An artificial intelligence-based financial data retrieval method and system thereof

Abstract

The application provides a financial data retrieval method and system based on artificial intelligence, which comprises the following steps: obtaining original query data and associated information data input by a user; inputting the associated information data into a user portrait model to obtain a user category output by the user portrait model; inputting the user category and the original query data into an intent recognition model to obtain a multi-layer retrieval intent output by the intent recognition model; determining a domain knowledge graph based on the user category, and performing disambiguation on terms in the multi-layer retrieval intent based on the domain knowledge graph to obtain a term disambiguation result; reconstructing a query retrieval logic tree based on the term disambiguation result, and determining a preliminary retrieval result based on the query retrieval logic tree; fusing and sorting the preliminary retrieval result based on the multi-layer retrieval intent to obtain a fused retrieval result, and screening the fused retrieval result to determine a target retrieval result. The application improves the speed and quality of the user in obtaining effective information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a financial data retrieval method and system based on artificial intelligence. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, more and more traditional industries have begun to explore and apply these advanced technologies to improve work efficiency and reduce error rates. In the field of financial management, financial data retrieval methods based on artificial intelligence are becoming a new trend. By using natural language processing, machine learning and other means, the required data can be quickly and accurately extracted from massive financial information, not only improving the speed of data processing, but also greatly enhancing the ability of decision support.

[0003] Although the financial data retrieval method based on artificial intelligence (AI) has shown great potential, there are still some problems to be solved. First, the user group of financial data retrieval can be divided into professionals and non-professionals, and there are essential differences in retrieval purpose and problem focus due to the significant differences in knowledge reserves and work needs. This difference leads to limitations of natural language processing technology in the application of financial data retrieval. On the one hand, financial field professional terms have polysemy, industry specificity, and semantics are often affected by context, making it difficult for AI models to accurately understand. On the other hand, the diverse expression habits and vague needs of users also increase the difficulty of semantic analysis, making it difficult for the model to accurately match the user's retrieval intent. In addition, the relationship between retrieval content and retrieval intent is complex. Although different users' retrieval content may overlap, the retrieval intent often differs, and current technology is difficult to accurately distinguish subtle intent differences, resulting in repeated retrieval content, wasting computing resources, and reducing retrieval efficiency, affecting the speed and quality of users obtaining effective information. SUMMARY

[0004] The present application provides a financial data retrieval method and system based on artificial intelligence to ensure that different types of users can obtain target retrieval results that meet their own needs, improving the speed and quality of users obtaining effective information.

[0005] In a first aspect, the present application provides a financial data retrieval method based on artificial intelligence, comprising:

[0006] Obtaining the original query data and associated information data input by the user, and inputting the associated information data into the user portrait model to obtain the user category output by the user portrait model; the user portrait model is trained based on sample associated information and its corresponding user category label result;

[0007] The user category and the original query data are input into the intent recognition model to obtain the multi-layered search intent output by the intent recognition model; the multi-layered search intent includes the main search intent and the sub-search intent; the intent recognition model is trained based on the sample user category, sample query data and their corresponding search intent label results;

[0008] Based on user categories, a domain knowledge graph is determined, and terms in the multi-layered search intent are disambiguated based on the domain knowledge graph to obtain term disambiguation results;

[0009] Based on the term disambiguation results, the query retrieval logic tree is reconstructed, and based on the query retrieval logic tree, preliminary retrieval results are determined;

[0010] The preliminary search results are merged and sorted based on multi-level search intent to obtain merged search results. The merged search results are then filtered based on user category to determine the target search results.

[0011] Secondly, the present invention also provides an artificial intelligence-based financial data retrieval system, applied to the artificial intelligence-based financial data retrieval method as described in the first aspect; the artificial intelligence-based financial data retrieval system includes:

[0012] The user profile building module is used to acquire the original query data and related information data input by the user; and input the related information data into the user profile model to obtain the user category output by the user profile model; the user profile model is trained based on the sample related information and its corresponding user category label results.

[0013] The intent recognition module is used to input the user category and the original query data into the intent recognition model to obtain the multi-layered search intent output by the intent recognition model; the multi-layered search intent includes the main search intent and the sub-search intent; the intent recognition model is trained based on the sample user category, sample query data and their corresponding search intent label results;

[0014] The terminology disambiguation module is used to determine the domain knowledge graph based on the user category, and to disambiguate the terms in the multi-layered search intent based on the domain knowledge graph to obtain the terminology disambiguation result;

[0015] The reconstructed retrieval module is used to reconstruct the query retrieval logic tree based on the term disambiguation results, and to determine the preliminary retrieval results based on the query retrieval logic tree;

[0016] The retrieval and filtering module is used to merge and sort the preliminary retrieval results based on multi-level retrieval intents to obtain merged retrieval results, and to filter the merged retrieval results based on user categories to determine the target retrieval results.

[0017] Thirdly, the present invention also provides an electronic device, comprising: a memory for storing computer software programs; and a processor for reading and executing the computer software programs, thereby realizing the artificial intelligence-based financial data retrieval method as described above.

[0018] Fourthly, the present invention also provides a non-transitory computer-readable storage medium storing a computer software program, which, when executed by a processor, implements the artificial intelligence-based financial data retrieval method described above.

[0019] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the artificial intelligence-based financial data retrieval method as described above.

[0020] The AI-based financial data retrieval method provided in this invention constructs user profiles by acquiring related information data. It employs different knowledge graphs for terminology disambiguation based on different user categories, effectively solving the semantic understanding challenges arising from the polysemy, specificity, and diverse user expressions in financial terminology, thus improving the accuracy of natural language processing in financial data retrieval. Furthermore, by deriving multi-layered search intents through an intent recognition model and integrating and ranking search results accordingly, it accurately distinguishes the different intents behind similar search content from different users, avoiding duplicate searches, improving retrieval efficiency, and reducing the waste of computing resources. Finally, from user profile construction, intent recognition, terminology disambiguation to search result processing, and personalized processing based on user categories, it ensures that different types of users can obtain target search results that meet their specific needs, improving the speed and quality of users obtaining effective information. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the artificial intelligence-based financial data retrieval method provided in an embodiment of the present invention;

[0022] Figure 2 This is a schematic diagram of the structure of the artificial intelligence-based financial data retrieval system provided in an embodiment of the present invention;

[0023] Figure 3 An embodiment diagram of the electronic device provided in this invention;

[0024] Figure 4 An embodiment diagram of a computer-readable storage medium provided in accordance with the present invention. Detailed Implementation

[0025] With reference to the drawings and the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0026] In the description of the present application, the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0027] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or description". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given so that any person skilled in the art can implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can realize the present application without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the shown embodiments, but is consistent with the broadest scope in accordance with the principles and characteristics disclosed.

