Enterprise consultation data retrieval method and system based on graph enhanced retrieval generation

Abstract

The invention provides an enterprise consultation data retrieval method and system based on graph enhanced retrieval generation, and relates to the technical field of knowledge graph construction and retrieval, and the method comprises the following steps: introducing segmented word semantic features, context features, segmented word position features and an entity similarity mechanism based on a structured semantic unit to carry out entity recognition and relation extraction; constructing an enterprise knowledge graph; obtaining user retrieval content, and performing retrieval in the enterprise knowledge graph by calculating the matching similarity between the user retrieval content and each entity node to obtain at least one entity node as a candidate matching result; optimizing the candidate matching result based on the similarity edge in the enterprise knowledge graph to obtain a final matching result; and extracting an answer corresponding to the target question of the user retrieval content from the entity node corresponding to the final matching result as a retrieval result. The method has the advantage that the expression ability and retrieval precision of the enterprise knowledge graph are improved.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph construction and retrieval technology, and more specifically, to a method and system for retrieving enterprise consulting data generated based on graph augmentation retrieval. Background Technology

[0002] With the development of intelligent applications in various industries, knowledge graph-based enterprise data retrieval has gradually become an important and common intelligent query method.

[0003] Enterprise knowledge graphs are typically constructed by performing entity recognition and relationship extraction on enterprise text data. Then, in retrieval enhancement generation, they respond to user queries based on entity name matching, keyword retrieval, or graph structure path search. However, existing enterprise consulting data retrieval methods based on knowledge graphs and retrieval enhancement still have the following shortcomings: they are mainly constructed based on entities and their external relationships, with a weak ability to capture potential associations between entities. This leads to limitations in the feature representation of the enterprise knowledge graph, resulting in insufficient retrieval matching accuracy when dealing with complex retrieval content or when there are many possible matching results, such as ambiguous matches. Furthermore, the construction of enterprise knowledge graphs is usually carried out through rule matching, keyword extraction, or traditional sequence labeling models, which further reduces the accuracy of the enterprise knowledge graph representation. In conventional knowledge graph-based retrieval enhancement generation, after obtaining candidate entities, they are usually directly sorted according to matching scores or path lengths to obtain relevant information needed for retrieval, which then guides the generation of the final text. This approach yields relatively coarse retrieval results, failing to filter out some false matches and resulting in redundant feedback results.

[0004] Therefore, we will continue to optimize the construction method of enterprise knowledge graph and the retrieval method based on retrieval enhancement to improve the expressive power of enterprise knowledge graph and the accuracy of retrieval. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for retrieving enterprise consulting data based on graph-enhanced retrieval, which improves the expressive power of enterprise knowledge graphs and the accuracy of retrieval.

[0006] This invention is achieved through the following technical solution:

[0007] The enterprise consulting data retrieval method based on graph augmentation retrieval includes the following steps: Based on structured semantic units, word segmentation semantic features, context features, word segmentation position features, and entity similarity mechanisms are introduced to perform entity recognition and relation extraction, and an enterprise knowledge graph is constructed. The enterprise knowledge graph includes entity nodes, attribute nodes, relation edges, similarity edges, and dependent edges. Entity nodes storing entities are connected to the corresponding attribute nodes storing entity attributes through dependent edges. Entity nodes are connected to each other through relation edges and similarity edges. The similarity edges are marked with the similarity between entity nodes. The system retrieves user search content, calculates the matching similarity between user search content and each entity node, and performs a search in the enterprise knowledge graph to obtain at least one entity node as a candidate matching result. The candidate matching results are optimized based on the similarity edges in the enterprise knowledge graph to obtain the final matching result; Extract the answer to the target question of the user's search content from the entity node corresponding to the final matching result as the search result.

[0008] Preferably, the method for constructing the enterprise knowledge graph is as follows: Acquire enterprise data and perform standardization processing, which includes syntactic segmentation, uniform formatting, and deduplication. Each sentence is segmented using a Hidden Markov Model to obtain structured semantic units based on words. Based on structured semantic units, entities and their corresponding entity attributes, as well as the relationships between entities and other entities, are extracted. Entities and their corresponding entity attributes are connected by dependent edges, and entities and entities with which they have a relationship are connected by relational edges, thus obtaining an initial enterprise knowledge graph. Based on the entity and its corresponding entity attributes, the identification parameters of each entity are obtained, and the similarity between entities is calculated based on the identification parameters. Entities with similarity greater than a preset threshold are connected by similarity edges, and the similarity is marked on the similarity edges to obtain the enterprise knowledge graph.

