Method and system for matching financial knowledge question and answer with vector representation

Abstract

The application provides a financial knowledge question and answer matching method and system combined with vector representation, relates to the technical field of artificial intelligence and natural language processing, and comprises the following steps: one, constructing a financial knowledge vector library: decomposing unstructured financial documents into a plurality of knowledge segments through a semantic segmentation algorithm; adopting a pre-trained language model to perform vectorization coding on the knowledge segments to generate multi-dimensional semantic vectors; based on vector similarity clustering, establishing a hierarchical index structure to form the financial knowledge vector library; two, vectorization processing of user questions: receiving a financial problem text input by a user, and generating a corresponding question vector by using the pre-trained language model.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing, and in particular to a financial knowledge question-answering matching method and system that combines vector representation. Background Technology

[0002] Traditional financial knowledge management systems mainly rely on keyword matching or rule engines to implement question-and-answer functions, which has the following drawbacks: 1. Low retrieval efficiency: It is difficult to quickly locate key information in a large number of unstructured financial documents (such as policies, regulations, and operation manuals), and users need to spend a lot of time filtering them; 2. Insufficient semantic understanding: User questions and knowledge base content fail to match due to differences in expression (such as "accrual of depreciation" and "amortization of fixed assets"); 3. Weak dynamic optimization capability: The knowledge base cannot be updated in real time based on user feedback, resulting in limited question and answer quality and high knowledge base maintenance costs; 4. Difficulty in processing multimodal data: Financial scenarios involve various data formats such as text, tables, and scanned vouchers, which are difficult for traditional systems to integrate effectively. Summary of the Invention

[0003] This invention relates to a financial knowledge question-and-answer matching method and system that combines vector representation, which achieves integrated efficiency retrieval, accurate question-and-answer, and automated auditing.

[0004] This invention provides a financial knowledge question-answering matching method and system that combines vector representation, including the following steps: I. Constructing a Financial Knowledge Vector Library: Unstructured financial documents are decomposed into several knowledge fragments using a semantic segmentation algorithm; a pre-trained language model is used to vectorize and encode the knowledge fragments to generate multi-dimensional semantic vectors; based on vector similarity clustering, a hierarchical index structure is established to form a financial knowledge vector library. 2. Vectorization of user questions: Receive financial question text input by the user and generate corresponding question vectors using the pre-trained language model; III. Dynamic Matching and Answer Generation: Calculate the similarity between the question vector and each vector in the knowledge vector base, and select the Top-N candidate answers; combine the contextual relevance and historical user feedback weights to rank the candidate answers and output the optimal result; IV. Feedback-driven knowledge base optimization: Collect user satisfaction ratings and click behavior data for answers; adjust the vector weights of knowledge fragments based on feedback data, and dynamically update the knowledge vector base.

[0005] The financial knowledge question-answering matching method based on vector representation includes the following modules: I. Knowledge Processing Module: Used for document segmentation, vector encoding, and index construction; II. Question Answering Engine Module: Includes a question vectorization unit, a similarity matching unit, and a sorting output unit; III. Feedback Optimization Module: Collects user behavior data and optimizes knowledge vector weights through reinforcement learning algorithms; IV. Automated Audit Module: Integrates OCR technology, rule engine, and machine learning model to perform structured parsing, field extraction, and abnormal behavior detection on uploaded financial vouchers.

[0006] Optionally, the semantic segmentation algorithm includes: Named Entity Recognition (NER) based on a financial terminology dictionary to segment the document into policy clauses, operational steps, and question-answer pairs; and constructing a semantic association graph between knowledge fragments using a graph neural network (GNN).

[0007] Optionally, the automated audit module further includes: a voucher template configuration unit and a risk warning unit. The voucher template configuration unit allows administrators to customize audit rules and field extraction templates. The risk warning unit trains an anomaly detection model based on historical data to identify duplicate payments and split payments.

[0008] Optionally, the contextual relevance is calculated in the following way: 1. Extract at least one type of contextual information from user queries, including time, amount, and policy version number; 2. Combine the contextual information with the metadata of the candidate answer fragments to generate a relevance score.