[0028] Referring to Figure 1 , Figure 1 is a flowchart of the financial data retrieval method based on artificial intelligence provided by the present application. The execution subject of the financial data retrieval method based on artificial intelligence in the embodiments of the present application is a financial data retrieval system. Therefore, the financial data retrieval method based on artificial intelligence comprises:

[0029] Step 10, obtaining the original query data and the associated information data input by the user; and inputting the associated information data into the user portrait model to obtain the user category output by the user portrait model; the user portrait model is trained based on sample associated information and corresponding user category label results.

[0030] Optionally, the financial data retrieval system first acquires the original query data directly input by the user, including the retrieval keywords or sentences. Then, the associated information data is collected, wherein the associated information data includes the user identity label, the historical retrieval record, and the current operation scene. Specifically, the user identity label, such as the user ID (financial analyst or sales manager) and the corresponding identity permission, is acquired, and the historical retrieval record (frequently retrieving sales cost related data in the past three months, the system will record the position information when logging in the retrieval system, and the record of the usual query is obtained according to the position information) and the current operation scene (the position information is the current position work, such as preparing for the quarterly financial report) are acquired according to the user identity label. After obtaining the associated information data, the associated information data is input into the pre-trained user portrait model. The model is supervised learning and trained based on a large number of sample associated information and corresponding user category labels (such as professional financial personnel, business department personnel, and ordinary company employees), and adopts machine learning, reinforcement learning, deep neural network, and other model architectures. By analyzing the professional information in the user identity label, the data type preference in the historical retrieval record, and the task demand in the current operation scene, the user category is output, such as determining that the user is a “business department personnel”.

[0031] In an embodiment, taking user A as an example, the original query data is “view the sales return situation of North China in the first quarter of 2024”, the associated information data shows that the user identity label of user A is “regional sales manager”, the historical retrieval record of user A is mostly concentrated on regional sales performance, customer return data, and the current operation scene is the preparation of the regional sales performance review meeting. The user portrait model classifies user A as a “business department personnel” according to these information.

[0032] Step 20, input the user category and the original query data into the intent recognition model to obtain the multi-layer retrieval intent output by the intent recognition model; the multi-layer retrieval intent includes a main retrieval intent and a sub-retrieval intent; the intent recognition model is trained based on sample user categories, sample query data, and corresponding retrieval intent label results.

[0033] Optionally, after acquiring the specific user category, the financial data retrieval system inputs the user category and the original query data into the intent recognition model. The model is trained based on sample user categories, sample query data, and corresponding retrieval intent label results, and the model adopts deep learning and other artificial intelligence architectures, such as Transformer. The model first performs word segmentation and word vector embedding on the original query data, combines the user category information, identifies the multi-layer retrieval intent through the multi-layer attention mechanism, and specifically, the multi-layer retrieval intent includes a main retrieval intent and a sub-retrieval intent. Through the multi-layer intent, the user demand can be refined, the fuzzy retrieval can be avoided, the subsequent retrieval can more accurately match the user demand, and the pertinence of the retrieval is improved.

[0034] In an embodiment, continuing the example of user A, the "business department personnel" category and the "view the sales return on investment in North China in the first quarter of 2024" original query data are input into the intent recognition model. Since user A is a regional sales manager, the model identifies that the main search intent is "evaluate the sales return on investment performance in North China", and the sub-search intents include "compare the target return on investment with the actual return on investment" and "analyze the return on investment delay of each customer".

[0035] Step 30, based on the user category, determine the domain knowledge graph, and based on the domain knowledge graph, disambiguate the terms in the multi-layer search intent to obtain the term disambiguation result.

[0036] Optionally, the financial data retrieval system first pre-sets a knowledge graph library, which contains the domain knowledge graph under the corresponding user category, and then selects the corresponding domain knowledge graph from the knowledge graph library according to the user category determined in step 20. For example, if the user is a "financial audit personnel", a professional financial knowledge graph containing detailed accounting standards and audit specifications is selected; if the user is a "business department personnel", a knowledge graph focusing on business terminology interpretation is selected. Then, since financial domain terms have different meanings in different scenarios and user groups, only literal understanding can easily cause ambiguity, so according to the determined domain knowledge graph, the nodes and relationships in the knowledge graph are used to eliminate ambiguous terms in the multi-layer search intent, as described in steps 301-305. For example, for the "travel expense" term, in the professional financial knowledge graph, it is clear that it contains expense details, reimbursement standards, etc. definition, disambiguation, and accurate term meaning, i.e. term disambiguation result. Effectively solve the problem of polysemy and specificity of financial terms.

[0037] In an embodiment, for user A (business department personnel), the "sales return" term appears in his multi-layer search intent, and from the business domain knowledge graph, it is matched that "sales return" refers to the actual payment of sales to the account amount, and it is clear that it does not include information such as unaccounted receivables, etc., completing term disambiguation.

[0038] Step 40, based on the term disambiguation result, reconstruct the query search logic tree, and based on the query search logic tree, determine the preliminary search result.

[0039] Optionally, the financial data retrieval system uses the logical tree construction method to convert each semantic unit or sentence in the term disambiguation result into a structured query search logic tree based on the obtained term disambiguation result, as described in steps 401-403. Then, the logical tree is converted into a database query statement, and the query is executed in the financial data system to obtain the preliminary search result. The user's search intent is converted into an executable database query logic, improving the accuracy and efficiency of the search.

