Intelligent query expansion method and system based on financial knowledge network embedding

Abstract

The invention relates to the technical field of artificial intelligence information retrieval, and discloses an intelligent query expansion method and system based on financial knowledge network embedding, and the method comprises the following steps: S1, knowledge network construction and embedding: S11, extracting entities and relationships from multi-source financial data, and constructing a financial knowledge network; and S12, mapping entities and relationships in the financial knowledge network to a vector space by adopting a double-layer embedded framework. According to the intelligent query expansion method based on financial knowledge network embedding, high-quality query expansion is realized by deeply fusing a knowledge network and a semantic understanding technology; the semantic understanding depth of query expansion is improved; through a double-layer knowledge embedding framework and the semantic comprehension capability of a large language model, potential semantics of query can be deeply mined, and the meaning of polysemy words in a specific context can be accurately identified, so that an extension item which better fits the intention of a user is generated, and the knowledge basis of query extension is enriched.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence information retrieval technology, in particular to an intelligent query expansion method and system based on financial knowledge network embedding. BACKGROUND

[0002] With the complexity of the financial market and the explosive growth of information, analysts and investment managers in asset management and investment research need to query and process a large amount of unstructured data (such as announcements, financial reports, news, public opinion, industry research, etc.), and extract valuable information from them. The query submitted by the user is usually short, vague, general and ambiguous, which is difficult to accurately express the information demand, resulting in the retrieval system difficult to return the results that meet the user's needs. Query expansion technology, as a key technology to improve information retrieval effect, has become a key component in investment research systems.

[0003] Although the current query expansion technology has made some progress, it still faces multiple challenges in practical application in the financial field, mainly in three aspects of shallow semantic understanding, insufficient knowledge utilization and poor context adaptability.

[0004] In terms of semantic understanding, traditional statistical-based query expansion methods mainly rely on the co-occurrence frequency of terms, which cannot effectively capture deep semantic associations. Even if word embedding technology is used, there is a problem of too coarse semantic representation granularity, which makes it difficult to accurately understand the real intention of the query. Especially for polysemous words and professional terms, existing methods often fail to accurately identify their semantics in a specific context, leading to deviation of the expansion from the user's intention. For example, when the user queries "Apple new product release", the system may mistakenly introduce expansion words related to the fruit "apple", rather than the technology company information that the user is really interested in.

[0005] In terms of knowledge utilization, although existing technologies have begun to try to introduce external knowledge resources, the utilization method is still relatively simple and one-sided. First, existing methods often treat knowledge graphs as simple synonym libraries, ignoring the rich structural information and semantic relationships in them. Second, most methods only focus on entity-level knowledge, ignoring concept-level and domain-level knowledge organization, making it difficult to support cross-level knowledge reasoning. The financial field often involves a large amount of company announcements, financial data, industry analysis reports and macroeconomic information. If these complex multi-source heterogeneous information cannot be effectively integrated, it is difficult to accurately provide appropriate expansion items for the query.

[0006] In the aspect of context adaptability, the existing query expansion methods often adopt static expansion strategies, which cannot dynamically adjust the expansion direction and strength according to the query context and personalized needs of users. On the one hand, the combination between query expansion and retrieval model is not close enough, and the expansion process is relatively independent of the retrieval process, so it is difficult to dynamically optimize the expansion strategy according to the retrieval feedback. On the other hand, the query intention of users may be different in different time and different scenarios, and the existing methods are difficult to adaptively adjust the expansion parameters according to the historical behavior of users and the current context. The financial query may have completely different meanings in different scenarios, especially involving different time periods, industries or market conditions. For example, the same query has a large semantic difference in bull market and bear market, and the existing system is difficult to automatically adjust the query expansion strategy according to the specific context.

[0007] In summary, the existing query expansion technology has the key problems of shallow semantic understanding, insufficient knowledge utilization, poor context adaptability and the like, and an intelligent query expansion method capable of deeply integrating knowledge network and semantic understanding and supporting context perception is urgently needed to improve the performance and user experience of information retrieval systems in the financial field. SUMMARY

[0008] (1) Technical problems to be solved In view of the deficiencies in the prior art, the present application provides an intelligent query expansion method and system based on financial knowledge network embedding, to solve the problems of shallow semantic understanding, insufficient knowledge utilization and poor context adaptability in the traditional query expansion method in the background art.

