characterize the scoring explanation for the scoring result; y aspt Real feature data used to characterize target object samples; It is used to characterize the self-reflection model, and based on the reflection basis, obtain the current reflection results for the feature data prediction result samples and the explanation information samples, that is, It is specifically used to characterize the current reflection results for feature data prediction result samples and explanation information samples.

[0239] Furthermore, in the disclosed embodiment, the self-reflective model may be a large language model whose model parameters are frozen. In a specific example, the initial large model may be trained in the following manner to obtain a target large model, thereby improving the performance of the target large model, so that when the target large model is used to obtain the feature data prediction results of the target object and the explanation information for the feature data prediction results, the accuracy of the feature data prediction results can be further improved, and the interpretability of the feature data prediction results can be further improved. The process may specifically include:

[0240] Obtain historical reflection results;

[0241] Based on the current reflection results and historical reflection results, long-term reflection data is obtained;

[0242] Based on long-term reflection data, the initial large model is trained to obtain the target large model.

[0243] Among them, the current reflection result can be regarded as the current short-term memory; the historical reflection result can be regarded as the historical short-term memory, which is stored in the memory bank; the long-term reflection data can be the long-term memory obtained by splicing the current reflection result and the historical reflection result, which can be specifically characterized as:

[0244]

[0245] in, Used to represent the current reflection results; Used to characterize the results of historical reflection; Used to represent long-term reflective data.

[0246] After obtaining the long-term reflection data, the model parameters of the initial large model can be adjusted based on the long-term reflection data to achieve the purpose of training the initial large model and obtaining the target large model. This process can also be regarded as the next round of parameter iteration process of the initial large model, which can be specifically characterized as:

[0247]

[0248] Among them, N S,τ ″ is used to characterize the daily news summary set sample in the news summary sample when performing the next round of parameter iteration; PN S,t ″ is used to represent the news summary sample of the current period in the news summary sample when performing the next round of parameter iteration; P S,t ″ is used to characterize the dynamic summary sample of feature data when performing the next round of parameter iteration; M t ″ is used to characterize the macro information summary sample when performing the next round of parameter iteration; F S "Used to characterize the fundamental summary sample when performing the next round of parameter iteration; Used to represent long-term reflective data; It is used to represent the first feature data prediction result sample of the target object sample obtained in the short-term future period when performing the next round of parameter iteration (that is, ), the second feature data prediction result sample of the target object sample in the long-term future period (that is, ), and the explanation information samples behind the first feature data prediction result samples and the second feature data prediction result samples (that is, ).

[0249] Through the above method, in the embodiment of the present disclosure, the real feature data for the target object sample can be obtained, and a reflection basis including multiple information summary samples, feature data prediction result samples, explanation information samples and real feature data can be generated, and then the self-reflection model is used to obtain the current reflection results for the feature data prediction result samples and the explanation information samples based on the reflection basis. Finally, based on the current reflection results, the initial large model is trained to obtain the target large model. That is to say, in the embodiment of the present disclosure, the prediction ability of the initial large model can be improved by allowing the initial large model to reflect on itself. Specifically, by connecting the initial large model and the self-reflection model, it is deployed into an automated agent model containing two neural network models, and the automated agent model can be improved through verbal self-reflection, cyclically and iteratively, that is, the parameters of the initial large model are fine-tuned. At the same time, when fine-tuning the parameters of the initial large model, it is not selected to use manual feedback, but based on the reflection results (including current reflection results and historical reflection results) generated by the automated agent model to cyclically and iteratively improve, so as to further improve the learning effect of the initial large model.

[0250] The following will be combined Fig.10 and 11 , the integrity process of a large model training method provided by an embodiment of the present disclosure is described.

[0251] Step S1001: Acquire multiple information summary samples related to the target object sample.

[0252] For the specific description of step S1001, please refer to the relevant description in the above step S901, which will not be repeated here.

[0253] Step S1002, using the initial large model, based on multiple information summary samples, obtain a feature data prediction result sample for the target object sample and an explanation information sample for the feature data prediction result sample.

[0254] For the specific description of step S1002, please refer to the relevant description in the above step S902, which will not be repeated here.