[0040] In an embodiment, for user A, a query retrieval logic tree is constructed according to the main retrieval intent and sub-retrieval intents of “Evaluate the sales return performance in North China region”. The root node is “Evaluation of sales return performance in North China region in the first quarter of 2024”, and the branch nodes include “Actual return amount query”, “Target return amount query”, “Customer return delay day calculation”, etc. Then the logic tree is converted into a SQL query statement, which is executed in the enterprise financial database to obtain the preliminary retrieval results containing fields such as customer return amount and return time.

[0041] Step 50, based on the multi-layer retrieval intent, the preliminary retrieval results are fused and sorted to obtain the fused retrieval results, and based on the user category, the fused retrieval results are filtered to determine the target retrieval results.

[0042] Optionally, the financial data retrieval system obtains the preliminary retrieval results according to step 40. Since the preliminary retrieval results may contain multiple related data, the multi-layer retrieval intent obtained in step 20 is used to fuse and sort the multiple related data. For example, the weight can be set according to the relevance and importance of the retrieval intent and the result data, the score of each result is calculated, and then the results are sorted according to the score to obtain the fused retrieval results. Through fusion and sorting, the results are filtered and sorted according to the importance of the retrieval intent, and the data related to the core needs of the user is highlighted. For details, see the description of steps 501-504. Then, according to the sorted fused retrieval results and the user category determined in step 10, the fused retrieval results are filtered again. Since different categories of users have different data needs and sensitivities, for example, enterprise executives may pay more attention to summary data, while financial analysts may need detailed detail data, therefore, through filtering, data that does not meet the user's knowledge level and business needs can be removed, so that the final target retrieval results are both accurate and meet the user's usage habits, and finally the target retrieval results are obtained. It should be noted that when filtering, the filtering rules can be set according to the user category. For example, for “business department personnel”, remove too professional financial analysis data and retain intuitive business indicator data.

[0043] The embodiment of the application constructs a user portrait by acquiring association information data, can adopt different knowledge graphs for term disambiguation for different user categories, effectively solves the semantic understanding problem caused by the polysemy and specificity of financial terms and the diversification of user expression, and improves the accuracy of natural language processing in financial data retrieval; in addition, the multi-layer retrieval intention is obtained through the intention recognition model, and the retrieval results are sorted based on this, which can accurately distinguish the different intention behind the similar retrieval content of different users, avoids repeated retrieval, improves the retrieval efficiency, and reduces the waste of computing resources; finally, from user portrait construction, intention recognition, term disambiguation to retrieval result processing, and combined with user category for personalized processing, it ensures that different types of users can obtain target retrieval results that meet their own needs, and improves the speed and quality of users to obtain effective information.

[0044] In an embodiment, steps 301-305 are described as follows:

[0045] Step 301, mapping the main retrieval intention based on the domain knowledge graph to obtain a top node set, and extracting the association relationship between the sub-retrieval intention and the top node set based on the domain knowledge graph to obtain a node sub-set.

[0046] Optionally, after the financial data retrieval system determines the domain knowledge graph, since the nodes of the knowledge graph represent various financial concepts, the edges represent the relationship between the concepts, and the main retrieval intention represents the core direction of the user's retrieval demand, therefore, mapping the main retrieval intention to the knowledge graph is to find the corresponding top concept nodes, which are relatively broad and abstract concepts in the knowledge graph. For example, in the financial knowledge graph, "financial statement analysis" may correspond to "financial statement" and "financial indicator analysis" top nodes, thereby forming a top node set C main The sub-retrieval intention is a refinement of the main retrieval intention, and the association relationship between the sub-retrieval intention and the nodes in the top node set is found through the knowledge graph, such as "balance sheet" is directly associated with "financial statement", and "debt paying ability indicator" is associated with "financial indicator analysis", thereby extracting the nodes related to the sub-retrieval intention to form a node sub-set. These nodes reflect the position and association of the sub-retrieval intention in the knowledge graph.

[0047] Step 302, determining the sub-intention weight of each sub-retrieval intention based on the importance of the main retrieval intention to the sub-retrieval intention.

[0048] Optionally, since different sub-retrieval intentions play different roles in the implementation of the main retrieval intention in the user's retrieval requirements, when the financial data retrieval system analyzes the contribution of each sub-retrieval intention to the implementation of the main retrieval intention, it can evaluate from multiple dimensions such as semantic relevance and business logic importance. For example, the main retrieval intention is "evaluate the profitability of the enterprise", and the sub-retrieval intention "analyze the composition of main business income" may be crucial to evaluating profitability, because main business income is a key component of profitability, while the sub-retrieval intention "statistical office supplies procurement cost" is also related to enterprise cost, but has less direct impact on profitability. Therefore, according to artificial financial professional knowledge and experience, different types of sub-retrieval intentions can be given different weights, such as sub-intentions involving core financial indicator analysis have high weights, while sub-intentions involving only basic information viewing of reports have low weights; or machine learning algorithms (such as multi-attribute decision analysis) can be used to learn the relationship between sub-retrieval intentions and user's final satisfaction, and each sub-retrieval intention is given a weight w i , the higher the priority, the greater the weight.

[0049] In an embodiment, taking the above "financial condition evaluation" main retrieval intention as an example, the sub-retrieval intentions "analyze the asset-liability ratio of the company in 2024" and "view the page number of the balance sheet". According to financial professional experience, "analyze the asset-liability ratio of the company in 2024" is crucial to evaluating financial condition, and is given a weight of 0.8; "view the page number of the balance sheet" is relatively unimportant, and is given a weight of 0.2.

[0050] Step 303, starting from the nodes in the top node set, diffusing along the edges of the domain knowledge graph to the lower layer nodes, determining the semantic similarity between each disambiguation term and each node on the diffusion path, and the sub-intention hierarchy to which each node belongs and the corresponding sub-intention weight; wherein the disambiguation term is any term in the multi-layer retrieval intention.