[0009] (2) Technical solutions To achieve the above purpose, the present application provides the following technical solutions: an intelligent query expansion method based on financial knowledge network embedding, comprising the following steps: S1, knowledge network construction and embedding, comprising: S11, extracting entities and relationships from multi-source financial data to construct a financial knowledge network; S12, adopting a double-layer embedding framework to map the entities and relationships in the financial knowledge network to a vector space, wherein the double-layer embedding framework comprises: a basic embedding layer adopting a TransE model, introducing entity type information and relationship symmetry and transitivity constraints; a semantic enhancement layer using a pre-trained language model BERT to encode the entity context to generate a semantic enhancement representation; S13, adaptively fusing the representations of the basic embedding layer and the semantic enhancement layer based on task performance to generate a unified entity and relationship embedding vector; S2, query processing and vectorization, comprising: S21, using a large language model LLM to identify key entities in the user query and link to the financial knowledge network; S22, input the user query text into the pre-trained language model RoBERTa to generate the semantic vector of the query text; S23, fuse the embedding vector of the linked entity with the semantic vector of the query text to generate a comprehensive query vector; S3, expansion generation and optimization, including: S31, in the vector space of the financial knowledge network, calculate the similarity between the comprehensive query vector and all entity vectors to generate candidate expansion entities; S32, based on historical query log data, use LambdaRank learning ranking method to screen and rearrange the candidate expansion entities; S33, combine the screened expansion entities with the original query to generate an expanded query, and dynamically adjust the expansion entity weight.

[0010] Preferably, in the knowledge network construction and embedding step, the multi-source financial data includes structured databases, semi-structured documents and unstructured texts; In the TransE model used in the basic embedding layer, entity type information is introduced through type constraint regularization term, and relationship symmetry and transitivity are introduced through relationship constraint regularization term.

[0011] Preferably, in the semantic enhancement layer, the encoding of the entity context by the BERT model includes: Collect the context of the entity in the financial text corpus; Use the BERT model to encode the context to obtain the context representation; Average pooling is performed on all context representations to generate semantic enhanced representation of the entity.

[0012] Preferably, in the adaptive fusion step, the weight parameter is automatically determined by optimizing the performance of entity classification and relationship prediction tasks on the validation set.

[0013] Preferably, in the query entity identification and linking step, the LLM identifies the key entities in the query through a preset prompt template, and performs entity disambiguation according to the candidate entity information in the knowledge network In the comprehensive query vector generation step, the fusion method is weighted fusion, and the weight is dynamically adjusted according to the relevance of the entity and the query text.

[0014] Preferably, in the expansion generation and optimization step, the similarity calculation uses cosine similarity, and the approximate nearest neighbor search algorithm is used to improve the calculation efficiency; And the extended query adopts a Boolean query form, and the extended entity weight is dynamically adjusted according to the relevance degree of the extended entity to the original query.

[0015] The system for implementing the intelligent query expansion method based on financial knowledge network embedding comprises: The knowledge network construction and embedding module is configured to perform the functions described in step S1. The query processing and vectorization module is configured to perform the functions described in step S2. The expansion generation and optimization module is configured to perform the functions described in step S3. The system forms a flywheel closed loop through data flow and control flow, and continuously optimizes model parameters and expansion effects.

[0016] Preferably, the knowledge network construction and embedding module comprises: The knowledge acquisition unit is configured to extract entities and relationships from multi-source financial data. The knowledge representation and embedding unit is configured to perform the double-layer embedding framework. The knowledge fusion unit is configured to perform the adaptive fusion step.

[0017] Preferably, the query processing and vectorization module comprises: The query entity identification and linking unit is configured to identify and link query entities by using the LLM. The query vector generation unit is configured to fuse query text semantic vectors and entity embedding vectors.