[0009] Preferably, the method for extracting the structured semantic units is as follows: Multiple entity description vectors are created and initialized to an empty set. The first element of the entity description vector is used to store the entity name, and each of the remaining elements stores a pair of attribute categories and corresponding attribute parameters. Establish an entity relationship matrix and initialize it to an empty set. The m-th row and n-th column of the entity relationship matrix represents the relationship between the m-th entity and the n-th entity. Initialize sentence number i=1, initialize entity description vector number j=1, and initialize element number of the j-th entity description vector. ; Perform the following operations until all sentences have been traversed: The classification label of each word in the i-th sentence is obtained through a word segmentation and classification model. The classification label includes entity, relationship between entities, attribute category and attribute parameter. Determine the number of entities in the i-th sentence. If the number of entities is one, check if the entity in the i-th sentence has already been stored in the entity description vector. If not, store the entity in the i-th sentence in the j-th entity description vector and update the value of j to j=j+1. Simultaneously, perform the following operations on the attribute categories and corresponding attribute parameters in the i-th sentence in sequence: [The text abruptly ends here, likely due to an incomplete translation or a missing section.] Each parameter, and update The value is If there are more than one entity, then determine whether each entity in the i-th sentence has been stored in the entity description vector. If not, store the entity of the i-th sentence in the j-th entity description vector and update the value of j to j=j+1. At the same time, store the entity relationships in the i-th sentence in the entity relationship matrix. Then, perform the following operations on the attribute categories and corresponding attribute parameters in the i-th sentence in sequence: store the entity in the j-th entity description vector... Each parameter, and update The value is ; Update the value of i to i = i + 1; After completing the traversal, the initial enterprise knowledge graph is constructed based on the entity description vector and the entity relationship matrix.

[0010] Preferably, the word segmentation and classification model includes: The input layer takes as input the word segment to be classified and all the word segments in the sentence. Feature extraction layer extracts semantic features from word segmentation. Contextual features and word segmentation position features Then, feature fusion is performed to obtain word segmentation fusion features. : ; ; ; ; ; ; in, It is a bidirectional long short-term memory neural network. To construct a character sequence function that is ordered character-by-character, This is a function that extracts the vector of each character from a character-based vector index table and concatenates them. For the word segmentation to be classified, The word preceding the segment to be classified. The next word in the segmentation to be classified. This represents the total number of words in the sentence containing the word to be classified. Number the segmented words to be classified sequentially within the sentence. and They are respectively and The semantic features of word segmentation, when the preceding or following word of the word to be classified does not exist, the corresponding... or Set the value to zero; The output layer outputs based on word segmentation fusion features. The probability of belonging to each of the stated category labels is calculated, and the label with the highest probability is selected as the classification label. The category tags mentioned above: ; ; ; in, and Represent Belongs to the kth type of category label The probability and rating, It is a natural exponential function. For intermediate parameters, , and For training weights, , and For training bias.

[0011] Preferably, the method for obtaining the identifier parameters of each entity and calculating the similarity between entities based on the identifier parameters is as follows: Calculate the identifier parameters based on the entity's entity description vector: ; in, For the first The entity description vector of the nth entity The attribute category number corresponding to each element. For the first The entity description vector of the nth entity The attribute parameters of the attribute category corresponding to each element. For the first The total number of elements in the entity description vector of each of the entities; Calculate the first The entity and the first The method for determining the similarity between entities is as follows: Obtain the first based on the identifier parameter The entity and the first The attribute categories that all entities possess form a common set of attribute categories; According to the The entity and the first The difference between the number of attribute categories and the corresponding attribute parameter values ​​for each entity is used to calculate the first... The entity and the first The similarity between entities : ; in, and The t-th attribute category in the set of common attribute categories is the t-th attribute category in the set of common attribute categories. The entity and the first The numerical values ​​of the corresponding attribute parameters in each entity, where T is the total number of elements in the common attribute category set.