[0009] Optionally, the feedback data further includes the user's modified answer text, which is then used as a new knowledge fragment, vectorized, and added to the financial knowledge vector library.

[0010] Optionally, the knowledge processing module further includes a multimodal data integration unit, which supports joint vectorization encoding of mixed documents containing text, tables, and images to generate a unified semantic representation.

[0011] Optionally, the feedback optimization module further includes a popularity analysis unit and an adaptive learning unit. The popularity analysis unit counts the frequency of user questions and the click rate of answers, generates a list of popular questions and displays them in priority. The adaptive learning unit updates the parameters of the pre-trained language model through online learning based on user feedback data.

[0012] Optionally, it also includes: a permission management module: based on a three-level model of role-permission-user, it realizes permission allocation and operation log recording for users in different positions.

[0013] Optionally, the access management module supports dynamically adjusting the scope of accessible knowledge bases based on the user's organization and position, and performing multi-factor authentication and operation logging for sensitive operations.

[0014] This invention provides a financial knowledge question-answering matching method and system that combines vector representation, which has the following beneficial effects: 1. The financial knowledge question-answering matching method and system proposed in this invention, which combines vector representation, demonstrates significant technical advantages and practical value in financial knowledge management, intelligent question answering, and automated auditing scenarios. By constructing a dynamically optimized financial knowledge vector library and combining multimodal data processing and feedback-driven mechanisms, the system can achieve millisecond-level semantic retrieval responses, reducing the response time of traditional keyword matching by more than 80%, effectively solving the problem of low efficiency in retrieving massive amounts of unstructured documents. In terms of matching accuracy, the vectorized encoding based on a pre-trained language model and the context-related weighted algorithm achieves a professional terminology recognition accuracy of over 95%, and the answer matching accuracy is improved by more than 40% compared to traditional methods, making it particularly suitable for scenarios where policy clauses are frequently updated. For example, when a user asks a question related to the "2023 fixed asset depreciation policy," the system can accurately associate it with the latest revised document content, avoiding incorrect answers due to outdated policies.

[0015] 2. The dynamic optimization capability of the system in this invention further enhances its adaptability. By collecting user satisfaction ratings, click behavior, and modified answer text, the feedback optimization module uses reinforcement learning algorithms to adjust the knowledge base weights in real time and dynamically supplements the database with new knowledge fragments contributed by users, shortening the system iteration cycle to within 24 hours. The popular question list generated by the popularity analysis unit not only improves the efficiency of user self-service but also provides data support for internal corporate training and policy dissemination. Simultaneously, the adaptive learning unit continuously optimizes the semantic encoding strategy by updating pre-trained model parameters online, ensuring that the system maintains high-precision output even when facing emerging financial terms or complex queries.

[0016] 3. In terms of automated auditing and risk control, this invention significantly improves the efficiency and accuracy of financial voucher processing by integrating OCR technology and machine learning models into the audit module. The voucher template configuration unit supports custom audit rules (such as invoice consistency verification and amount comparison), and the field extraction accuracy combined with OCR parsing can reach over 90%, improving efficiency by 60% compared to traditional manual review. The risk warning unit, based on an anomaly detection model trained on historical data, can effectively identify hidden risk behaviors such as duplicate payments and split payments, with a warning accuracy exceeding 85%, helping enterprises reduce financial risks by more than 60% in the payment process. For example, when the system detects that a payment has been split into multiple small transactions, it can immediately generate a risk report and block the abnormal process, avoiding compliance risks from the source.

[0017] 4. In this invention, the access control module achieves fine-grained security control through a three-tiered model of role-permission-user and a multi-factor authentication mechanism. Users in different positions (such as financial personnel in provincial-level industrial units and headquarters administrators) can only access the knowledge base and functional modules relevant to their responsibilities. Sensitive operations (such as modifying audit rules and exporting data) require dynamic verification and full traceability to ensure the traceability and compliance of system operations. This design, while ensuring data security, reduces operational errors caused by permission abuse and is suitable for the complex management needs of large group enterprises with multiple levels and cross-regional operations.

