CN-122023011-A - Financial business processing method, apparatus, computer device and program product

CN122023011ACN 122023011 ACN122023011 ACN 122023011ACN-122023011-A

Abstract

The present application relates to the field of financial service processing technologies, and in particular, to a financial service processing method, apparatus, computer device, and program product. The method comprises the steps of encoding client queries into content vectors and condition vectors, determining an applicable rule area through weighted fusion by calculating the distance from the condition vectors to preset navigation anchor points, wherein the navigation anchor points are clustering centers after clustering rules with similar applicable conditions in a knowledge base, calculating information entropy of the condition vectors, determining a knowledge boundary range according to the information entropy, expanding the knowledge boundary range to contain a plurality of candidate rules when the information entropy is high, shrinking the knowledge boundary range to lock accurate rules when the information entropy is low, and adjusting attention weights of decoders according to the applicable rule area and the knowledge boundary range to generate answers in the knowledge boundary range. The application can improve the communication experience of the clients when processing financial business.

Inventors

XU KEFEI
Yu xiaoqiao
CHEN SHUHUI
FENG SIYUAN
ZHOU FUKEN
FU YINHUI

Assignees

广东东软学院
广东国相云科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251209

Claims (10)

1. The financial business processing method is characterized by comprising the following steps: Acquiring a client query; Encoding the client query into a content vector and a condition vector, the content vector characterizing semantic content of a question, the condition vector characterizing applicable conditions upon which the question is answered; The applicable rule area is determined by calculating the distance from the condition vector to a preset navigation anchor point, wherein the navigation anchor point is a clustering center after the rule clustering with similar applicable conditions in the knowledge base; Calculating the information entropy of the condition vector; Determining a knowledge boundary range according to the information entropy, expanding the knowledge boundary range to contain a plurality of candidate rules when the information entropy is high, and narrowing the knowledge boundary range to lock accurate rules when the information entropy is low; and adjusting the attention weight of the decoder according to the applicable rule area and the knowledge boundary range, and generating an answer in the knowledge boundary range.
2. The financial transaction processing method according to claim 1, further comprising the steps of: maintaining a hypothesis stack in the dialogue process, and recording implicit applicable conditions and corresponding historical answer identifications, on which answers are generated, wherein the implicit applicable conditions are extracted from applicable conditions of rules after matching answer texts with business contents of rules in a knowledge base; acquiring a condition vector corresponding to a new round of client query, and calculating the inconsistency degree of the condition vector and corresponding dimensions of each implicit applicable condition in the hypothetical stack; Triggering a backtracking mechanism when the inconsistency exceeds a preset opposite threshold, extracting historical answers and subsequent answers thereof depending on the negated implicit applicable conditions from a hypothesis stack, and marking the historical answers and subsequent answers as to-be-reevaluated; Updating the condition vector according to a new round of customer inquiry, recalculating information entropy and boundary radius for the affected condition dimension, and redefining an applicable rule area, and keeping the original boundary range for the unaffected dimension; And re-evaluating the marked historical answers in the updated applicable rule area and knowledge boundary range, and generating corrected answers for the historical answers with the matching degree lower than a preset confidence threshold.
3. The financial transaction processing method according to claim 1, wherein said encoding the customer query into a content vector and a condition vector comprises: Customer queries are processed through a dual encoder architecture: Performing self-attention processing on the client query by a content encoder to extract problem semantics to generate the content vector; performing cross-attention processing on the client query by a condition encoder to identify a condition constraint to generate the condition vector; the double encoder is trained through comparison learning, positive sample pairs are different problem expressions under the same applicable conditions, negative sample pairs are semantic similarity problems under different applicable conditions, and the content vector and the condition vector form orthogonal representation dimensions in a feature space through comparison objective function optimization.
4. The financial transaction processing method according to claim 1, wherein the determining the applicable rule area by calculating a distance from the condition vector to a preset navigation anchor point through weighted fusion includes: In the pre-training stage, clustering rules in the knowledge base according to the similarity of the applicable conditions, and marking the clustering center of each cluster as a navigation anchor point; calculating Euclidean distance from the condition vector to each navigation anchor point; calculating the fusion weight of each navigation anchor point by adopting a Gaussian kernel function, wherein the calculation formula of the fusion weight is as follows: Wherein the method comprises the steps of Representing the euclidean distance of the condition vector to the ith navigation anchor point, Is a bandwidth parameter; And carrying out weighted fusion according to the fusion weight of each navigation anchor point, and determining the positioning of the condition vector in the knowledge base, wherein the knowledge base area corresponding to the positioning is the applicable rule area.
5. The financial transaction processing method according to claim 1, wherein determining a knowledge boundary range from the information entropy includes: Respectively calculating the information entropy of the condition vector in each condition dimension: Wherein, the The probability of the ith condition dimension on the jth possible value; calculating the boundary radius of each dimension according to the information entropy of each condition dimension: So that the lower information entropy corresponding to the defined condition dimension leads to smaller boundary radius, and the higher information entropy corresponding to the undefined condition dimension leads to larger boundary radius; the boundary radii of the dimensions are combined to form a knowledge boundary range in the high-dimensional space.
6. The financial transaction processing method according to claim 1, wherein said adjusting the attention weight of the decoder according to the applicable rule area and the knowledge boundary range comprises: determining sampling weights of the content retrieved from the rules corresponding to the navigation anchor points according to the fusion weights of the navigation anchor points in the applicable rule area; calculating the whole width of the knowledge boundary range according to the boundary radius of each current condition dimension; Determining an answer generation strategy according to the whole width, adopting a multi-candidate display strategy when the whole width is larger than a preset width threshold value, and adopting an accurate answer strategy when the whole width is smaller than the preset width threshold value; Performing dimension filtering on the rule to be searched according to the boundary radius of each condition dimension, and only searching the rule matched with the dimension condition for the defined dimension with the boundary radius smaller than the preset radius threshold value, and reserving a plurality of values of the rule for the undefined dimension with the boundary radius larger than the preset radius threshold value; Retrieving content from the filtered rules according to the sampling weights and the answer generation strategy, and adjusting the attention weight of the retrieved content when it is input to the decoder.
7. The financial transaction processing method according to claim 2, wherein the implicit applicable conditions are extracted from applicable conditions of the rules by matching answer text with the transaction contents of each rule in the knowledge base, comprising: encoding the generated answer text into a vector representation; Calculating the similarity between the answer vector and the business content vector of each rule in the knowledge base, and identifying the rule with the highest similarity; extracting the applicable conditions of the rule labels, and comparing the applicable conditions with the explicit dimensions in the current condition vector; and recording the condition which does not appear in the explicit dimension in the applicable conditions as an implicit applicable condition to a hypothesis stack and associating the corresponding historical answer identification.
8. A financial transaction processing device, characterized in that it comprises at least one module for performing the financial transaction processing method according to any one of claims 1 to 7.
9. Computer device, characterized in that it comprises a processor for executing a computer program stored in a memory for implementing the financial transaction processing method according to any of claims 1-7.
10. Computer program product containing instructions which, when executed by a computer device, cause the computer device to perform the financial transaction method according to any of claims 1 to 7.