[0051] Optionally, the financial data retrieval system determines the semantic similarity between each disambiguation term and each node on the diffusion path according to the top node set C mainFrom each node in the top-level node set, the diffusion process traverses and spreads to its lower-level nodes along the edges between nodes in the knowledge graph. During the diffusion process, for each disambiguation term in the multi-layer retrieval intent (e.g., the word "control" in "cost control" can have different meanings in different contexts), the semantic similarity between it and the concepts represented by the nodes on the diffusion path is calculated. A path length-based method (e.g., the shorter the shortest path distance between two concepts in the knowledge graph, the higher the semantic similarity) or a vector embedding-based method (mapping concepts to a vector space and measuring semantic similarity by calculating the cosine similarity between vectors) can be used. At the same time, the sub-intent hierarchy to which each node belongs (i.e., which sub-retrieval intent the node is associated with) and the corresponding sub-intent weight are recorded.

[0052] Further, let the disambiguation term be t, the top-level node set be C main , and the nodes on the diffusion path be c i . The vector embedding-based semantic similarity calculation can use the cosine similarity formula where and are the vector representations of the disambiguation term t and the node c i , respectively. At the same time, let the sub-intent hierarchy to which the node c i belongs be l i , and the corresponding sub-intent weight be

[0053] In an embodiment, for the disambiguation term "liabilities" under the main retrieval intent "financial condition evaluation", the diffusion process starts from the top-level node "asset-liability analysis" and obtains lower-level nodes such as "current liabilities" and "long-term liabilities". The vector embedding method is used to calculate the semantic similarity between "liabilities" and "current liabilities" as 0.9 and the semantic similarity between "liabilities" and "long-term liabilities" as 0.8. "Current liabilities" belongs to the sub-intent hierarchy "analyze the asset-liability ratio of the company in 2024" and has a corresponding sub-intent weight of 0.8; "long-term liabilities" also belongs to the same sub-intent hierarchy and has a corresponding sub-intent weight of 0.8.

[0054] Step 304: For the node set obtained by diffusion, the association strength of the node with the disambiguation sentence and the hierarchical depth of the node in the domain knowledge graph are determined, and the node set is filtered based on the association strength and the hierarchical depth to obtain a candidate node set.

[0055] Optionally, the financial data retrieval system first determines the association strength of each node with the disambiguation sentence (i.e., the search intent sentence containing the disambiguation term) for the node set obtained in step 303. The association strength of the node with the disambiguation sentence can be determined based on the weight of the edge in the knowledge graph, which generally reflects the closeness of the association between concepts. For example, in the financial knowledge graph, the weight of the edge between "asset-liability ratio" and "liability" can be high, indicating that they are closely associated. At the same time, the hierarchical depth of each node in the domain knowledge graph is determined, and the deeper the hierarchical depth, the more specific and subdivided the concept. The hierarchical depth represents the number of layers from the top node, and a node with too deep hierarchical depth can contain concepts that are too subdivided and not closely related to the user's main requirements. A suitable association strength threshold and hierarchical depth threshold are set in advance to filter out nodes with low association strength and too deep hierarchical depth, and to retain more relevant nodes as the candidate node set C can .

[0056] Further, let the node set obtained by diffusion be C = {c1, c2, …, c n}, the association strength of node c i with the disambiguation sentence be Rel(t, c i ) (obtained through the edge weight of the knowledge graph), and the hierarchical depth of node c i in the domain knowledge graph be Depth(c i ). Set the association strength threshold to be τ rel and the hierarchical depth threshold to be τ depth , then the candidate node set C can = {c i | Rel(t, c i ) ≥ τ rel and Depth(c i ) ≤ τ deph , c i ∈ C}.

[0057] In an embodiment, for the node set related to "liability" obtained by diffusion, the association strength (according to the edge weight) of "accounts payable" with "liability" is 0.7, and the association strength of "wages payable-bonuses" with "liability" is 0.3. The hierarchical depth of "accounts payable" is 2, and the hierarchical depth of "wages payable-bonuses" is 3. Assuming that the association strength threshold is set to 0.4 and the hierarchical depth threshold is set to 3, "wages payable-bonuses" is filtered out, and nodes such as "accounts payable" that meet the conditions are retained in the candidate node set.

[0058] In step 305, for each node in the candidate node set, a comprehensive score is determined based on a pre-set comprehensive score function, and the candidate node set is screened based on the comprehensive score to obtain a term disambiguation result.

[0059] Optionally, the financial data retrieval system performs semantic similarity calculations on each node in the candidate node set obtained in step 204, based on the previously calculated semantic similarity Sim(t,c). i Sub-intent weights Correlation strength Rel(t,c) i Substitute information such as ) into the preset comprehensive scoring function: Calculate the overall score, where c represents any node in the candidate node set; n represents the number of sub-semantic hierarchy graphs associated with candidate node c; w i Represented as the sub-intent weight at the i-th sub-intent level; Sim(t,c i ) represents the node c corresponding to the i-th sub-intent level of the term t to be disambiguated. i Semantic similarity; Depth(c i ) represents the node c corresponding to the i-th sub-intention level. i Hierarchical depth in domain knowledge graphs; Rel(t,c i ) represents the node c corresponding to the i-th sub-intention level. i The strength of association between the node and the term t to be disambiguated. A higher score indicates a higher degree of match between the concept represented by the node and the term to be disambiguated. Then, the candidate nodes are ranked according to the comprehensive score, and the concept represented by the node with the highest score is selected as the term disambiguation result. If the comprehensive scores are all below a certain preset threshold, multiple candidate manual verification may be triggered, and the accurate disambiguation result will be determined by human judgment.

[0060] In one embodiment, the candidate node set includes "Inventory" and "Accounts Receivable" nodes. For the "Inventory" node, a comprehensive function is used to determine its value. The overall score was calculated to be 0.6; for the "Accounts Receivable" node, the overall score was calculated to be 0.3. The "Inventory" node scored higher, so "Inventory" was used as the disambiguation result for the "Inventory" term in "Inventory-Related Financial Indicators".

[0061] This invention maps the main search intent to a top-level node set and then extracts a subset of nodes by combining the relationships between sub-search intents. This allows for rapid location of regions in the knowledge graph related to the search intent, providing a precise search range for term disambiguation and improving disambiguation efficiency. Furthermore, it considers multiple dimensions such as sub-intent weight, semantic similarity, association strength, and hierarchical depth to comprehensively evaluate term meaning, avoiding disambiguation bias caused by a single factor and making the disambiguation results more accurate and reliable. In addition, it filters noisy concepts by considering association strength and hierarchical depth, reducing interference from irrelevant information and further improving the quality and efficiency of disambiguation. Finally, it filters candidate nodes based on a comprehensive score, determining the disambiguation results quantitatively to ensure objectivity. Meanwhile, a manual verification mechanism provides a reliable solution for special cases.