[0255] Step S1003, using the target scoring model, obtaining scoring results for the feature data prediction result samples and the explanation information samples, as well as scoring explanations for the scoring results.

[0256] For the specific description of step S1003, please refer to the relevant description in the above step S903, which will not be repeated here.

[0257] Step S1004, generating a reflection basis including a plurality of information summary samples, feature data prediction result samples, explanation information samples and real feature data.

[0258] For the specific description of step S1004, please refer to the relevant description in the above step S903, which will not be repeated here.

[0259] Step S1005, using the self-reflection model, based on the reflection basis, obtain the current reflection results for the feature data prediction result sample and the explanation information sample.

[0260] For the specific description of step S1005, please refer to the relevant description in the above step S903, which will not be repeated here.

[0261] Step S1006, obtaining historical reflection results; based on the current reflection results and historical reflection results, obtaining long-term reflection data.

[0262] For the specific description of step S1006, please refer to the relevant description in the above step S903, which will not be repeated here.

[0263] Step S1007, based on the long-term reflection data, the initial large model is trained to obtain the target large model.

[0264] For the specific description of step S1007, please refer to the relevant description in the above step S903, which will not be repeated here.

[0265] See also Fig.12 , which is a scenario diagram of a large model training method provided in an embodiment of the present disclosure.

[0266] As mentioned above, the large model training method provided by the embodiment of the present disclosure is applied to an electronic device. The electronic device may be a server or a terminal device. Here, the terminal device may be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a vehicle-mounted computer, etc.), a personal digital processing or other similar computing device.

[0267] Electronic devices are used for:

[0268] Acquire multiple information summary samples related to the target object sample; wherein the multiple information summary samples have different information categories, so as to characterize the performance and / or impact degree of the target object sample in different dimensions;

[0269] Using the initial large model, based on multiple information summary samples, a feature data prediction result sample for the target object sample and an explanation information sample for the feature data prediction result sample are obtained;

[0270] Based on the feature data prediction result samples and the explanation information samples, the initial large model is trained to obtain the target large model.

[0271] It should be noted that in the embodiments of the present disclosure, Fig.12The scene diagram shown is only illustrative and not restrictive. Those skilled in the art can Fig.12 Various obvious changes and / or substitutions are made to the examples, and the obtained technical solutions still fall within the scope of the disclosure of the embodiments of the present disclosure.

[0272] In order to better implement the data prediction method, the embodiment of the present disclosure also provides a data prediction device, which can be integrated in an electronic device. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a mainframe computer, a conventional computer (for example, a desktop computer, a laptop computer, a car computer, etc.), a personal digital processing or other similar computing device. Fig.13 The schematic structural block diagram shown illustrates a data prediction device 1300 provided in the disclosed embodiment.

[0273] The data prediction device 1300 includes:

[0274] The summary information acquisition unit 1301 is used to acquire multiple information summaries related to the target object; wherein the multiple information summaries have different information categories, so as to characterize the performance and / or impact degree of the target object in different dimensions;

[0275] The feature data prediction unit 1302 is used to obtain the feature data prediction result of the target object and the explanation information for the feature data prediction result based on multiple information summaries by using the target large model.

[0276] In some optional implementations, the feature data prediction unit 1302 is used to:

[0277] Get answer guidance information;

[0278] Based on the answer guidance information, the chain logic reasoning steps are obtained;

[0279] Based on multiple information summaries and following the chain logic reasoning steps, the feature data prediction results of the target object and the explanatory information for the feature data prediction results are obtained.

[0280] In some optional implementations, the multiple information summaries include a news summary and a feature data dynamic summary; the chain logic reasoning step includes a first step and a second step; the feature data prediction unit 1302 is used to:

[0281] Performing a first step to obtain a first inference result based on the news summary;

[0282] Performing a second step to obtain a second reasoning result based on the dynamic summary of the feature data;

[0283] Based on the first reasoning result and the second reasoning result, a first feature data prediction result of the target object in a short-term future period and explanation information for the first feature data prediction result are obtained.