[0012] Preferably, the method for searching in the enterprise knowledge graph is as follows: Semantic parsing is performed based on the user's search content to extract the search dimensions, which include entity name dimension, attribute dimension, relationship dimension, and search target. Taking each entity node as a unit, the following operations are performed on each entity node: Matching searches are performed on the enterprise knowledge graph according to the retrieval dimensions. Cosine similarity is used to obtain entity similarity scores, multiple attribute similarity scores, and multiple relation similarity scores. The entity similarity score is the similarity between the entity stored in the entity node and the entity name dimension. The attribute similarity score is the maximum value of the similarity between the attribute dimension involved in the user's search content and all related attribute nodes. Related attribute nodes are the attribute nodes connected to the entity node through the subordinate edge. The relation similarity is the similarity between the relation dimension involved in the user's search content and all relation edges of the entity node. The matching similarity of the entity node is obtained by weighted summation of relation similarity, node similarity, and attribute similarity; After obtaining the matching similarity of all the entity nodes, the entity nodes with matching similarity greater than a preset threshold are selected as the candidate matching results.

[0013] Preferably, the method for optimizing candidate matching results based on similarity edges in the enterprise knowledge graph involves performing the following operations for each candidate matching result: Obtain all entity nodes connected to the entity nodes corresponding to the candidate matching results through similarity edges, and use them as optimization nodes; Obtain the matching similarity between the user's search content and each optimization node, and calculate the optimization parameters of the candidate matching results; The final score of the candidate matching result is calculated based on the optimization parameters and the matching similarity. After obtaining the final scores of all the candidate matching results, the candidate matching results with a final score greater than a preset threshold are selected as the final matching results.

[0014] Preferably, the method for calculating the optimization parameters of the candidate matching results is as follows: ; in, For the first The optimization parameters for the candidate matching results. For the first The entity node number corresponding to each of the candidate matching results. This represents the entity node number corresponding to the d-th optimization node. For the first The first entity node and the first The similarity score marked on the similar edges of each entity node. For the user's search content and the first The matching similarity of each entity node.

[0015] Preferably, the method for calculating the final score of the candidate matching results is as follows: ; in, For the first The final score of each of the candidate matching results. It is a natural exponential function. For the user's search content and the first The matching similarity of each entity node.

[0016] This invention also provides a business information data retrieval system based on graph-enhanced retrieval, applied to the aforementioned business information data retrieval method based on graph-enhanced retrieval, comprising: The enterprise knowledge graph construction module is used to construct an enterprise knowledge graph based on entity attribute structure and entity similarity association mechanism through entity recognition and relation extraction. The enterprise knowledge graph includes entity nodes, attribute nodes, relation edges, similarity edges, and dependent edges. Entity nodes storing entities are connected to the corresponding attribute nodes storing entity attributes through dependent edges. Entity nodes are connected to each other through relation edges and similarity edges. The similarity edges are marked with the similarity between entity nodes. The preliminary retrieval module is used to obtain the user's search content, calculate the matching similarity between the user's search content and each entity node, and search the enterprise knowledge graph to obtain at least one entity node as a candidate matching result. The retrieval optimization module is used to optimize candidate matching results based on similarity edges in the enterprise knowledge graph to obtain the final matching result; Extract the answer to the target question of the user's search content from the entity node corresponding to the final matching result as the search result.

[0017] The technical solution of the present invention has at least the following advantages and beneficial effects: This invention introduces character-level word segmentation semantic features, context features, and word segmentation position features to achieve refined discrimination of the category to which each word belongs, significantly reducing the probability of confusion between entity categories, relationship categories, and attribute categories, improving the recognition accuracy of each node and each edge of the enterprise knowledge graph, and thus improving the structural consistency and semantic reliability of the knowledge graph. This invention reduces reliance on manual rules, manual annotation templates, and domain dictionaries in the process of building enterprise knowledge graphs, thereby improving the automation and scalability of the enterprise knowledge graph construction process. The enterprise knowledge graph structure of this invention is clear and highly interpretable, and it also helps to discover redundant and duplicate content through similarity edges, thereby improving the graph storage efficiency. This invention introduces a similarity mechanism, which can help further optimize search results during the retrieval stage, filter out redundant preliminary search results, improve search accuracy, and reduce the redundancy of search results. This invention is reasonably designed and has a simple structure, making it applicable to the establishment of enterprise knowledge graphs in different types and environments, and easy to promote and implement. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the enterprise consulting data retrieval method based on graph augmentation retrieval provided in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the structure of the enterprise consulting data retrieval system based on graph-enhanced retrieval provided in Embodiment 2 of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0020] Example 1 This embodiment provides a method for retrieving enterprise consulting data generated based on graph augmentation retrieval. (See also...) Figure 1 This includes the following steps: Step S1: Based on structured semantic units, introduce word segmentation semantic features, context features, word segmentation position features, and entity similarity mechanisms to perform entity recognition and relation extraction, constructing an enterprise knowledge graph. The enterprise knowledge graph includes entity nodes, attribute nodes, relation edges, similarity edges, and dependent edges. Entity nodes storing entities are connected to their corresponding attribute nodes via dependent edges. Entity nodes are connected to each other via relation edges and similarity edges, with the similarity between entity nodes marked on the similarity edges. It is important to note that Step S1 is executed only once in the entire process, constructing the enterprise knowledge graph for use in subsequent steps. Alternatively, the enterprise knowledge graph can be periodically reconstructed or updated using the method described in Step S1.