[0062] In an embodiment, steps 401-403 are described as follows:

[0063] Step 401, all semantic units in the term disambiguation result are taken as candidate nodes of the logical tree. For each pair of candidate nodes, a connectivity index is determined based on a connectivity function.

[0064] Optionally, the financial data retrieval system first extracts all semantic units in the term disambiguation result as candidate nodes of the logical tree. For example, if the term disambiguation result is “analyze the cost impact of the straight-line depreciation method of fixed assets of an enterprise”, the semantic units “enterprise”, “fixed assets”, “straight-line depreciation method”, and “cost impact” will all become candidate nodes. Then, for each pair of candidate nodes, the connectivity index is calculated according to the connectivity function: The shortest path length between nodes n1 and n2 is determined by a graph structure algorithm of the knowledge graph (such as Dijkstra's algorithm); the semantic similarity S lian (n1, n2) is determined by a semantic similarity calculation method (such as cosine similarity based on word vectors); and the intent dependency D lian (n1, n2) between nodes n1 and n2 is determined according to user search intent analysis. For example, “fixed assets” and “straight-line depreciation method” are connected by a related path in the knowledge graph, and their shortest path length is calculated. The semantic similarity between the two is calculated by word vectors, and the intent dependency is determined according to the close degree of the two concepts in the user search intent, and finally the connectivity index of the two is obtained.

[0065] In an embodiment, it is assumed that the term disambiguation result contains three semantic units “net profit”, “revenue”, and “cost” as candidate nodes. For the pair of candidate nodes “net profit” and “revenue”, the shortest path length between them is 3 (assuming that the path length between nodes in the graph is measured in the knowledge graph), the semantic similarity is 0.6 calculated by vectors, and the intent dependency is 0.7 determined according to business logic. According to the connectivity function: The connectivity index

[0066] Step 402, the candidate nodes are combined from high to low based on the connectivity index, and a tree topology structure is obtained.

[0067] Optionally, the financial data retrieval system combines the candidate nodes in order of the connectivity index from high to low according to the connectivity index calculated in step 401. High connectivity index means that the semantic association between the nodes is close and the degree of mutual dependence in meeting the user's intent is high. The nodes with high connectivity are combined first, and a tree topology is gradually constructed. For example, the nodes of "net profit" and "cost" with the highest connectivity are combined first because they are closely associated in financial analysis, and then other nodes are sequentially added to form a tree structure with clear hierarchy and close association between nodes.

[0068] In an embodiment, continuing the example of step 401, assume that the connectivity index of "net profit" and "cost" is the highest, which is 5.5; the connectivity index of "net profit" and "revenue" is 7.14; and the connectivity index of "revenue" and "cost" is 8.2. Then, "net profit" and "cost" are combined first, and then "revenue" is added in a suitable manner to form a tree topology, such as "net profit" as the parent node and "cost" and "revenue" as the child nodes (the specific structure is determined according to business logic and connection relationship).

[0069] In step 403, the tree topology is simplified and deleted, and the candidate node with the highest connectivity index is taken as the root node to construct a query retrieval logic tree.

[0070] Optionally, after the tree topology is constructed, the financial data retrieval system deletes edges with a connectivity index lower than a threshold value to simplify the tree structure and avoid excessive complexity. At the same time, the node with the highest connectivity index and consistent with the main intent is selected from the candidate nodes as the root node, because the root node is the core of the logic tree, and high connectivity and consistency with the main intent can better guide the entire logic tree, make it revolve around the core concept, and more accurately express the user's retrieval intent.

[0071] In an embodiment, the connectivity threshold is set to 7. In the above example, the connectivity index of "revenue" and "cost" is 8.2, which is higher than the threshold value and is retained; the connectivity index of "net profit" and "revenue" is 7.14, which is higher than the threshold value and is retained; and the connectivity index of "net profit" and "cost" is 5.5, which is lower than the threshold value and is deleted. Assume that the connectivity index of "net profit" is the highest among all nodes and is consistent with the main intent "financial profit analysis", then "net profit" is taken as the root node, and the final query retrieval logic tree is constructed, with "revenue" and "cost" as the child nodes hanging under the "net profit" node.

[0072] The embodiment of the application starts from the disambiguated semantic units, constructs a logical tree based on a connectivity function, can accurately map the semantic and concept correlation in the user search intention, makes the logical tree accurately express the user demand, and constructs a logical tree with reasonable structure, simplicity and clarity by sorting and combining nodes and simplifying and deleting operations through the connectivity index, highlights the key concepts, avoids redundancy, and improves the search efficiency and accuracy.

[0073] In an embodiment, steps 501-504 are described as follows:

[0074] In step 501, the main search intention, the sub-search intention and the preliminary search result are respectively subjected to semantic vector conversion to obtain a main search intention semantic vector, a sub-search intention semantic vector and a preliminary search result semantic vector.

[0075] Optionally, the financial data search system converts the keywords in the main search intention and the sub-search intention and the text information in the preliminary search result into semantic vectors by using a word vector model such as Word2Vec or GloVe. For example, the keywords such as “enterprise” and “financial situation” in the main search intention “analyze the financial situation of an enterprise” are converted into semantic vectors; the keyword “profit rate” in the sub-search intention “calculate the profit rate” is converted into a semantic vector; and the keywords in the preliminary search result “2024 net profit rate data of enterprise” are also converted into semantic vectors. In this way, the text information is converted into numerical vector form that can be processed and compared by a computer, and is prepared for subsequent similarity calculation.

[0076] In an embodiment, it is assumed that the main search intention is “analyze the revenue situation of an enterprise”, and the keywords “enterprise” and “revenue situation” are converted into semantic vectors and The sub-search intention is to view the quarterly revenue data, and the keywords “quarterly” and “revenue data” are converted into semantic vectors and The preliminary search result is the revenue report of the enterprise in the second quarter of 2024, and the keywords “enterprise”, “2024”, “second quarter” and “revenue report” are converted into semantic vectors

[0077] In step 502, the main search intention semantic vector and the sub-search intention semantic vector are respectively subjected to similarity processing with the preliminary search result semantic vector to obtain a first similarity score and a second similarity score.