Description

Financial business processing method, apparatus, computer device and program product Technical Field The present application relates to the field of financial service processing technologies, and in particular, to a financial service processing method, apparatus, computer device, and program product. Background The application of large language models in professional fields such as financial consultation requires a well-defined knowledge boundary. In the field of financial business consultation, differentiated terms are usually formulated for different customer groups, and a model must determine whether a specific answer is applicable to a current customer. The model learns various business rules in the pre-training stage, but can give out answers which are correct in facts and are not applicable when actually answering. For example, when a customer asks for a loan interest rate, the model outputs the corresponding interest rate of the standard customer, who actually belongs to the VIP group, to which the VIP applies. The knowledge base based matching scheme can explicitly label the applicable conditions of each rule, but its fixed boundary requires that all necessary conditions are known to give an answer. In the early stages of a conversation, a customer generally only can say that I want to deal with a credit card, but a matching scheme based on a knowledge base needs a plurality of conditions such as credit investigation conditions, income level, liability conditions and the like to be matched with specific rules. Only answering or requiring customers to answer a series of questions first, interactive mode machinery under fixed knowledge boundaries. For example, after a customer provides a monthly revenue of 8000 yuan, the credit status is still unknown, although the revenue level is known. The fixed boundary cannot utilize this partial information to shrink the answer range to a rule set applicable to medium income groups, yet still requires the customer to supplement all remaining conditions. The prior art still has some problems, which lead to poor customer communication experience in large model-based financial business processes. Disclosure of Invention In order to solve the above technical problems or at least partially solve the above technical problems, the present application provides a financial service processing method, apparatus, computer device and program product, which can improve the communication experience of clients when processing financial services. In a first aspect, the present application provides a financial transaction processing method, including the steps of: Acquiring a client query; Encoding the client query into a content vector and a condition vector, the content vector characterizing semantic content of a question, the condition vector characterizing applicable conditions upon which the question is answered; The applicable rule area is determined by calculating the distance from the condition vector to a preset navigation anchor point, wherein the navigation anchor point is a clustering center after the rule clustering with similar applicable conditions in the knowledge base; Calculating the information entropy of the condition vector; Determining a knowledge boundary range according to the information entropy, expanding the knowledge boundary range to contain a plurality of candidate rules when the information entropy is high, and narrowing the knowledge boundary range to lock accurate rules when the information entropy is low; and adjusting the attention weight of the decoder according to the applicable rule area and the knowledge boundary range, and generating an answer in the knowledge boundary range. Optionally, the financial service processing method further includes the following steps: maintaining a hypothesis stack in the dialogue process, and recording implicit applicable conditions and corresponding historical answer identifications, on which answers are generated, wherein the implicit applicable conditions are extracted from applicable conditions of rules after matching answer texts with business contents of rules in a knowledge base; acquiring a condition vector corresponding to a new round of client query, and calculating the inconsistency degree of the condition vector and corresponding dimensions of each implicit applicable condition in the hypothetical stack; Triggering a backtracking mechanism when the inconsistency exceeds a preset opposite threshold, extracting historical answers and subsequent answers thereof depending on the negated implicit applicable conditions from a hypothesis stack, and marking the historical answers and subsequent answers as to-be-reevaluated; Updating the condition vector according to a new round of customer inquiry, recalculating information entropy and boundary radius for the affected condition dimension, and redefining an applicable rule area, and keeping the original boundary range for the unaffected dimension; And re-ev