[0021] In this embodiment, the method for constructing the enterprise knowledge graph is as follows: Step S101: Acquire and standardize enterprise data, such as basic enterprise information, operational data, consultation records, business reports, and industry texts. Standardization includes syntactic segmentation, unified formatting, and deduplication. The aim is to divide long texts into independent sentence units and to standardize time formats, monetary formats, encoding methods, etc. This step effectively improves the consistency and structure of text data, providing effective input for subsequent semantic analysis.

[0022] Step S102: Perform word segmentation on each sentence using a Hidden Markov Model (HMM) to obtain structured semantic units based on words. The HMM is a common word segmentation model that can model word boundaries and adaptively identify word boundaries without manual rule intervention.

[0023] Step S103: Extract entities and their corresponding entity attributes, as well as the relationships between entities and other entities, based on structured semantic units. Connect entities and their corresponding entity attributes through dependent edges, and connect entities and entities with which they have a relationship through relational edges to obtain the initial enterprise knowledge graph. The preferred method in this embodiment is as follows: Multiple entity description vectors are created and initialized to an empty set. The first element of the entity description vector is used to store the entity name, and each of the remaining elements stores a pair of attribute categories and corresponding attribute parameters. Establish an entity relationship matrix and initialize it to an empty set. The m-th row and n-th column of the entity relationship matrix represents the relationship between the m-th entity and the n-th entity. Initialize sentence number i=1, initialize entity description vector number j=1, and initialize element number of the j-th entity description vector. ; Perform the following operations until all sentences have been traversed: The classification label of each word in the i-th sentence is obtained through a word segmentation and classification model. The classification label includes entity, relationship between entities, attribute category and attribute parameter. Determine the number of entities in the i-th sentence. If the number of entities is one, check if the entity in the i-th sentence has already been stored in the entity description vector. If not, store the entity in the i-th sentence in the j-th entity description vector and update the value of j to j=j+1. Simultaneously, perform the following operations on the attribute categories and corresponding attribute parameters in the i-th sentence in sequence: [The text abruptly ends here, likely due to an incomplete translation or a missing section.] Each parameter, and update The value is If there are more than one entity, then determine whether each entity in the i-th sentence has been stored in the entity description vector. If not, store the entity of the i-th sentence in the j-th entity description vector and update the value of j to j=j+1. At the same time, store the entity relationships in the i-th sentence in the entity relationship matrix. Then, perform the following operations on the attribute categories and corresponding attribute parameters in the i-th sentence in sequence: store the entity in the j-th entity description vector... Each parameter, and update The value is ; Update the value of i to i = i + 1; After completing the traversal, the initial enterprise knowledge graph is constructed based on the entity description vector and the entity relationship matrix.

[0024] Specifically, the preferred word segmentation and classification model used in the above scheme includes: The input layer takes as input the word segment to be classified and all the word segments in the sentence. Feature extraction layer extracts semantic features from word segmentation. Contextual features and word segmentation position features Then, feature fusion is performed to obtain word segmentation fusion features. : ; ; ; ; ; ; in, It is a bidirectional long short-term memory neural network. To construct a character sequence function that is ordered character-by-character, This is a function that extracts the vector of each character from a character-based vector index table and concatenates them. For the word segmentation to be classified, The word preceding the segment to be classified. The next word in the segmentation to be classified. This represents the total number of words in the sentence containing the word to be classified. Number the segmented words to be classified sequentially within the sentence. and They are respectively and The semantic features of word segmentation, when the preceding or following word of the word to be classified does not exist, the corresponding... or Set the value to zero; The above feature extraction method can enhance the context awareness of semantic discrimination and avoid ambiguity caused by relying solely on isolated words. For example, "register" may be an action relationship or an attribute category, and relying on ordinary recognition may not be able to accurately determine it.