[0284] In some optional implementations, the multiple information summaries include a macro information summary and a fundamental summary; the chain logic reasoning step includes a third step and a fourth step; the feature data prediction unit 1302 is used to:

[0285] Performing a third step to obtain a third reasoning result based on the macro information summary;

[0286] Performing a fourth step to obtain a fourth reasoning result based on the fundamental summary;

[0287] Based on the third reasoning result and the fourth reasoning result, a second feature data prediction result of the target object in a long-term future period and explanatory information for the second feature data prediction result are obtained.

[0288] In some optional implementations, the multiple information summaries include a news summary, a feature data dynamic summary, a macro information summary, and a fundamental summary; the chain logic reasoning steps include a first step, a second step, a third step, and a fourth step; the feature data prediction unit 1302 is used to:

[0289] Performing a first step to obtain a first inference result based on the news summary;

[0290] Performing a second step to obtain a second reasoning result based on the dynamic summary of the feature data;

[0291] Performing a third step to obtain a third reasoning result based on the macro information summary;

[0292] Performing a fourth step to obtain a fourth reasoning result based on the fundamental summary;

[0293] Based on the first reasoning result, the second reasoning result, the third reasoning result and the fourth reasoning result, the first feature data prediction result of the target object in the short-term future period, the second feature data prediction result of the target object in the long-term future period, and the explanatory information for the first feature data prediction result and the second feature data prediction result are obtained.

[0294] In some optional implementations, the multiple information summaries include news summaries, and the news summaries include daily news summaries; the summary information acquisition unit 1301 is used to:

[0295] Obtain N groups of daily public opinion news related to the target object; wherein the N groups of daily public opinion news correspond to N independent dates in the current period one by one; N ≥ 2 and is an integer;

[0296] Each group of daily public opinion news in the N groups of daily public opinion news is used as the public opinion news to be processed, and the public opinion news to be processed is processed by using the first news summarizer to obtain the daily news summary corresponding to the public opinion news to be processed.

[0297] In some optional implementations, the news summary also includes a news summary for the current period; and the summary information acquisition unit 1301 is further configured to:

[0298] Based on N daily news summaries corresponding to N groups of daily public opinion news, a daily news summary set is obtained;

[0299] Get the historical period news summary corresponding to the historical period;

[0300] The second news summarizer is used to process the daily news summary set based on the historical time period news summary to obtain the current time period news summary corresponding to the current time period.

[0301] In some optional implementations, the multiple information summaries include a dynamic summary of feature data; the summary information acquisition unit 1301 is used to:

[0302] Get the first market data of the target object;

[0303] Acquire second market data of a first target number of reference objects similar to the target object;

[0304] The feature data dynamic summarizer is used to obtain the feature data dynamic summary based on the first market data of the target object, the second market data of the first target number of reference objects, and the specific reference index.

[0305] In some optional implementations, the multiple information summaries include a macro information summary; the summary information acquisition unit 1301 is used to:

[0306] Using the large language model, the second target number of macro research reports are processed to obtain the second target number of macro research summaries corresponding to the second target number of macro research reports;

[0307] Use the large language model to process the macro indicator data and obtain the macro indicator summary;

[0308] Using the large language model, a macro information summary is obtained based on the second target number of macro research summaries and macro indicator summaries.

[0309] In some optional implementations, the multiple information summaries include a basic surface summary; the summary information acquisition unit 1301 is used to:

[0310] Using the large language model, the management analysis results of the target entity are processed to obtain a summary of the management fundamentals; the target entity is the economic organization that issues the target object;

[0311] Using the large language model, we process analyst research reports related to the target entity and obtain the analyst fundamental summary;

[0312] Using a large language model, a fundamental summary is obtained based on management fundamental summary, analyst fundamental summary, and specific data of the target entity.

[0313] In some optional implementations, the summary information obtaining unit 1301 is used to:

[0314] Obtain the first analyst research report of the target entity;

[0315] Obtain second analyst research reports on a third target number of reference entities that are similar to the target entity;

[0316] The first analyst research report of the target entity and the second analyst research report of the third target quantitative reference entity are taken as analyst research reports related to the target entity, and the analyst research reports related to the target entity are processed to obtain the analyst fundamental summary.

[0317] For the description of the specific functions and examples of each unit of the data prediction device 1300 of the embodiment of the present disclosure, reference can be made to the relevant description of the corresponding steps in the aforementioned data prediction method embodiment, which will not be repeated here.

