CN-122021862-A - Financial intelligent wind control large model system and reasoning enhancement method

CN122021862ACN 122021862 ACN122021862 ACN 122021862ACN-122021862-A

Abstract

The application provides a financial intelligent wind control big model system and an inference enhancement method, which relate to the technical field of natural language processing, and are used for screening historical risk event text data related to financial risk events based on word frequency correlation of risk features and semantic association relation among texts, further constructing a risk event set after retrieval enhancement, guiding a financial intelligent wind control big model to generate a multi-view risk inference chain through a thinking chain, further constructing weight matrixes reflecting importance of different risk elements by combining the multi-view risk inference chain and occurrence frequencies of the risk elements in the financial risk events, adjusting inference weights of the different risk elements through the weight matrixes to obtain a dynamic inference strategy adapting to the financial risk events, and carrying out enhancement inference on the financial risk events according to the dynamic inference strategy to obtain a risk assessment result meeting financial intelligent wind control requirements. By adopting the scheme of the application, the enhanced reasoning of the financial risk can be realized based on the self-adaptive adjustment of the risk factor reasoning weight.

Inventors

LIU SHUO
WANG HONGBO
SHI WANGJUN
DONG BAICHEN
YUAN ZHEN
LIU GUANG
CHEN SHAOZHEN

Assignees

国网汇通金财（北京）信息科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251029

Claims (10)

1. An reasoning enhancement method applied to a financial intelligent wind control large model system is characterized by comprising the following steps: identifying financial risk events which are preliminarily determined to be high risk from the financial text data; screening historical risk event text data related to the financial risk event based on word frequency correlation of risk features and semantic association relation between texts, and further constructing a risk event set after retrieval enhancement; According to the risk event set, a financial intelligent wind control large model is guided through a thinking chain to generate a multi-view risk reasoning chain, and then a weight matrix reflecting the importance of different risk elements is constructed by combining the multi-view risk reasoning chain and the occurrence frequency of the risk elements in the financial risk event; the reasoning weights of different risk factors are adjusted through the weight matrix, so that a dynamic reasoning strategy adapting to the financial risk event is obtained; and carrying out enhanced reasoning on the financial risk event according to the dynamic reasoning strategy and combining the causal logic relationship between the risk factors and the reasoning conclusion to obtain a risk assessment result meeting the financial intelligent wind control requirement.
2. The method of claim 1, wherein identifying a financial risk event from the financial text data that is initially determined to be high risk comprises: extracting keywords containing risk features from the financial text data, wherein the risk features comprise transaction abnormal fluctuation, overdue repayment, illegal operation and negative public opinion labels; matching the extracted risk feature keywords with a preset high risk feature library, and counting the number of successfully matched risk features; And when the number of the risk features successfully matched exceeds a preset risk threshold, judging the corresponding financial event as a preliminary high-risk financial risk event.
3. The method of claim 1, wherein screening historical risk event text data related to the financial risk event based on word frequency correlation of risk features and semantic association between texts, and further constructing a search-enhanced risk event set specifically comprises: extracting core risk characteristics of financial risk events which are preliminarily judged to be high risk, wherein the core risk characteristics comprise high-frequency risk keywords and risk type labels; calculating word frequency correlation scores of the core risk features and the risk features in the historical risk event text data, wherein the word frequency correlation scores are calculated based on the occurrence frequency of feature words and the frequency of inverse documents; generating semantic vectors of the financial risk event and the historical risk event text data through a pre-trained text semantic model, and calculating semantic association degrees between the semantic vectors through similarity; The word frequency relevance score and the semantic relevance are fused to obtain a comprehensive relevance score, and historical risk event text data with the comprehensive relevance score higher than a set threshold value is screened out; and merging the screened historical risk event text data with the preliminarily determined high-risk financial risk event to construct a risk event set after retrieval enhancement.
4. The method of claim 1, wherein the step of generating a multi-view risk inference chain by guiding a financial intelligent wind-driven big model through a thought chain according to the risk event set comprises the following steps: designing a thinking chain prompt template comprising a multi-dimensional reasoning angle; the step requirements of reasoning are clearly defined in the thinking chain prompt template, and a complete reasoning process from risk factor identification to risk conclusion is guided to be output by the financial intelligent wind control big model; Inputting the risk event set into a financial intelligent wind control big model, and triggering the financial intelligent wind control big model to conduct step-by-step reasoning through the thinking chain prompt template; And obtaining a plurality of groups of reasoning results generated by the financial intelligent wind control big model, wherein each group of reasoning results corresponds to one reasoning angle, and a multi-view risk reasoning chain comprising intermediate reasoning nodes and logic association is formed.
5. The method of claim 4, wherein the inference angle comprises a risk formation cause, a risk conduction path, a risk impact range.
6. The method of claim 1, wherein constructing a weight matrix reflecting importance of different risk elements in combination with a multi-view risk inference chain and occurrence frequencies of risk elements in the financial risk event specifically comprises: extracting all related risk factors from a multi-view risk reasoning chain; counting the occurrence times of each risk element in all inference chains, and further obtaining the occurrence frequency of each risk element; analyzing the position weight of each risk element in the inference chain; Determining an importance score of each risk element by combining the occurrence frequency and the position weight of the risk element; and constructing a weight matrix reflecting the importance of different risk elements according to the importance scores of all the risk elements.
7. The method of claim 1, wherein adjusting the inference weights of different risk factors by the weight matrix to obtain a dynamic inference policy that adapts to the financial risk event specifically comprises: Determining initial inference weights of risk elements in the financial risk event; adjusting the initial reasoning weight of each risk element based on the importance score of each risk element in the weight matrix to obtain the revised reasoning weight of each risk element; And determining the priority of each risk element in the reasoning process through the revised reasoning weights of all the risk elements to form a dynamic reasoning strategy adapting to the financial risk event.
8. The system comprises an reasoning enhancement unit, and is characterized in that the reasoning enhancement unit comprises: The identification module is used for identifying financial risk events which are preliminarily judged to be high risk from the financial text data; The processing module is used for screening historical risk event text data related to the financial risk event based on word frequency correlation of risk features and semantic association relation among texts, and further constructing a risk event set after retrieval enhancement; the processing module is further used for guiding a financial intelligent wind control large model to generate a multi-view risk reasoning chain through a thinking chain according to the risk event set, and further constructing a weight matrix reflecting the importance of different risk elements by combining the multi-view risk reasoning chain and the occurrence frequency of the risk elements in the financial risk event; The processing module is further used for adjusting the reasoning weights of different risk factors through the weight matrix to obtain a dynamic reasoning strategy adapting to the financial risk event; And the execution module is used for carrying out enhanced reasoning on the financial risk event according to the dynamic reasoning strategy and combining the causal logic relationship between the risk element and the reasoning conclusion to obtain a risk assessment result which meets the financial intelligent wind control requirement.
9. A computer device comprising a memory storing code and a processor, wherein the processor is configured to obtain the code and perform the inference enhancement method of any of claims 1 to 7.
10. A computer readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the inference enhancement method of any one of claims 1 to 7.