[0018] 5. The technical solution of this invention is significantly superior to traditional systems in terms of efficiency, accuracy, and adaptability. Actual test data shows that vectorized retrieval reduces the question-and-answer response time to less than 200 milliseconds, achieves a professional terminology recognition accuracy rate exceeding 95%, and achieves an abnormal payment behavior detection accuracy rate of over 90%. The system can be widely applied in corporate finance departments, provincial-level industrial units, and financial institutions for intelligent question answering, voucher auditing, and risk warning. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below.

[0020] The accompanying drawings described below are only related to some embodiments of the invention and are not intended to limit the invention.

[0021] In the attached diagram: Figure 1 The overall architecture diagram of the present invention is shown; Figure 2 An application architecture diagram of the present invention is shown; Figure 3 The data architecture diagram of the present invention is shown; Figure 4 The technical architecture diagram of the present invention is shown; Figure 5 The design diagram of the intelligent question-answering solution of the present invention is shown; Figure 6 The diagram shows the design of the financial automated verification scheme of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the described embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example: Please refer to Figures 1 to 6 : This invention proposes a financial knowledge question-answering matching method and system that combines vector representation, including the following steps: I. Constructing a Financial Knowledge Vector Library: Unstructured financial documents are decomposed into several knowledge fragments using a semantic segmentation algorithm; a pre-trained language model is used to vectorize and encode the knowledge fragments to generate multi-dimensional semantic vectors; based on vector similarity clustering, a hierarchical index structure is established to form a financial knowledge vector library. II. Vectorization of User Questions: Receive financial question text input by users and generate corresponding question vectors using a pre-trained language model; III. Dynamic Matching and Answer Generation: Calculate the similarity between the question vector and each vector in the knowledge vector base, and select the Top-N candidate answers; combine the contextual relevance and historical user feedback weights to rank the candidate answers and output the optimal result; IV. Feedback-driven knowledge base optimization: Collect user satisfaction ratings and click behavior data for answers; adjust the vector weights of knowledge fragments based on feedback data, and dynamically update the knowledge vector base.

[0024] The financial knowledge question-answering matching method based on vector representation includes the following modules: I. Knowledge Processing Module: Used for document segmentation, vector encoding, and index construction; II. Question Answering Engine Module: Includes a question vectorization unit, a similarity matching unit, and a sorting output unit; III. Feedback Optimization Module: Collects user behavior data and optimizes knowledge vector weights through reinforcement learning algorithms; IV. Automated Audit Module: Integrates OCR technology, rule engine, and machine learning model to perform structured parsing, field extraction, and abnormal behavior detection on uploaded financial vouchers.

[0025] The semantic segmentation algorithm includes: Named Entity Recognition (NER) based on a financial terminology dictionary to segment documents into policy clauses, operational steps, and question-answer pairs; and the use of graph neural networks (GNN) to construct a semantic association graph between knowledge fragments.

[0026] The automated audit module further includes a voucher template configuration unit and a risk warning unit. The voucher template configuration unit allows administrators to customize audit rules and field extraction templates. The risk warning unit trains an anomaly detection model based on historical data to identify duplicate payments and split payments.

[0027] The contextual relevance is calculated in the following way: 1. Extract at least one type of contextual information from user queries, including time, amount, and policy version number; 2. Combine contextual information with metadata of candidate answer fragments to generate a relevance score.

[0028] The feedback data further includes the user's modified answer text, which is then used as a new knowledge fragment, vectorized, and added to the financial knowledge vector library.

[0029] The knowledge processing module also includes a multimodal data integration unit, which supports joint vectorization encoding of mixed documents containing text, tables, and images to generate a unified semantic representation.

[0030] The feedback optimization module further includes a popularity analysis unit and an adaptive learning unit. The popularity analysis unit counts the frequency of user questions and the click rate of answers, generates a list of popular questions and displays them in priority. The adaptive learning unit updates the parameters of the pre-trained language model through online learning based on user feedback data.