[0318] In order to better implement the large model training method, the embodiment of the present disclosure also provides a large model training device, which can be integrated in an electronic device. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a mainframe computer, a conventional computer (for example, a desktop computer, a laptop computer, a car computer, etc.), a personal digital processing or other similar computing device. Fig.14 The schematic structural block diagram shown illustrates a large model training device 1400 provided in the disclosed embodiment.

[0319] The large model training device 1400 includes:

[0320] The first sample acquisition unit 1401 is used to acquire a plurality of information summary samples related to the target object sample; wherein the plurality of information summary samples have different information categories, so as to characterize the performance and / or the degree of influence of the target object sample in different dimensions;

[0321] The second sample acquisition unit 1402 is used to obtain a feature data prediction result sample for the target object sample and an explanation information sample for the feature data prediction result sample based on the multiple information summary samples using the initial large model;

[0322] The large model training unit 1403 is used to train the initial large model based on the feature data prediction result samples and the explanation information samples to obtain the target large model.

[0323] In some optional implementations, the large model training unit 1403 is used to:

[0324] Obtaining real feature data for target object samples;

[0325] Generate reflection evidence including multiple information summary samples, feature data prediction result samples, explanation information samples and real feature data;

[0326] Using the self-reflection model, based on the reflection basis, the current reflection results for the feature data prediction result samples and the explanation information samples are obtained;

[0327] Based on the current reflection results, the initial large model is trained to obtain the target large model.

[0328] In some optional implementations, the large model training unit 1403 is used to:

[0329] Using the target scoring model, obtaining scoring results for feature data prediction result samples and explanation information samples, as well as scoring explanations for the scoring results;

[0330] Generates reflection evidence including multiple information summary samples, feature data prediction result samples, explanation information samples, scoring results, scoring explanations and real feature data.

[0331] In some optional implementations, the large model training apparatus 1400 further includes a scoring large model training unit, which is used to:

[0332] Acquire a first sample pair; wherein the first sample pair includes a first feature data prediction sample and a first explanation information sample, and first real feature data corresponding to the first feature data prediction sample; and the similarity between the first feature data prediction sample and the first real feature data is greater than or equal to a first preset similarity threshold;

[0333] Acquire a second sample pair; wherein the second sample pair includes a second feature data prediction sample and a second interpretation information sample, and second real feature data corresponding to the second feature data prediction sample; the similarity between the second feature data prediction sample and the second real feature data is less than a second similarity threshold, and the second similarity threshold is less than a first preset similarity threshold;

[0334] Based on the first sample pair and the second sample pair, the initial scoring module is trained to obtain a target scoring model.

[0335] In some optional implementations, the large model training unit 1403 is used to:

[0336] Obtain historical reflection results;

[0337] Based on the current reflection results and historical reflection results, long-term reflection data is obtained;

[0338] Based on long-term reflection data, the initial large model is trained to obtain the target large model.

[0339] The specific functions and examples of each unit of the large model training device 1400 of the embodiment of the present disclosure can be found in the relevant description of the corresponding steps in the aforementioned large model training method embodiment, which will not be repeated here. In the technical solution of the present disclosure, the acquisition, storage and application of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good customs.

[0340] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0341] Fig.15 A schematic block diagram of an example electronic device 1500 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0342] like Fig.15As shown, the device 1500 includes a computing unit 1501, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1502 or a computer program loaded from a storage unit 1508 to a random access memory (RAM) 1503. In the RAM 1503, various programs and data required for the operation of the device 1500 can also be stored. The computing unit 1501, the ROM 1502, and the RAM 1503 are connected to each other via a bus 1504. An input / output (I / O) interface 1505 is also connected to the bus 1504.

[0343] A number of components in the device 1500 are connected to the I / O interface 1505, including: an input unit 1506, such as a keyboard, a mouse, etc.; an output unit 1507, such as various types of displays, speakers, etc.; a storage unit 1508, such as a disk, an optical disk, etc.; and a communication unit 1509, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1509 allows the device 1500 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0344] The computing unit 1501 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSP), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 1501 performs the various methods and processes described above, such as data prediction methods and / or large model training methods. For example, in some embodiments, the data prediction method and / or large model training method may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 1508. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 1500 via the ROM 1502 and / or the communication unit 1509. When the computer program is loaded into RAM 1503 and executed by the computing unit 1501, one or more steps in the data prediction method and / or large model training method described above may be performed. Alternatively, in other embodiments, the computing unit 1501 may be configured as a data prediction method and / or a large model training method in any other appropriate manner (e.g., by means of firmware).