Description

Financial intelligent wind control large model system and reasoning enhancement method Technical Field The application relates to the technical field of natural language processing, in particular to a financial intelligent wind control large model system and an inference enhancement method. Background The natural language processing (Natural Language Processing, NLP) technology has important significance in the financial wind control field, can realize automatic identification and classification of potential risk events by carrying out semantic understanding and structural analysis on massive financial text data, and the NLP can not only extract risk elements from unstructured data such as notices, news, transaction records and the like, but also assist in modeling causal relations among financial events, so that the coverage range and response speed of risk monitoring are improved, and a financial institution can realize deep understanding of complex events by means of semantic representation and text vectorization technology, provide data support for risk assessment, early warning and decision making, and improve the overall wind control efficiency and accuracy. The existing financial risk reasoning method mainly depends on static rules or weight setting of single characteristics, the importance of different risk factors in specific events is difficult to dynamically reflect, the contribution of part of key risk factors is underestimated in complex financial events, edge factors or noise information can be excessively amplified, the false alarm rate of risk reasoning is higher, the accuracy and reliability of decision making are affected, in addition, the existing method lacks a uniform weighting mechanism in multi-event and multi-dimensional information fusion, traceable and logically clear reasoning chains are difficult to form, and risk assessment results are difficult to interpret and verify. Therefore, how to implement enhanced reasoning of financial risk based on adaptive adjustment of risk factor reasoning weights becomes a difficult problem for the industry. Disclosure of Invention The application provides a financial intelligent wind control large model system and an inference enhancement method, which can realize the enhancement inference of financial risk based on the self-adaptive adjustment of the inference weight of risk factors. In a first aspect, the present application provides an inference enhancement method applied to a financial intelligent wind-controlled large model system, the method comprising the steps of: identifying financial risk events which are preliminarily determined to be high risk from the financial text data; screening historical risk event text data related to the financial risk event based on word frequency correlation of risk features and semantic association relation between texts, and further constructing a risk event set after retrieval enhancement; According to the risk event set, a financial intelligent wind control large model is guided through a thinking chain to generate a multi-view risk reasoning chain, and then a weight matrix reflecting the importance of different risk elements is constructed by combining the multi-view risk reasoning chain and the occurrence frequency of the risk elements in the financial risk event; the reasoning weights of different risk factors are adjusted through the weight matrix, so that a dynamic reasoning strategy adapting to the financial risk event is obtained; and carrying out enhanced reasoning on the financial risk event according to the dynamic reasoning strategy and combining the causal logic relationship between the risk factors and the reasoning conclusion to obtain a risk assessment result meeting the financial intelligent wind control requirement. Preferably, identifying the financial risk event preliminarily determined to be high risk from the financial text data specifically includes: extracting keywords containing risk features from the financial text data, wherein the risk features comprise transaction abnormal fluctuation, overdue repayment, illegal operation and negative public opinion labels; matching the extracted risk feature keywords with a preset high risk feature library, and counting the number of successfully matched risk features; And when the number of the risk features successfully matched exceeds a preset risk threshold, judging the corresponding financial event as a preliminary high-risk financial risk event. Preferably, the step of screening historical risk event text data related to the financial risk event based on word frequency correlation of risk features and semantic association relation between texts, and the step of constructing a risk event set after retrieval enhancement specifically comprises the following steps: extracting core risk characteristics of financial risk events which are preliminarily judged to be high risk, wherein the core risk characteristics comprise high-frequency risk k