[0078] Optionally, the financial data retrieval system usually adopts cosine similarity calculation method when calculating the similarity, that is, the cosine value of the angle between two vectors is calculated to measure their similarity, and the value closer to 1 indicates that the two vectors are more similar. Specifically, for the main retrieval intention semantic vector and the preliminary retrieval result semantic vector, the cosine similarity between them is calculated to obtain a first similarity score, which reflects the matching degree of the preliminary retrieval result and the main retrieval intention. Similarly, for each sub-retrieval intention semantic vector and the preliminary retrieval result semantic vector, the cosine similarity is calculated respectively to obtain multiple second similarity scores, which reflect the matching degree of the preliminary retrieval result and each sub-retrieval intention.

[0079] In an embodiment, the example continues: the main retrieval intention semantic vector (composed of and ) and the preliminary retrieval result semantic vector (composed of and ), according to the cosine similarity formula , the first similarity score is calculated to be 0.7. The sub-retrieval intention semantic vector and , the second similarity score is calculated to be 0.6, and , the second similarity score is calculated to be 0.5.

[0080] Step 503, based on the first similarity score and the second similarity score, a similarity matrix is constructed; each row in the similarity matrix represents a preliminary retrieval result, and each column corresponds to the similarity score of the main retrieval intention and each sub-retrieval intention.

[0081] Figure 1 Optionally, after the financial data retrieval system obtains the similarity scores of all preliminary retrieval results and the main retrieval intention and the sub-retrieval intention, the scores are organized into a matrix form. Each row of the matrix corresponds to a preliminary retrieval result, and each column corresponds to the similarity score of the main retrieval intention or a sub-retrieval intention. For example, assuming that there are m preliminary retrieval results, the main retrieval intention is I , and the sub-retrieval intention is k, then the constructed similarity matrix is an m*(k+1) matrix. The similarity between each preliminary retrieval result and different retrieval intentions can be clearly shown, which facilitates subsequent sorting operations.

[0082] In an embodiment, assuming that there are 3 preliminary retrieval results R1, R2, R3, the main retrieval intention is I main , and the sub-retrieval intention is I sub1 , I sub2 . The first similarity score of R1 and I main is 0.7, and the first similarity score of R1 and I sub1R2 has a second similarity score of 0.5 with I sub2 R2 has a second similarity score of 0.5 with I main R2 has a second similarity score of 0.5 with I sub1 R2 has a second similarity score of 0.5 with I sub2 R2 has a second similarity score of 0.5 with I main R2 has a second similarity score of 0.5 with I sub1 R2 has a second similarity score of 0.5 with I sub2 R2 has a second similarity score of 0.5 with I

[0083] Main search intent similarity I sub1 similarity I sub2 similarity [R1] 0.7 0.6 0.5 [R2] 0.6 0.4 0.3 [R3] 0.8 0.7 0.6 .

[0084] Step 504, weighting and ranking the preliminary search results in the similarity matrix based on the preset nested ranking strategy to obtain the fusion search results.

[0085] Optionally, the nested ranking strategy preset by the financial data search system is formulated according to the importance and hierarchical relationship of the main search intent and the sub-search intents. Generally, the first similarity score (i.e., the similarity score with the main search intent) is arranged in descending order first, because the main search intent represents the main direction of the user search, and the preliminary search results with high matching degree with the main search intent should be ranked in the front. When the first similarity scores are the same, the second similarity scores (i.e., the similarity scores with the sub-search intents) are further arranged in descending order. The weighting and comparison or the sequential comparison can be set according to the priority order of the sub-search intents. For example, the second similarity scores corresponding to the important sub-search intents are compared first, and if the scores are still the same, the scores of other sub-search intents are compared, and finally the fusion search results are obtained, so that the search results are arranged in descending order according to the matching degree with the user search intents.

[0086] In an embodiment, for the above similarity matrix, according to the nested ranking strategy, the main search intent similarity scores are arranged in descending order first, R3 (0.8) is ranked in the front, then R1 (0.7), and finally R2 (0.6). In this example, the first similarity scores can already distinguish the order, and there is no need to further compare the sub-search intent similarity scores. If there is a case that the first similarity scores are the same, such as R4, the main search intent similarity score of which is also 0.7, the sub-search intent similarity scores need to be further compared. It is assumed that the sub-search intents I sub1 has a higher priority than I sub2 , and the similarity score of R1 with I sub1 is 0.6, which is higher than the similarity score of R4 with I sub1If the similarity score of R1 and R4 is 0.5, R1 is placed in front of R4, and the final fusion retrieval result is R3, R1, R2 (assuming that R4 does not exist in the final result).

[0087] The embodiment of the application further mines the semantic association between the retrieval intention and the retrieval result by semantic vector conversion and similarity calculation, avoids the limitation of sorting only from surface keyword matching, improves the understanding and processing capability of semantics, and sorts the retrieval result gradually and accurately from the whole to the part based on the hierarchical structure of the main retrieval intention and the sub-retrieval intention by using a nested sorting strategy, so that the sorting result is more in line with the logic of user retrieval demand, and the quality and practicability of the retrieval result are improved. Finally, the similarity information is structured by constructing a similarity matrix, a clear and ordered data structure is provided for the sorting operation, numerical calculation and comparison are facilitated, and the efficiency and accuracy of sorting are improved.

[0088] Further, the financial data retrieval system based on artificial intelligence provided by the application is described below, and the financial data retrieval system based on artificial intelligence described below can be correspondingly referred to the financial data retrieval method based on artificial intelligence described above.

[0089] Optionally, referring to Figure 2 , Figure 2 is a structural schematic diagram of the financial data retrieval system based on artificial intelligence provided by the application, and the financial data retrieval system based on artificial intelligence comprises:

[0090] The user portrait construction module 210 is configured to acquire original query data and associated information data input by a user, input the associated information data into a user portrait model, and obtain a user category output by the user portrait model. The user portrait model is trained based on sample associated information and corresponding user category label results.