[0025] Extracting semantic features of word segmentation In this process, word segmentation is broken down into character sequences, then a LookupTable is used to map the characters into vectors, and finally, BiLSTM is used to model the dependencies between characters, thereby outputting the morphological and semantic representation of the word. This method can infer the category of words that are not in the vocabulary through character semantics. Contextual features In the extraction process, semantic features of the preceding and following words of the target word are extracted separately to establish local contextual semantic awareness. For example, in the previously mentioned "registration," "registered capital" can be determined from the following context to be an attribute category, while "register a company" can be determined from the following context to be a behavioral relationship. (Word segmentation position features) It is a sentence-level ordinal feature. According to traditional grammar, the position of each word can be used as a feature to express certain word attributes. Therefore, positional features can enhance the ability to judge structure.

[0026] The output layer outputs based on word segmentation fusion features. The probability of belonging to each of the stated category labels is calculated, and the label with the highest probability is selected as the classification label. The category tags mentioned above: ; ; ; in, and Represent Belongs to the kth type of category label The probability and rating, It is a natural exponential function. For intermediate parameters, , and For training weights, , and For training bias.

[0027] Output layer through The initial classification label is determined based on the semantic features of the word segmentation itself, and then the weights are adaptively adjusted according to the context information, that is, coefficients are assigned. On the one hand, it avoids excessive influence of context on classification results based on its own semantics; on the other hand, it can optimize and enhance judgment accuracy through context.

[0028] In step S103, each segment is defined by traversing each segment and the relationship between segments is extracted based on the characteristic that the segments belong to the same sentence. A sentence is equivalent to a business fact. This method is equivalent to using the sentence as the boundary of the semantic scope, reducing erroneous relationship edges and avoiding semantic drift.

[0029] Step S104: Based on the entity and its corresponding entity attributes, obtain the identification parameters of each entity and calculate the similarity between entities based on the identification parameters. Connect entities with similarity greater than a preset threshold through similarity edges and mark the similarity on the similarity edges to obtain the enterprise knowledge graph.

[0030] Based on the above scheme, the method for obtaining the identifier parameters of each entity and calculating the similarity between entities based on the identifier parameters is as follows: Calculate the identifier parameters based on the entity's entity description vector: ; in, For the first The entity description vector of the nth entity The attribute category number corresponding to each element. For the first The entity description vector of the nth entity The attribute parameters of the attribute category corresponding to each element. For the first The total number of elements in the entity description vector of each of the entities; Calculate the first The entity and the first The method for determining the similarity between entities is as follows: Obtain the first based on the identifier parameter The entity and the first The attribute categories that all entities possess form a common set of attribute categories; According to the The entity and the first The difference between the number of attribute categories and the corresponding attribute parameter values ​​for each entity is used to calculate the first... The entity and the first The similarity between entities : ; in, and The t-th attribute category in the set of common attribute categories is the t-th attribute category in the set of common attribute categories. The entity and the first The numerical values ​​of the corresponding attribute parameters in each entity, where T is the total number of elements in the common attribute category set.

[0031] Step S104 converts entity and attribute information into computable entity identifier parameters, and constructs similarity relationships between entities based on these parameters, thereby enhancing the relevance, reasoning ability, and retrieval capability of the knowledge graph. The similarity between entities includes both the similarity of overlapping attributes and the degree of attribute overlap. Used to measure the relative differences in values ​​for the same attribute. The similarity is measured by the proportion of common attributes between two entities. Both values ​​are better the smaller they are. The average value is then subtracted from 1 to obtain the similarity score, which is better the larger it is.

[0032] Step S2: Obtain the user's search content, calculate the matching similarity between the user's search content and each entity node, and search in the enterprise knowledge graph to obtain at least one entity node as a candidate matching result.