[0345] Various embodiments of the systems and techniques described above herein may be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0346] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0347] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a RAM, a ROM, an erasable programmable read-only memory (EPROM) or a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0348] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a cathode ray tube (CRT) display or a liquid crystal display (LCD)) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices are also used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0349] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: Local Area Network (LAN), Wide Area Network (WAN), and the Internet.

[0350] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship to each other. The server may be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0351] An embodiment of the present disclosure also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to execute a data prediction method and / or a large model training method.

[0352] The embodiments of the present disclosure also provide a computer program product, including a computer program, which implements a data prediction method and / or a large model training method when executed by a processor.

[0353] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in the present disclosure can be executed in parallel, sequentially or in a different order, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, and this document is not limited here. In addition, in the present disclosure, relational terms such as "first", "second", "third", etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. In addition, "multiple" in the present disclosure can be understood as at least two.

[0354] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the principles of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A data prediction method, include: Acquire multiple information summaries related to the target object; wherein the multiple information summaries have different information categories, so as to characterize the performance and / or impact degree of the target object in different dimensions; The target macro model is used to obtain the feature data prediction result of the target object and the explanation information for the feature data prediction result based on the multiple information summaries.

2. The method according to claim 1, in, The step of obtaining the feature data prediction result of the target object based on the multiple information summaries and explanation information for the feature data prediction result includes: Get answer guidance information; Based on the answer guidance information, chain logic reasoning steps are obtained; Based on the multiple information summaries and in accordance with the chain logic reasoning steps, the feature data prediction results of the target object and the explanation information for the feature data prediction results are obtained.

3. The method according to claim 2, in, The multiple information summaries include news summaries and feature data dynamic summaries; the chain logic reasoning step includes a first step and a second step; based on the multiple information summaries, according to the chain logic reasoning step, the feature data prediction result of the target object and the explanation information for the feature data prediction result are obtained, including: Execute the first step to obtain a first reasoning result based on the news summary; Executing the second step to obtain a second reasoning result based on the dynamic summary of the feature data; Based on the first reasoning result and the second reasoning result, a first feature data prediction result of the target object in a short-term future period and explanation information for the first feature data prediction result are obtained.

4. The method according to claim 2, in, The multiple information summaries include a macro information summary and a fundamental summary; the chain logic reasoning step includes a third step and a fourth step; based on the multiple information summaries, according to the chain logic reasoning step, the feature data prediction result of the target object and the explanation information for the feature data prediction result are obtained, including: Executing the third step to obtain a third reasoning result based on the macro information summary; Executing the fourth step to obtain a fourth reasoning result based on the fundamentals summary; Based on the third reasoning result and the fourth reasoning result, a second feature data prediction result of the target object in a long-term future period and explanation information for the second feature data prediction result are obtained.

5. The method according to claim 2, in, The multiple information summaries include news summaries, feature data dynamic summaries, macro information summaries, and fundamentals summaries; the chain logic reasoning steps include a first step, a second step, a third step, and a fourth step; based on the multiple information summaries, according to the chain logic reasoning steps, the feature data prediction results of the target object and the explanation information for the feature data prediction results are obtained, including: Execute the first step to obtain a first reasoning result based on the news summary; Executing the second step to obtain a second reasoning result based on the dynamic summary of the feature data; Executing the third step to obtain a third reasoning result based on the macro information summary; Executing the fourth step to obtain a fourth reasoning result based on the fundamentals summary; Based on the first reasoning result, the second reasoning result, the third reasoning result and the fourth reasoning result, a first feature data prediction result of the target object in a short-term future period, a second feature data prediction result of the target object in a long-term future period, and explanatory information for the first feature data prediction result and the second feature data prediction result are obtained.