[0091] The intention recognition module 220 is configured to input the user category and the original query data into an intention recognition model, and obtain a multi-layer retrieval intention output by the intention recognition model. The multi-layer retrieval intention comprises a main retrieval intention and a sub-retrieval intention. The intention recognition model is trained based on sample user categories, sample query data, and corresponding retrieval intention labels.

[0092] The term disambiguation module 230 is configured to determine a domain knowledge graph based on the user category, disambiguate terms in the multi-layer retrieval intention based on the domain knowledge graph, and obtain a term disambiguation result.

[0093] The restructured retrieval module 240 is configured to restructure a query retrieval logic tree based on the term disambiguation result, and determine a preliminary retrieval result based on the query retrieval logic tree.

[0094] The retrieval screening module 250 is used for fusing and sorting the preliminary retrieval results based on the multi-layer retrieval intention to obtain fused retrieval results, and screening the fused retrieval results based on the user category to determine the target retrieval results.

[0095] The embodiment of the application can construct a user portrait by acquiring associated information data, can adopt different knowledge graphs for term disambiguation for different user categories, effectively solves the semantic understanding problem caused by the polysemy and specificity of financial terms and the diversification of user expressions, and improves the accuracy of natural language processing in financial data retrieval; in addition, the multi-layer retrieval intention is obtained through the intention recognition model, and the retrieval results are fused and sorted based on the multi-layer retrieval intention, which can accurately distinguish the different intention behind the similar retrieval content of different users, avoid repeated retrieval, improve the retrieval efficiency, and reduce the waste of computing resources; finally, from user portrait construction, intention recognition, term disambiguation to retrieval result processing, and combined with user category for personalized processing, it is ensured that different types of users can obtain target retrieval results meeting their own needs, and the speed and quality of users obtaining effective information are improved.

[0096] Please refer to Figure 3 , Figure 3 The embodiment of the electronic device provided by the embodiment of the application is shown in the figure. Figure 3 As shown in the figure, the embodiment of the application provides an electronic device 300, which includes a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320, and the processor 320 implements the following steps when executing the computer program 311:

[0097] Obtain the original query data and associated information data input by the user; and input the associated information data into the user portrait model to obtain the user category output by the user portrait model; the user portrait model is trained based on sample associated information and corresponding user category label results;

[0098] Input the user category and the original query data into the intention recognition model to obtain the multi-layer retrieval intention output by the intention recognition model; the multi-layer retrieval intention includes a main retrieval intention and a sub-retrieval intention; the intention recognition model is trained based on sample user categories, sample query data and corresponding retrieval intention label results;

[0099] Based on the user category, determine the domain knowledge graph, and disambiguate the terms in the multi-layer retrieval intention based on the domain knowledge graph to obtain the term disambiguation result;

[0100] Based on the term disambiguation result, reconstruct the query retrieval logic tree, and based on the query retrieval logic tree, determine the preliminary retrieval results;

[0101] The primary retrieval result is fused and sorted based on the multi-layer retrieval intention to obtain a fusion retrieval result, and the fusion retrieval result is filtered based on the user category to determine a target retrieval result.

[0102] Please refer to Figure 4 , Figure 4 An embodiment of a computer-readable storage medium provided for the embodiment of the application is shown in the figure. Figure 4 As shown in the figure, the embodiment provides a computer-readable storage medium 400, which stores a computer program 311, and the computer program 311 is executed by a processor to implement the following steps:

[0103] Obtain the original query data and the associated information data input by the user, and input the associated information data into a user portrait model to obtain the user category output by the user portrait model; the user portrait model is trained based on sample associated information and corresponding user category label results;

[0104] Input the user category and the original query data into an intention recognition model to obtain the multi-layer retrieval intention output by the intention recognition model; the multi-layer retrieval intention includes a main retrieval intention and a sub-retrieval intention; the intention recognition model is trained based on sample user categories, sample query data and corresponding retrieval intention label results;

[0105] Based on the user category, determine a domain knowledge graph, and disambiguate the terms in the multi-layer retrieval intention based on the domain knowledge graph to obtain a term disambiguation result;

[0106] Based on the term disambiguation result, reconstruct a query retrieval logic tree, and based on the query retrieval logic tree, determine a primary retrieval result;

[0107] Based on the multi-layer retrieval intention, fuse and sort the primary retrieval result to obtain a fusion retrieval result, and filter the fusion retrieval result based on the user category to determine a target retrieval result.

[0108] On the other hand, the application also provides a computer program product, which includes a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer can execute the artificial intelligence-based financial data retrieval method provided by each method, and the method includes:

[0109] Obtain the original query data and the associated information data input by the user, and input the associated information data into a user portrait model to obtain the user category output by the user portrait model; the user portrait model is trained based on sample associated information and corresponding user category label results;

[0110] Input the user category and the original query data into the intention recognition model to obtain a multi-layer retrieval intention output by the intention recognition model; the multi-layer retrieval intention includes a main retrieval intention and a sub-retrieval intention; the intention recognition model is trained based on sample user categories, sample query data, and retrieval intention label results thereof;

[0111] Based on the user category, determine a domain knowledge graph, and disambiguate terms in the multi-layer retrieval intention based on the domain knowledge graph to obtain a term disambiguation result;

[0112] Based on the term disambiguation result, reconstruct a query retrieval logic tree, and based on the query retrieval logic tree, determine a preliminary retrieval result;

[0113] Based on the multi-layer retrieval intention, fuse and rank the preliminary retrieval result to obtain a fused retrieval result, and based on the user category, filter the fused retrieval result to determine a target retrieval result.

[0114] The system embodiments described above are merely illustrative, wherein the units shown as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, i.e., they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0115] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus the necessary general hardware platform, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods of the various embodiments or some parts of the embodiments.