[0033] The preferred method for searching the enterprise knowledge graph in this embodiment is as follows: Step S201: Perform semantic parsing based on the user's search content, and extract the search dimensions, including entity name dimension, attribute dimension, relationship dimension, and search target; Step S202: Taking the entity node as a unit, perform the following operations for each entity node: Matching searches are performed on the enterprise knowledge graph according to the retrieval dimensions. Cosine similarity is used to obtain entity similarity scores, multiple attribute similarity scores, and multiple relation similarity scores. The entity similarity score is the similarity between the entity stored in the entity node and the entity name dimension. The attribute similarity score is the maximum value of the similarity between the attribute dimension involved in the user's search content and all related attribute nodes. Related attribute nodes are the attribute nodes connected to the entity node through the subordinate edge. The relation similarity is the similarity between the relation dimension involved in the user's search content and all relation edges of the entity node. The matching similarity of the entity node is obtained by weighted summation of relation similarity, node similarity, and attribute similarity; Step S203: After obtaining the matching similarity of all the entity nodes, select the entity nodes whose matching similarity is greater than a preset threshold as the candidate matching results.

[0034] The above scheme first decomposes the natural language retrieval content into structured retrieval dimensions. The retrieval dimensions correspond to the constituent modules of the enterprise knowledge graph in this embodiment. Therefore, matching can be performed on each node to form a unified matching similarity score. Finally, entity nodes with high matching similarity are output as candidate results, reducing noise and improving accuracy, and obtaining candidate matching results as the initial results.

[0035] Step S3: Optimize the candidate matching results based on the similarity edges in the enterprise knowledge graph to obtain the final matching results.

[0036] Specifically, the method for optimizing candidate matching results based on similarity edges in the enterprise knowledge graph involves performing the following operations for each candidate matching result: Step S301: Obtain all entity nodes connected by similarity edges to the entity nodes corresponding to the candidate matching results through the enterprise knowledge graph, and use them as optimization nodes.

[0037] Step S302: Obtain the matching similarity between the user's search content and each optimization node, and calculate the optimization parameters of the candidate matching results. The method for calculating the optimization parameters of the candidate matching results is as follows: ; in, For the first The optimization parameters for the candidate matching results. For the first The entity node number corresponding to each of the candidate matching results. This represents the entity node number corresponding to the d-th optimization node. For the first The first entity node and the first The similarity score marked on the similar edges of each entity node. For the user's search content and the first The matching similarity of each entity node.

[0038] Step S303: Calculate the final score of the candidate matching result based on the optimization parameters and the matching similarity, using the following method: ; in, For the first The final score of each of the candidate matching results. It is a natural exponential function. For the user's search content and the first The matching similarity of each entity node.

[0039] Step S304: After obtaining the final scores of all the candidate matching results, select the candidate matching results with a final score greater than a preset threshold as the final matching results.

[0040] Unlike traditional methods that rely solely on the similarity between user search content and candidate matching results, this scheme introduces similarity edges from the enterprise knowledge graph as a further constraint. Specifically, it uses similar entity nodes of candidate entities to perform consistency checks and score corrections on the candidate matching results. If the user's search intent matches the trend of the candidate matching results and similar entity nodes, the candidate entity is considered more trustworthy; otherwise, a significant deviation indicates a possible mismatch. The Sigmoid function is used to avoid sudden changes in scores. In other words, this embodiment enhances the stability and reliability of search results through a multi-node semantic voting mechanism.

[0041] Step S4: Extract the answer corresponding to the target question of the user's search content from the entity nodes corresponding to the final matching results as the search results. Specifically, this can be achieved by capturing the question intent through semantic parsing and then retrieving the corresponding answer based on the final matched entity nodes.

[0042] Example 2 This embodiment provides a business information data retrieval system based on graph augmentation retrieval, applied to the aforementioned business information data retrieval method based on graph augmentation retrieval. See [link / reference]. Figure 2 ,include: The enterprise knowledge graph construction module is used to construct an enterprise knowledge graph based on entity attribute structure and entity similarity association mechanism through entity recognition and relation extraction. The enterprise knowledge graph includes entity nodes, attribute nodes, relation edges, similarity edges, and dependent edges. Entity nodes storing entities are connected to the corresponding attribute nodes storing entity attributes through dependent edges. Entity nodes are connected to each other through relation edges and similarity edges. The similarity edges are marked with the similarity between entity nodes. The preliminary retrieval module is used to obtain the user's search content, calculate the matching similarity between the user's search content and each entity node, and search the enterprise knowledge graph to obtain at least one entity node as a candidate matching result. The retrieval optimization module is used to optimize candidate matching results based on similarity edges in the enterprise knowledge graph to obtain the final matching result; Extract the answer to the target question of the user's search content from the entity node corresponding to the final matching result as the search result.