6. The method according to claim 1 or 2, in, The multiple information summaries include news summaries, and the news summaries include daily news summaries; the acquiring of multiple information summaries related to the target object includes: Obtain N groups of daily public opinion news related to the target object; wherein the N groups of daily public opinion news correspond one-to-one to N independent dates in the current period; N ≥ 2 and is an integer; Each group of daily public opinion news in the N groups of daily public opinion news is used as the public opinion news to be processed, and the public opinion news to be processed is processed by using a first news summarizer to obtain a daily news summary corresponding to the public opinion news to be processed.

7. The method according to claim 6, in, The news summary also includes a news summary for the current period; the step of obtaining a plurality of information summaries related to the target object also includes: Based on the N daily news summaries corresponding to the N groups of daily public opinion news, a daily news summary set is obtained; Get the historical period news summary corresponding to the historical period; The second news summarizer is used to process the daily news summary set based on the historical time period news summary to obtain a current time period news summary corresponding to the current time period.

8. The method according to claim 1 or 2, in, The plurality of information summaries include a dynamic summary of characteristic data; The obtaining of multiple information summaries related to the target object includes: Acquire first market data of the target object; Acquire second market data of a first target number of reference objects similar to the target object; The feature data dynamic summarizer is used to obtain the feature data dynamic summary based on the first market data of the target object, the second market data of the first target number of reference objects, and a specific reference index.

9. The method according to claim 1 or 2, in, The multiple information summaries include macro information summaries; the acquiring multiple information summaries related to the target object includes: Using the large language model, the second target number of macro research reports are processed to obtain a second target number of macro research summaries corresponding to the second target number of macro research reports; Using the large language model, the macro-indicator data is processed to obtain a macro-indicator summary; The macro information summary is obtained by utilizing the large language model based on the second target number of macro research summaries and the macro indicator summaries.

10. The method according to claim 1 or 2, in, The multiple information summaries include a basic surface summary; the obtaining of multiple information summaries related to the target object includes: Using a large language model, the management level analysis results of the target entity are processed to obtain a management level fundamentals summary; wherein the target entity is the economic organization that issues the target object; Using the large language model, processing analyst research reports related to the target entity to obtain an analyst fundamental summary; The fundamental summary is obtained by utilizing the large language model based on the management fundamental summary, the analyst fundamental summary, and the specific data of the target entity.

11. The method according to claim 10, in, The analyst research reports related to the target entity are processed to obtain the analyst fundamental summary, including: Obtaining a first analyst research report on the target entity; Obtaining second analyst research reports on a third target number of reference entities that are similar to the target entity; The first analyst research report of the target entity and the second analyst research report of the third target reference entity are used as analyst research reports related to the target entity, and the analyst research reports related to the target entity are processed to obtain the analyst fundamental summary.

12. A large model training method, include: Acquire multiple information summary samples related to the target object sample; wherein the multiple information summary samples have different information categories, so as to characterize the performance and / or impact degree of the target object sample in different dimensions; Using the initial large model, based on the multiple information summary samples, obtain a feature data prediction result sample for the target object sample and an explanation information sample for the feature data prediction result sample; Based on the feature data prediction result samples and the explanation information samples, the initial large model is trained to obtain a target large model.

13. The method according to claim 12, in, The step of training the initial large model based on the feature data prediction result sample and the explanation information sample to obtain a target large model includes: Acquire real feature data for the target object sample; Generate a reflection basis including the plurality of information summary samples, the feature data prediction result sample, the explanation information sample and the real feature data; Using the self-reflection model, based on the reflection basis, obtain the current reflection result for the feature data prediction result sample and the explanation information sample; Based on the current reflection result, the initial large model is trained to obtain a target large model.

14. The method according to claim 13, in, The generation includes the plurality of information summary samples, the feature data prediction result samples, the explanation information samples and the reflection basis of the real feature data, including: Using the target scoring model, obtaining scoring results for the feature data prediction result sample and the explanation information sample, as well as scoring explanations for the scoring results; Generate a reflection basis including the multiple information summary samples, the feature data prediction result samples, the explanation information samples, the scoring results, the scoring explanations and the real feature data.