[0116] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features thereof; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A financial data retrieval method based on artificial intelligence, characterized in that, include: Obtain the user's original query data and related information data; The associated information data is then input into the user profile model to obtain the user category output by the user profile model; The user profile model is trained based on sample association information and its corresponding user category label results; The user category and the original query data are input into the intent recognition model to obtain the multi-layered search intent output by the intent recognition model; The multi-layered search intent includes the main search intent and sub-search intent; The intent recognition model is trained based on sample user categories, sample query data, and their corresponding search intent tag results. Based on user categories, a domain knowledge graph is determined, and terms in the multi-layered search intent are disambiguated based on the domain knowledge graph to obtain term disambiguation results; Based on the term disambiguation results, the query retrieval logic tree is reconstructed, and based on the query retrieval logic tree, preliminary retrieval results are determined; The preliminary search results are merged and sorted based on multi-level search intent to obtain merged search results. The merged search results are then filtered based on user category to determine the target search results.

2. The financial data retrieval method based on artificial intelligence according to claim 1, characterized in that, The associated information data includes user identity tags, historical search records, and current operation scenarios. The user identity tags include user ID and its corresponding identity permissions.

3. The financial data retrieval method based on artificial intelligence according to claim 1, characterized in that, The disambiguation of terms in the multi-layered search intent based on the domain knowledge graph, to obtain term disambiguation results, includes: The main search intent is mapped based on the domain knowledge graph to obtain a top-level node set, and the association between the sub-search intent and the top-level node set is extracted based on the domain knowledge graph to obtain a node subset; Based on the importance of the sub-search intent to the main search intent, the sub-intent weight of each sub-search intent is determined; Starting from the nodes in the top-level node set, the algorithm diffuses down to the lower-level nodes along the edges of the domain knowledge graph to determine the semantic similarity between each term to be disambiguated and each node on the diffusion path, as well as the sub-intent level to which each node belongs and the corresponding sub-intent weight; the term to be disambiguated is any term in the multi-level retrieval intent. For the node set obtained from diffusion, determine the association strength between the node and the statement to be disambiguated, as well as the hierarchical depth of the node in the domain knowledge graph. Then, filter the node set based on the association strength and hierarchical depth to obtain a candidate node set. For each node in the candidate node set, a comprehensive score is determined based on a preset comprehensive score function, and the candidate node set is filtered based on the comprehensive score to obtain the term disambiguation result.

4. The financial data retrieval method based on artificial intelligence according to claim 3, characterized in that, The synthesized function is: Where c represents any node in the candidate node set; n represents the number of sub-meaning hierarchy graphs associated with candidate node c; w i Represented as the sub-intent weight at the i-th sub-intent level; Sim(t,c i ) represents the node c corresponding to the i-th sub-intent level of the term t to be disambiguated. i Semantic similarity; Depth(c i ) represents the node c corresponding to the i-th sub-intention level. i Hierarchical depth in domain knowledge graphs; Rel(t,c i ) represents the node c corresponding to the i-th sub-intention level. i The strength of association with the term t to be disambiguated.

5. The financial data retrieval method based on artificial intelligence according to claim 1, characterized in that, The process of reconstructing the query retrieval logic tree based on the term disambiguation results includes: All semantic units in the term disambiguation results are used as candidate nodes of the logic tree. For each pair of candidate nodes, a connectivity index is determined based on the connectivity function. Candidate nodes are combined from high to low based on the connectivity index to obtain a tree-like topology. The tree-like topology is simplified and deleted, and the candidate node with the highest connectivity index is used as the root node to construct the query retrieval logic tree.

6. The financial data retrieval method based on artificial intelligence according to claim 5, characterized in that, The connectivity function is: Among them, C lian (n1, n2) represents the connectivity index between nodes n1 and n2; L lian (n1, n2) represents the shortest path length between nodes n1 and n2; S lian (n1, n2) represents the semantic similarity between nodes n1 and n2; D lian (n1, n2) represents the intent dependency graph of nodes n1 and n2.

7. The financial data retrieval method based on artificial intelligence according to claim 1, characterized in that, The preliminary search results are merged and sorted based on multi-level search intent to obtain merged search results, including: The main search intent, sub-search intent, and preliminary search results are respectively transformed into semantic vectors to obtain the semantic vectors of the main search intent, sub-search intent, and preliminary search results. The semantic vectors of the main search intent and the sub-search intent are compared with the semantic vectors of the preliminary search results to obtain a first similarity score and a second similarity score. Based on the first similarity score and the second similarity score, a similarity matrix is ​​constructed; each row in the similarity matrix represents a preliminary search result, and each column corresponds to the similarity score of the main search intent and each sub-search intent, respectively. The preliminary search results in the similarity matrix are weighted and sorted based on a preset nested sorting strategy to obtain the fused search results.

8. A financial data retrieval system based on artificial intelligence, characterized in that, The method for retrieving financial data based on artificial intelligence as described in any one of claims 1 to 7 is applicable; the system for retrieving financial data based on artificial intelligence comprises: The user profile building module is used to acquire the original query data and related information data input by the user; and input the related information data into the user profile model to obtain the user category output by the user profile model; the user profile model is trained based on the sample related information and its corresponding user category label results. The intent recognition module is used to input the user category and the original query data into the intent recognition model to obtain the multi-layered search intent output by the intent recognition model; the multi-layered search intent includes the main search intent and the sub-search intent; the intent recognition model is trained based on the sample user category, sample query data and their corresponding search intent label results; The terminology disambiguation module is used to determine the domain knowledge graph based on the user category, and to disambiguate the terms in the multi-layered search intent based on the domain knowledge graph to obtain the terminology disambiguation result; The reconstructed retrieval module is used to reconstruct the query retrieval logic tree based on the term disambiguation results, and to determine the preliminary retrieval results based on the query retrieval logic tree; The retrieval and filtering module is used to merge and sort the preliminary retrieval results based on multi-level retrieval intents to obtain merged retrieval results, and to filter the merged retrieval results based on user categories to determine the target retrieval results.

9. An electronic device, comprising: Memory, used to store computer software programs; A processor for reading and executing the computer software program, characterized in that, when the processor executes the computer software program, it implements the artificial intelligence-based financial data retrieval method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, wherein a computer software program is stored therein, characterized in that, When the computer software program is executed by the processor, it implements the artificial intelligence-based financial data retrieval method as described in any one of claims 1 to 7.

Images