[0031] It also includes: a permission management module: based on a three-level model of role-permission-user, it realizes permission allocation and operation log recording for users in different positions.

[0032] The access control module supports dynamically adjusting the scope of the knowledge base that can be accessed based on the user's organization and position, as well as performing multi-factor authentication and operation logging for sensitive operations.

[0033] The following provides further explanation and clarification of the above content: A financial knowledge question-answering matching method combining vector representation: The construction of the financial knowledge base takes unstructured documents (such as policies, regulations, and operational procedures) as input and decomposes them into fine-grained knowledge fragments using semantic segmentation algorithms. Semantic segmentation combines named entity recognition (NER) from a financial terminology dictionary to accurately segment structured content such as "policy clauses" and "operational procedures," and constructs a semantic relationship graph between knowledge fragments using graph neural networks (GNNs) to ensure the coherence of contextual logic. Subsequently, a pre-trained language model (such as BERT) is used to vectorize the knowledge fragments, generating high-dimensional semantic vectors. A hierarchical index structure is then built based on vector similarity clustering (e.g., using the Faiss library), significantly improving retrieval efficiency. This process not only supports text data but can also process mixed formats such as tables and scanned vouchers through a multimodal data integration unit to generate a unified semantic representation.

[0034] User-input financial questions (such as "How to accrue depreciation on fixed assets?") are converted into semantic vectors by a pre-trained language model. The system then selects Top-N candidate answers by calculating the cosine similarity between these vectors and those in the knowledge base. To improve matching accuracy, the system introduces a contextual relevance scoring mechanism: it extracts key metadata such as time, amount, and policy version number from user questions and performs correlation analysis with attributes like the effective date and scope of application of candidate answers. For example, when a user's question contains "2023," the system prioritizes matching the most recently revised policy content. Furthermore, historical user feedback weights (such as click-through rate and satisfaction rating) are dynamically incorporated into the ranking algorithm to ensure that high-value answers are displayed first.

[0035] The system continuously optimizes its knowledge base using user behavior data. User satisfaction ratings for answers, click behavior, and modified text are all collected and analyzed. If an answer's rating falls below a threshold, its weight is reduced to decrease its subsequent matching priority; user-modified answers are added to the database as new knowledge fragments. The adaptive learning unit updates pre-trained model parameters through online learning, for example, by using reinforcement learning algorithms to adjust vector encoding strategies, thereby improving the accuracy of semantic matching. Simultaneously, the popularity analysis unit generates a "Top 20 Hottest Questions" list to help users quickly locate common issues.

[0036] A financial knowledge question-answering matching system combining vector representation: The knowledge processing module is responsible for parsing and vectorizing multimodal data, supporting the joint encoding of text, tables, and images to ensure unified processing of complex data such as financial vouchers. The question-answering engine module achieves millisecond-level retrieval through hierarchical indexing and generates the final answer by combining contextual relevance and feedback weights. For example, when a user uploads a scanned invoice, the OCR parsing unit extracts key fields (such as amount and seller) and automatically compares them with audit rules in the knowledge base.

[0037] The automated audit module deeply integrates OCR technology and machine learning models, allowing administrators to customize audit rules (such as invoice consistency verification and signature integrity detection) through the voucher template configuration unit. The risk warning unit trains anomaly detection models based on historical data to identify high-risk behaviors such as duplicate payments and split payments. For example, when the system detects that a payment has been split into multiple smaller transactions, it automatically triggers an alert and generates a report, assisting finance personnel in quickly identifying risks.

[0038] The access control module is based on a three-tiered model of roles, permissions, and users, enabling fine-grained permission allocation. For example, users of provincial-level industrial units can only access the knowledge base of their respective institutions, while headquarters administrators can configure audit rules globally. Sensitive operations (such as rule modification and data export) require multi-factor authentication, and operation logs are kept throughout the process to ensure system security and compliance.

[0039] The following points should be noted in this article: 1. The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention; other structures can refer to general designs.