[0043] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for retrieving enterprise consulting data based on graph-enhanced retrieval, characterized in that: Includes the following steps: Based on structured semantic units, word segmentation semantic features, context features, word segmentation position features, and entity similarity mechanisms are introduced to perform entity recognition and relation extraction, and an enterprise knowledge graph is constructed. The enterprise knowledge graph includes entity nodes, attribute nodes, relation edges, similarity edges, and dependent edges. Entity nodes storing entities are connected to the corresponding attribute nodes storing entity attributes through dependent edges. Entity nodes are connected to each other through relation edges and similarity edges. The similarity edges are marked with the similarity between entity nodes. The system retrieves user search content, calculates the matching similarity between user search content and each entity node, and performs a search in the enterprise knowledge graph to obtain at least one entity node as a candidate matching result. The candidate matching results are optimized based on the similarity edges in the enterprise knowledge graph to obtain the final matching result; Extract the answer to the target question of the user's search content from the entity node corresponding to the final matching result as the search result.

2. The enterprise information data retrieval method based on graph augmentation retrieval according to claim 1, characterized in that, The method for constructing the enterprise knowledge graph is as follows: Acquire enterprise data and perform standardization processing, which includes syntactic segmentation, uniform formatting, and deduplication. Each sentence is segmented using a Hidden Markov Model to obtain structured semantic units based on words. Based on the structured semantic units, entities and their corresponding entity attributes, as well as the relationships between entities and other entities, are extracted. Entities and their corresponding entity attributes are connected by dependent edges, and entities and entities with which they have a relationship are connected by relational edges, thus obtaining an initial enterprise knowledge graph. Based on the entity and its corresponding entity attributes, the identification parameters of each entity are obtained, and the similarity between entities is calculated based on the identification parameters. Entities with similarity greater than a preset threshold are connected by similarity edges, and the similarity is marked on the similarity edges to obtain the enterprise knowledge graph.

3. The enterprise information data retrieval method based on graph augmentation retrieval according to claim 2, characterized in that, The method for extracting the structured semantic units is as follows: Multiple entity description vectors are created and initialized to an empty set. The first element of the entity description vector is used to store the entity name, and each of the remaining elements stores a pair of attribute categories and corresponding attribute parameters. Establish an entity relationship matrix and initialize it to an empty set. The m-th row and n-th column of the entity relationship matrix represents the relationship between the m-th entity and the n-th entity. Initialize sentence number i=1, initialize entity description vector number j=1, and initialize element number of the j-th entity description vector. ; Perform the following operations until all sentences have been traversed: The classification label of each word in the i-th sentence is obtained through a word segmentation and classification model. The classification label includes entity, relationship between entities, attribute category and attribute parameter. Determine the number of entities in the i-th sentence. If the number of entities is one, check if the entity in the i-th sentence has already been stored in the entity description vector. If not, store the entity in the i-th sentence in the j-th entity description vector and update the value of j to j=j+1. Simultaneously, perform the following operations on the attribute categories and corresponding attribute parameters in the i-th sentence in sequence: [The text abruptly ends here, likely due to an incomplete translation or a missing section.] Each parameter, and update The value is If there are more than one entity, then determine whether each entity in the i-th sentence has been stored in the entity description vector. If not, store the entity of the i-th sentence in the j-th entity description vector and update the value of j to j=j+1. At the same time, store the entity relationships in the i-th sentence in the entity relationship matrix. Then, perform the following operations on the attribute categories and corresponding attribute parameters in the i-th sentence in sequence: store the entity in the j-th entity description vector... Each parameter, and update The value is ; Update the value of i to i = i + 1; After completing the traversal, the initial enterprise knowledge graph is constructed based on the entity description vector and the entity relationship matrix.