[0018] Preferably, the expansion generation and optimization module comprises: The candidate expansion generation unit is configured to generate candidate expansion entities based on vector similarity. The expansion screening and rearrangement unit is configured to screen and rearrange the candidate expansions based on the LambdaRank learning ranking method. The query rewriting unit is configured to combine the screened expansion entities with the original query to generate an extended query.

[0019] (Three) beneficial effects Compared with the prior art, the present application provides an intelligent query expansion method and system based on financial knowledge network embedding, which has the following beneficial effects: The present application realizes high-quality query expansion by deeply fusing knowledge network and semantic understanding technology, improves the semantic understanding depth of query expansion, and through the semantic understanding ability of the double-layer knowledge embedding framework and the large language model, can deeply mine the potential semantics of the query, accurately identify the meaning of polysemous words in a specific context, and thus generate more user-intended expansion items.

[0020] The application enriches the knowledge base of query expansion; through constructing a multi-source heterogeneous knowledge network and mapping it to a unified vector space, efficient integration and utilization of knowledge are realized. At the same time, based on the adaptive knowledge fusion mechanism, the system can dynamically adjust the weight of knowledge representation according to different task requirements, ensuring the optimal fusion effect; finally, the application significantly enhances the context adaptability of query expansion. Through the generation of query vector, multi-dimensional expansion screening and rearrangement, and the continuous optimization mechanism based on user feedback, the expansion strategy can be adaptively adjusted according to the specific context of the query and the historical behavior of the user. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A step schematic diagram of the intelligent query expansion method based on the financial knowledge network embedding of the application; Figure 2 A prompt template schematic diagram of entity recognition of the application; Figure 3 A system block diagram of the intelligent query expansion system based on the financial knowledge network embedding of the application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0023] Referring to the accompanying Figure 1 The intelligent query expansion method based on the financial knowledge network embedding comprises the following steps: S1, knowledge network construction and embedding, comprising: S11, extracting entities and relationships from multi-source financial data to construct a financial knowledge network; In the knowledge network construction and embedding step, the multi-source financial data comprises a structured database, a semi-structured document and an unstructured text First, entities and relationships are directly extracted from the structured database. For each entity, record its unique identifier, name, type, attribute, etc. For each relationship, record its type, head entity, tail entity, etc.

[0024] Secondly, entities and relationships are extracted from semi-structured data (such as XML documents, HTML web pages, etc.) by parsing structured tags. For example, the relationship between entities can be extracted from the table of an HTML web page; the hierarchical relationship of entities can be extracted from the tag structure of an XML document.

[0025] Again, from unstructured text (such as news articles, academic papers, etc.) through natural language processing and document parsing technology, entities and relationships are extracted. The system uses a BERT-based named entity recognition (NER) model for entity extraction. Specifically, a pre-trained BERT model is used, which is fine-tuned on financial domain corpus to identify entities such as companies, stock codes, industry terms, sectors, events, etc., and label their types (such as "company", "industry classification", "sector", etc.). The relationship extraction uses a Transformer-based relationship extraction model, which encodes the context of entity pairs in the text to identify relationships between entities (such as "shareholder relationship", "upstream and downstream relationship", etc.). To improve extraction accuracy, the model combines dependency syntax analysis to enhance the ability to capture relationships in complex sentence structures, and adapts to specific relationship types in the financial domain through few-shot learning techniques.

[0026] Finally, the entities and relationships extracted from different data sources are integrated to build a unified knowledge network. During the integration process, problems such as entity alignment, relationship mapping, and conflict resolution need to be solved. When integrating multi-source data, the entity alignment uses a two-stage method: first, use a BERT-based semantic embedding model to generate vector representations of entity names, calculate cosine similarity to identify potential matches (such as "Apple Inc." and "Apple Inc."); second, combine the Levenshtein edit distance algorithm to handle name spelling differences (such as case, abbreviation), set a similarity threshold (for example, 0.9) for preliminary alignment. For complex scenarios, a graph neural network (GNN) is used to build a knowledge graph alignment model, which propagates node features (entity attributes) and edge features (relationship types) to optimize entity matching across data sources. Relationship mapping uses a predefined financial domain ontology to map relationship types from different data sources to a unified semantic framework, and uses rule matching and RoBERTa-based semantic classification models for relationship semantic consistency verification. Conflict resolution is based on data source credibility and frequency voting mechanism: prefer entities and relationships from high credibility data sources; when there is a conflict (such as the same entity having different attributes in different data sources), determine the final entity attribute through a weighted mechanism (weights based on data source credibility and update time). The fusion process uses a weighted average method to integrate the entity and relationship embedding vectors extracted from structured, semi-structured and unstructured data, and the weights are automatically determined by optimizing the performance of entity classification and relationship prediction tasks on the validation set, ensuring that the unified financial knowledge network representation balances semantic accuracy and structural integrity.