15. The method according to claim 14, further comprising: include: Acquire a first sample pair; wherein the first sample pair includes a first feature data prediction sample and a first explanation information sample, and first real feature data corresponding to the first feature data prediction sample; the similarity between the first feature data prediction sample and the first real feature data is greater than or equal to a first preset similarity threshold; Acquire a second sample pair; wherein the second sample pair includes a second feature data prediction sample and a second interpretation information sample, and second real feature data corresponding to the second feature data prediction sample; the similarity between the second feature data prediction sample and the second real feature data is less than a second similarity threshold, and the second similarity threshold is less than the first preset similarity threshold; Based on the first sample pair and the second sample pair, an initial scoring module is trained to obtain the target scoring model.

16. The method according to claim 13, in, The step of training the initial large model based on the current reflection result to obtain the target large model includes: Obtain historical reflection results; Based on the current reflection result and the historical reflection result, long-term reflection data is obtained; Based on the long-term reflection data, the initial large model is trained to obtain the target large model.

17. A data prediction device, include: A summary information acquisition unit, used to acquire a plurality of information summaries related to the target object; wherein the plurality of information summaries have different information categories, so as to characterize the performance and / or the degree of influence of the target object in different dimensions; The feature data prediction unit is used to obtain the feature data prediction result of the target object and the explanation information for the feature data prediction result based on the multiple information summaries by using the target large model.

18. The device according to claim 17, in, The feature data prediction unit is used for: Get answer guidance information; Based on the answer guidance information, chain logic reasoning steps are obtained; Based on the multiple information summaries and in accordance with the chain logic reasoning steps, the feature data prediction results of the target object and the explanation information for the feature data prediction results are obtained.

19. The device according to claim 18, in, The multiple information summaries include news summaries and feature data dynamic summaries; the chain logic reasoning steps include a first step and a second step; the feature data prediction unit is used to: Execute the first step to obtain a first reasoning result based on the news summary; Executing the second step to obtain a second reasoning result based on the dynamic summary of the feature data; Based on the first reasoning result and the second reasoning result, a first feature data prediction result of the target object in a short-term future period and explanation information for the first feature data prediction result are obtained.

20. The device according to claim 18, in, The multiple information summaries include a macro information summary and a fundamental information summary; the chain logic reasoning step includes a third step and a fourth step; the feature data prediction unit is used to: Executing the third step to obtain a third reasoning result based on the macro information summary; Executing the fourth step to obtain a fourth reasoning result based on the fundamentals summary; Based on the third reasoning result and the fourth reasoning result, a second feature data prediction result of the target object in a long-term future period and explanation information for the second feature data prediction result are obtained.

21. The device according to claim 18, in, The multiple information summaries include news summaries, feature data dynamic summaries, macro information summaries and fundamentals summaries; the chain logic reasoning steps include a first step, a second step, a third step and a fourth step; the feature data prediction unit is used to: Execute the first step to obtain a first reasoning result based on the news summary; Executing the second step to obtain a second reasoning result based on the dynamic summary of the feature data; Executing the third step to obtain a third reasoning result based on the macro information summary; Executing the fourth step to obtain a fourth reasoning result based on the fundamentals summary; Based on the first reasoning result, the second reasoning result, the third reasoning result and the fourth reasoning result, a first feature data prediction result of the target object in a short-term future period, a second feature data prediction result of the target object in a long-term future period, and explanatory information for the first feature data prediction result and the second feature data prediction result are obtained.

22. The device according to claim 17 or 18, in, The plurality of information summaries include news summaries, and the news summaries include daily news summaries; and the summary information acquisition unit is used to: Obtain N groups of daily public opinion news related to the target object; wherein the N groups of daily public opinion news correspond one-to-one to N independent dates in the current period; N ≥ 2 and is an integer; Each group of daily public opinion news in the N groups of daily public opinion news is used as the public opinion news to be processed, and the public opinion news to be processed is processed by using a first news summarizer to obtain a daily news summary corresponding to the public opinion news to be processed.

23. The device according to claim 22, in, The news summary also includes a news summary for the current period; the summary information acquisition unit is also used for: Based on the N daily news summaries corresponding to the N groups of daily public opinion news, a daily news summary set is obtained; Get the historical period news summary corresponding to the historical period; The second news summarizer is used to process the daily news summary set based on the historical time period news summary to obtain a current time period news summary corresponding to the current time period.