[0040] 2. Where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other to obtain new embodiments.

[0041] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A financial knowledge question-and-answer matching method combining vector representation, characterized in that, Includes the following steps: I. Constructing a Financial Knowledge Vector Library: Unstructured financial documents are decomposed into several knowledge fragments using a semantic segmentation algorithm; a pre-trained language model is used to vectorize and encode the knowledge fragments to generate multi-dimensional semantic vectors; based on vector similarity clustering, a hierarchical index structure is established to form a financial knowledge vector library.

2. Vectorization of user questions: Receive financial question text input by the user and generate corresponding question vectors using the pre-trained language model; III. Dynamic Matching and Answer Generation: Calculate the similarity between the question vector and each vector in the knowledge vector base, and select the Top-N candidate answers; combine the contextual relevance and historical user feedback weights to rank the candidate answers and output the optimal result; IV. Feedback-driven knowledge base optimization: Collect user satisfaction ratings and click behavior data for answers; The vector weights of knowledge fragments are adjusted based on feedback data, and the knowledge vector library is dynamically updated.

2. A financial knowledge question-and-answer matching system combining vector representation, used to implement the financial knowledge question-and-answer matching method combining vector representation as described in claim 1, characterized in that, Includes the following modules: I. Knowledge Processing Module: Used for document segmentation, vector encoding, and index construction; II. Question Answering Engine Module: Includes a question vectorization unit, a similarity matching unit, and a sorting output unit; III. Feedback Optimization Module: Collects user behavior data and optimizes knowledge vector weights through reinforcement learning algorithms; IV. Automated Audit Module: Integrates OCR technology, rule engine, and machine learning model to perform structured parsing, field extraction, and abnormal behavior detection on uploaded financial vouchers.

3. The financial knowledge question-answering matching method combining vector representation according to claim 1, characterized in that, The semantic segmentation algorithm includes: Named Entity Recognition (NER) based on a financial terminology dictionary to segment documents into policy clauses, operational steps, and question-answer pairs; and using Graph Neural Networks (GNNs) to construct a semantic association graph between knowledge fragments.

4. The financial knowledge question-answering matching system combining vector representation according to claim 2, characterized in that, The automated audit module further includes a voucher template configuration unit and a risk warning unit. The voucher template configuration unit allows administrators to customize audit rules and field extraction templates. The risk warning unit trains an anomaly detection model based on historical data to identify duplicate payments and split payments.

5. The financial knowledge question-answering matching method combining vector representation according to claim 1, characterized in that, The contextual relevance is calculated in the following way:

1. Extract at least one type of contextual information from user queries, including time, amount, and policy version number; 2. Combine the contextual information with the metadata of the candidate answer fragments to generate a relevance score.

6. The financial knowledge question-answering matching method combining vector representation according to claim 1, characterized in that, The feedback data further includes the user's modified answer text, which is then used as a new knowledge fragment, vectorized, and added to the financial knowledge vector library.

7. The financial knowledge question-answering matching system combining vector representation according to claim 2, characterized in that, The knowledge processing module also includes a multimodal data integration unit, which supports joint vectorization encoding of mixed documents containing text, tables, and images to generate a unified semantic representation.

8. The financial knowledge question-answering matching system combining vector representation according to claim 2, characterized in that, The feedback optimization module further includes a popularity analysis unit and an adaptive learning unit. The popularity analysis unit counts the frequency of user questions and the click rate of answers, generates a list of popular questions and displays them in priority. The adaptive learning unit updates the parameters of the pre-trained language model through online learning based on user feedback data.

9. The financial knowledge question-answering matching system combining vector representation according to claim 2, characterized in that, Also includes: Access control module: Based on a three-tier model of role-permission-user, it enables the allocation of permissions and recording of operation logs for users in different roles.

10. The financial knowledge question-answering matching system combining vector representation according to claim 9, characterized in that, The access control module supports dynamically adjusting the scope of the knowledge base that can be accessed based on the user's organization and position, and performs multi-factor authentication and operation logging for sensitive operations.

Images