[0027] S12, using a double-layer embedding framework, mapping entities and relationships in the financial knowledge network to a vector space, wherein the double-layer embedding framework comprises: The basic embedding layer adopts a TransE model, and entity type information and relation symmetry and transitivity constraints are introduced; In the TransE model adopted by the basic embedding layer, entity type information is introduced through a type constraint regularization term, and relation symmetry and transitivity are introduced through a relation constraint regularization term.

[0028] The basic assumption of the TransE model is that for a triple (head entity, relation, tail entity) in a knowledge graph, the head entity vector plus the relation vector should be close to the tail entity vector.

[0029] The TransE model is improved in the application, including the following aspects: Entity type information is introduced, and the vector representation of entities of the same type should be closer. The symmetry and transitivity of the relationship are considered. If both relationships exist, this constraint should be considered. Based on the improved model, the company's business line, the upstream and downstream relationship between companies, and the association between stocks and macro economy are captured.

[0030] The semantic enhancement layer uses a pre-trained language model BERT to encode the context of the entity to generate a semantic enhancement representation. In the semantic enhancement layer, the encoding of the entity context by the BERT model includes: Collect the context of the entity in the financial text corpus; Use the BERT model to encode the context to obtain a context representation; Average pooling is performed on all context representations to generate a semantic enhancement representation of the entity.

[0031] The semantic enhancement layer uses a pre-trained language model to capture the contextual semantics of entities and concepts.

[0032] For each entity, first, its context in the text corpus needs to be collected; then, the pre-trained language model BERT is used to encode the context to obtain a context representation; After obtaining the representation with context information, all context representations need to be averaged and pooled to obtain the semantic representation of the entity. This can capture the semantic information of the entity in different contexts and make up for the possible semantic gap in the basic embedding layer.

[0033] S13, based on the task performance, adaptively fusing the representations of the basic embedding layer and the semantic enhancement layer to generate unified entity and relation embedding vectors; The weight parameters are automatically determined by optimizing the performance of entity classification and relation prediction tasks on the validation set; Knowledge fusion is responsible for fusing the representations of the basic embedding layer and the semantic enhancement layer into a unified knowledge representation. The invention adopts a weighted average method for fusion.

[0034] In order to adaptively determine the weight parameter, the invention designs a weight learning method based on task performance. The specific steps include: First, evaluate the performance of different weight values in entity classification and relationship prediction and other downstream tasks on the validation set; Second, select the weight value that optimizes the task performance as the final weight parameter.

[0035] In this way, the weights of the basic representation and the semantic representation can be adaptively adjusted according to different task requirements, thereby obtaining the optimal fused representation.

[0036] Not only effectively fuses structured knowledge and contextual semantics, solving the problem of shallow semantic understanding in traditional knowledge embedding methods, but also solves the problem of insufficient knowledge utilization in traditional methods, and realizes the effective integration of multi-source heterogeneous knowledge.

[0037] S2, query processing and vectorization, including: S21, using a large language model (LLM) to identify key entities in the user query and link to the financial knowledge network; The LLM identifies key entities in the query through a pre-set prompt template, and performs entity disambiguation based on candidate entity information in the knowledge network; The invention uses a large language model (LLM) as an agent to complete the entire process of entity recognition and linking, making full use of the powerful semantic understanding and knowledge reasoning capabilities of LLM.

[0038] First, input the original query text into the large language model. Based on the pre-designed prompt template, guide the large language model to identify the key entities in the query.