24. The device according to claim 17 or 18, in, The plurality of information summaries include a dynamic summary of feature data; the summary information acquisition unit is used to: Acquire first market data of the target object; Acquire second market data of a first target number of reference objects similar to the target object; The feature data dynamic summarizer is used to obtain the feature data dynamic summary based on the first market data of the target object, the second market data of the first target number of reference objects, and a specific reference index.

25. The device according to claim 17 or 18, in, The multiple information summaries include macro information summaries; the summary information acquisition unit is used to: Using the large language model, the second target number of macro research reports are processed to obtain a second target number of macro research summaries corresponding to the second target number of macro research reports; Using the large language model, the macro-indicator data is processed to obtain a macro-indicator summary; The macro information summary is obtained by utilizing the large language model based on the second target number of macro research summaries and the macro indicator summaries.

26. The device according to claim 17 or 18, in, The plurality of information summaries include a basic surface summary; the summary information acquisition unit is used for: Using a large language model, the management level analysis results of the target entity are processed to obtain a management level fundamentals summary; wherein the target entity is the economic organization that issues the target object; Using the large language model, processing analyst research reports related to the target entity to obtain an analyst fundamental summary; The fundamental summary is obtained by utilizing the large language model based on the management fundamental summary, the analyst fundamental summary, and the specific data of the target entity.

27. The device according to claim 26, in, The summary information acquisition unit is used for: Obtaining a first analyst research report on the target entity; Obtaining second analyst research reports on a third target number of reference entities that are similar to the target entity; The first analyst research report of the target entity and the second analyst research report of the third target reference entity are used as analyst research reports related to the target entity, and the analyst research reports related to the target entity are processed to obtain the analyst fundamental summary.

28. A large model training device, include: A first sample acquisition unit is used to acquire a plurality of information summary samples related to the target object sample; wherein the plurality of information summary samples have different information categories, so as to characterize the performance and / or the degree of influence of the target object sample in different dimensions; A second sample acquisition unit is used to obtain a feature data prediction result sample for the target object sample and an explanation information sample for the feature data prediction result sample based on the multiple information summary samples using the initial large model; The large model training unit is used to train the initial large model based on the feature data prediction result sample and the explanation information sample to obtain a target large model.

29. The device according to claim 28, in, The large model training unit is used for: Acquire real feature data for the target object sample; Generate a reflection basis including the plurality of information summary samples, the feature data prediction result sample, the explanation information sample and the real feature data; Using the self-reflection model, based on the reflection basis, obtain the current reflection result for the feature data prediction result sample and the explanation information sample; Based on the current reflection result, the initial large model is trained to obtain a target large model.

30. The device according to claim 29, in, The large model training unit is used for: Using the target scoring model, obtaining scoring results for the feature data prediction result sample and the explanation information sample, as well as scoring explanations for the scoring results; Generate a reflection basis including the multiple information summary samples, the feature data prediction result samples, the explanation information samples, the scoring results, the scoring explanations and the real feature data.

31. The apparatus according to claim 30, further comprising a scoring large model training unit, for: Obtaining a first sample pair; in, The first sample pair includes a first feature data prediction sample and a first explanation information sample, and first real feature data corresponding to the first feature data prediction sample; the similarity between the first feature data prediction sample and the first real feature data is greater than or equal to a first preset similarity threshold; Acquire a second sample pair; wherein the second sample pair includes a second feature data prediction sample and a second interpretation information sample, and second real feature data corresponding to the second feature data prediction sample; the similarity between the second feature data prediction sample and the second real feature data is less than a second similarity threshold, and the second similarity threshold is less than the first preset similarity threshold; Based on the first sample pair and the second sample pair, an initial scoring module is trained to obtain the target scoring model.

32. The device according to claim 29, in, The large model training unit is used for: Obtain historical reflection results; Based on the current reflection result and the historical reflection result, long-term reflection data is obtained; Based on the long-term reflection data, the initial large model is trained to obtain the target large model.

33. An electronic device, include: at least one processor; a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the method according to any one of claims 1 to 16.

34. A non-transitory computer-readable storage medium storing computer instructions, in, The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 16.

35. A computer program product, comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 16.

Images