[0039] The prompt template, as shown in Figure 2 It contains clear task instructions, entity type definitions, examples and format requirements, ensuring that the LLM can accurately identify standardized entities and their types in the query.

[0040] Second, by constructing an entity linking prompt template, guide the large language model to establish a link between the identified entities and the nodes in the knowledge network.

[0041] The prompt template contains candidate entity information in the knowledge network, requiring the LLM to perform entity disambiguation based on context and semantic understanding.

[0042] Finally, the large language model will output the linking results and confidence of each entity, and the system will filter the final entity link S22, input the user query text into the pre-trained language model RoBERTa to generate a semantic vector of the query text; By converting the processed query into a vector representation, a foundation is provided for subsequent query expansion. The application combines query vector generation mechanisms to organically combine semantic information and entity knowledge information of the query text.

[0043] First, the semantic vector of the query text is calculated. A pre-trained language model RoBerta is used to encode the query text into a vector representation. The model encodes the query text into a context-dependent word vector sequence, and then performs a pooling operation on the word vector sequence to obtain a vector representation of the query text.

[0044] S23, fuse the embedding vector of the linked entity with the semantic vector of the query text to generate a comprehensive query vector; the fusion method is weighted fusion, and the weight is dynamically adjusted according to the relevance of the entity and the query text; Obtain the knowledge vector of the entity in the query, identify the key entity in the query through the query entity recognition and linking submodule, and establish a link between the key entity and the node in the knowledge network; for each linked entity, obtain its fusion embedding representation in the knowledge network.

[0045] If the query contains multiple entities, the fusion embedding representations of these entities are weighted averaged to obtain the knowledge vector of the query entity, and finally a comprehensive query vector is generated.

[0046] The application adopts a weighted fusion method to fuse the semantic vector of the query text and the knowledge vector of the query entity into a unified query representation.

[0047] Through the pre-designed prompt template, the powerful semantic understanding and knowledge reasoning ability of the large language model is utilized. This method breaks through the limitations of traditional entity recognition and linking methods, and can more accurately understand the query intent, and realize high-quality entity recognition and disambiguation.

[0048] The semantic vector of the query text is extracted by RoBerta, and the fusion embedding representation of the entity in the knowledge network is combined to generate a comprehensive query vector that can fully express the user's query intent. This mechanism effectively solves the problem of incomplete query representation in traditional methods, and improves the accuracy of query understanding.

[0049] S3, expansion generation and optimization, including: S31, in the vector space of the financial knowledge network, calculate the similarity of the comprehensive query vector and all entity vectors to generate candidate expansion entities; Candidate expansions are generated based on the semantic similarity of the vector space. Specifically, in the vector space of the knowledge network, the similarity of the query vector and all entity vectors is calculated, and the N entities with the highest similarity are selected as candidate expansions; The similarity calculation can be calculated by cosine similarity; the cosine similarity has a value range of [-1, 1], and the greater the value, the more similar the directions of the two vectors, that is, the more relevant the semantics of the query and the entity.

[0050] In order to improve the calculation efficiency, the approximate nearest neighbor search algorithm is adopted to quickly find the entity vector most similar to the query vector. In this way, it can be avoided to calculate one by one with all entities in the knowledge network, and the response speed of the system is greatly improved.

[0051] S32, based on the historical query log data, the LambdaRank learning ranking method is used to screen and rearrange the candidate expansion entity; Firstly, the correlation score of each candidate expansion and the original query is calculated, and then the screened candidate expansion is sorted, and the correlation degree of the candidate expansion and the query is considered in the sorting; The sorting algorithm adopts the learning ranking method based on LambdaRank, and is trained through a large amount of historical query log data.

[0052] S33, the screened expansion entity and the original query are combined to generate an expansion query, and the expansion entity weight is dynamically adjusted; and the expansion query adopts a Boolean query form, and the expansion entity weight is dynamically adjusted according to the correlation degree with the original query; By directly adding the expansion word to the original query; the expansion query adopts a Boolean query form, such as "(original query) AND (expansion word 1 OR expansion word 2 OR...)"; this structure can ensure that the retrieval result is related to the original query, and also considers the new information brought by the expansion word; Finally, the weight of each part in the expansion query is allocated; the weight of the original query should be higher than that of the expansion word, so as to ensure the core relevance of the retrieval result; at the same time, the weights of different expansion words can also be different, which can be dynamically adjusted according to the correlation degree with the original query.

[0053] It should be noted here that the financial knowledge network embedding in the present application specifically refers to the technology of mapping the knowledge nodes (such as companies, stocks, industries, economic indicators, financial events, etc.) and their relationships (such as shareholder relationship, investment relationship, industry association, etc.) in the financial field to a low-dimensional vector space, so that the semantic relationship in the financial knowledge network can be represented and calculated through vector operation.

[0054] The query expansion in the present application specifically refers to the technology of enriching and improving the query representation by adding a financial sub-intention query related to the original query in the information retrieval process, aiming to improve the recall rate and precision rate of the retrieval system.

[0055] The application can be widely applied to the field of AI finance, especially in asset management (fund management) and investment research (investment research) scenarios, to improve the accuracy and efficiency of intelligent investment research and automated investment decision-making.

[0056] The application realizes high-quality query expansion by combining knowledge network embedding technology with semantic understanding, thereby improving the recall rate and precision of the information retrieval system.

[0057] The application has the following characteristics: deeply utilizing the structure and semantic information of the knowledge network, fusing entity-level and concept-level knowledge representation, and considering the context information and user historical behavior of the query. These characteristics enable the application to effectively solve the technical problems of shallow semantic understanding, insufficient knowledge utilization, and poor context adaptability in traditional query expansion methods, thereby improving the performance of the information retrieval system.

[0058] Referring to the accompanying Figure 1 and Figure 3 , the system for implementing the intelligent query expansion method based on financial knowledge network embedding includes: The knowledge network construction and embedding module is used to perform the functions described in step S1; The knowledge network construction and embedding module is mainly responsible for constructing a knowledge network from multiple sources of financial data and mapping entities and relationships in the knowledge network to a vector space; The query processing and vectorization module is used to perform the functions described in step S2; The query processing and vectorization module is responsible for preprocessing the user query, identifying the key entities and intent in the query, and converting the query into a vector representation; The expansion generation and optimization module is used to perform the functions described in step S3; The expansion generation and optimization module generates high-quality query expansions based on the query vector and the knowledge network, and continuously optimizes the model parameters through user feedback The system forms a flywheel closed loop through data flow and control flow to continuously optimize the model parameters and expansion effect; Through the close connection of data flow and control flow, a flywheel closed loop is formed. The knowledge network construction and embedding module provides basic knowledge resources for the query processing and vectorization module, the query processing and vectorization module provides query representation for the expansion generation and optimization module, and the expansion generation and optimization module optimizes the model parameters through user feedback, thereby affecting the knowledge network construction and embedding module and the query processing and vectorization module.

[0059] In this embodiment, the knowledge network construction and embedding module includes: The knowledge acquisition unit is used to extract entities and relationships from multiple sources of financial data; The knowledge representation and embedding unit is used to execute the double-layer embedding framework; a knowledge fusion unit configured to perform the adaptive fusion step.

[0060] In this embodiment, the query processing and vectorization module comprises: a query entity recognition and linking unit configured to recognize and link query entities using the LLM; a query vector generation unit configured to fuse query text semantic vectors and entity embedding vectors.

[0061] In this embodiment, the expansion generation and optimization module comprises: a candidate expansion generation unit configured to generate candidate expansion entities based on vector similarity; an expansion screening and rearrangement unit configured to screen and rearrange candidate expansions based on a LambdaRank learning ranking method; a query rewriting unit configured to combine screened expansion entities and the original query to generate an expanded query.

[0062] Finally, it should be noted that: the above is only the preferred embodiment of the present application, and is not used to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it still can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An intelligent query expansion method based on financial knowledge network embedding, characterized in that, The method comprises the following steps: S1, knowledge network construction and embedding, comprising: S11, extracting entities and relationships from multi-source financial data to construct a financial knowledge network; S12, using a double-layer embedding framework to map entities and relationships in the financial knowledge network to a vector space, wherein the double-layer embedding framework comprises: a basic embedding layer using a TransE model, introducing entity type information and relationship symmetry and transitivity constraints; a semantic enhancement layer using a pre-trained language model BERT to encode the context of entities to generate semantic enhancement representations; S13, based on task performance, adaptively fusing the representations of the basic embedding layer and the semantic enhancement layer to generate unified entity and relationship embedding vectors; S2, query processing and vectorization, comprising: S21, using a large language model LLM to identify key entities in the user query and link them to the financial knowledge network; S22, inputting the user query text into a pre-trained language model RoBERTa to generate semantic vectors of the query text; S23, fusing the embedding vectors of the linked entities with the semantic vectors of the query text to generate comprehensive query vectors; S3, expansion generation and optimization, comprising: S31, calculating the similarity between the comprehensive query vectors and all entity vectors in the vector space of the financial knowledge network to generate candidate expansion entities; S32, based on historical query log data, using a LambdaRank learning ranking method to filter and reorder the candidate expansion entities; S33, combining the filtered expansion entities with the original query to generate an expanded query and dynamically adjusting the expansion entity weight. 2.The method of claim 1, wherein: In the knowledge network construction and embedding step, the multi-source financial data includes structured databases, semi-structured documents, and unstructured texts; In the TransE model used by the basic embedding layer, entity type information is introduced through a type constraint regularization term, and relationship symmetry and transitivity are introduced through a relationship constraint regularization term. 3.The method of claim 1, wherein: In the semantic enhancement layer, the encoding of the entity context by the BERT model comprises: collecting the context of the entity in the financial text corpus; using the BERT model to encode the context to obtain a context representation; averaging and pooling all context representations to generate a semantic enhancement representation of the entity. 4.The method of claim 1, wherein: In the adaptive fusion step, the weight parameter is automatically determined by optimizing the performance of entity classification and relationship prediction tasks on the validation set. 5.The method of claim 1, wherein: In the query entity identification and linking step, LLM identifies key entities in the query through a pre-set prompt template and performs entity disambiguation based on candidate entity information in the knowledge network In the comprehensive query vector generation step, the fusion method is weighted fusion, and the weight is dynamically adjusted according to the relevance of the entity and the query text. 6.The method of claim 1, wherein: In the expansion generation and optimization step, the similarity calculation uses cosine similarity, and the approximate nearest neighbor search algorithm is used to improve the calculation efficiency; And the expansion query uses a Boolean query form, and the expansion entity weight is dynamically adjusted according to its relevance to the original query.

7. A system for implementing the method of intelligent query expansion based on financial knowledge network embedding according to any one of claims 1-6, characterized in that, It comprises: a knowledge network construction and embedding module for performing the functions described in step S1; a query processing and vectorization module for performing the functions described in step S2; An expansion generation and optimization module is configured to perform the functions described in step S3. The system forms a flywheel closed loop through data flow and control flow to continuously optimize model parameters and expansion effects. 8.The financial knowledge network embedding based intelligent query expansion system of claim 7, wherein: The knowledge network construction and embedding module includes: A knowledge acquisition unit is configured to extract entities and relationships from multi-source financial data; A knowledge representation and embedding unit is configured to perform the double-layer embedding framework; A knowledge fusion unit is configured to perform the adaptive fusion step. 9.The financial knowledge network embedding based intelligent query expansion system of claim 7, wherein: The query processing and vectorization module includes: A query entity recognition and linking unit is configured to recognize and link query entities using an LLM; A query vector generation unit is configured to fuse query text semantic vectors and entity embedding vectors. 10.The financial knowledge network embedding based intelligent query expansion system of claim 7, wherein: The expansion generation and optimization module includes: A candidate expansion generation unit is configured to generate candidate expansion entities based on vector similarity; An expansion screening and rearrangement unit is configured to screen and rearrange candidate expansions based on a LambdaRank learning ranking method; A query rewriting unit is configured to combine screened expansion entities with the original query to generate an expanded query.

Images