4. The enterprise information data retrieval method based on graph augmentation retrieval according to claim 3, characterized in that, The word segmentation and classification model includes: The input layer takes as input the word segment to be classified and all the word segments in the sentence. Feature extraction layer extracts semantic features from word segmentation. Contextual features and word segmentation position features Then, feature fusion is performed to obtain word segmentation fusion features. : ； ； ； ； ； ； in, It is a bidirectional long short-term memory neural network. To construct a character sequence function that is ordered character-by-character, This is a function that extracts the vector of each character from a character-based vector index table and concatenates them. For the word segmentation to be classified, The word preceding the segment to be classified. The next word in the segmentation to be classified. This represents the total number of words in the sentence containing the word to be classified. Number the segmented words to be classified sequentially within the sentence. and They are respectively and The semantic features of word segmentation, when the preceding or following word of the word to be classified does not exist, the corresponding... or Set the value to zero; The output layer outputs based on word segmentation fusion features. The probability of belonging to each of the stated category labels is calculated, and the label with the highest probability is selected as the classification label. The category tags mentioned above: ； ； ； in, and Represent Belongs to the kth type of category label The probability and rating, It is a natural exponential function. For intermediate parameters, , and For training weights, , and For training bias.

5. The enterprise information data retrieval method based on graph augmentation retrieval according to claim 4, characterized in that, The method for obtaining the identifier parameters of each entity and calculating the similarity between entities based on the identifier parameters is as follows: Calculate the identifier parameters based on the entity's entity description vector: ； in, For the first The entity description vector of the nth entity The attribute category number corresponding to each element. For the first The entity description vector of the nth entity The attribute parameters of the attribute category corresponding to each element. For the first The total number of elements in the entity description vector of each of the entities; Calculate the first The entity and the first The method for determining the similarity between entities is as follows: Obtain the first based on the identifier parameter The entity and the first The attribute categories that all entities possess form a common set of attribute categories; According to the The entity and the first The difference between the number of attribute categories and the corresponding attribute parameter values ​​for each entity is used to calculate the first... The entity and the first The similarity between entities : ； in, and The t-th attribute category in the set of common attribute categories is the t-th attribute category in the set of common attribute categories. The entity and the first The numerical values ​​of the corresponding attribute parameters in each entity, where T is the total number of elements in the common attribute category set.

6. The enterprise information data retrieval method based on graph augmentation retrieval according to claim 1, characterized in that, The method for searching in the enterprise knowledge graph is as follows: Semantic parsing is performed based on the user's search content to extract the search dimensions, which include entity name dimension, attribute dimension, relationship dimension, and search target. Taking each entity node as a unit, the following operations are performed on each entity node: Matching searches are performed on the enterprise knowledge graph according to the retrieval dimensions. Cosine similarity is used to obtain entity similarity scores, multiple attribute similarity scores, and multiple relation similarity scores. The entity similarity score is the similarity between the entity stored in the entity node and the entity name dimension. The attribute similarity score is the maximum value of the similarity between the attribute dimension involved in the user's search content and all related attribute nodes. Related attribute nodes are the attribute nodes connected to the entity node through the subordinate edge. The relation similarity is the similarity between the relation dimension involved in the user's search content and all relation edges of the entity node. The matching similarity of the entity node is obtained by weighted summation of relation similarity, node similarity, and attribute similarity; After obtaining the matching similarity of all the entity nodes, the entity nodes with matching similarity greater than a preset threshold are selected as the candidate matching results.

7. The enterprise information data retrieval method based on graph augmentation retrieval according to claim 1, characterized in that, The method for optimizing candidate matching results based on similarity edges in the enterprise knowledge graph involves performing the following operations for each candidate matching result: Obtain all entity nodes connected to the entity nodes corresponding to the candidate matching results through similarity edges, and use them as optimization nodes; Obtain the matching similarity between the user's search content and each optimization node, and calculate the optimization parameters of the candidate matching results; The final score of the candidate matching result is calculated based on the optimization parameters and the matching similarity. After obtaining the final scores of all the candidate matching results, the candidate matching results with a final score greater than a preset threshold are selected as the final matching results.

8. The enterprise information data retrieval method based on graph augmentation retrieval according to claim 7, characterized in that, The method for calculating the optimization parameters of the candidate matching results is as follows: ； in, For the first The optimization parameters for the candidate matching results. For the first The entity node number corresponding to each of the candidate matching results. This represents the entity node number corresponding to the d-th optimization node. For the first The first entity node and the first The similarity score marked on the similar edges of each entity node. For the user's search content and the first The matching similarity of each entity node.

9. The enterprise information data retrieval method based on graph augmentation retrieval according to claim 8, characterized in that, The method for calculating the final score of the candidate matching results is as follows: ； in, For the first The final score of each of the candidate matching results. It is a natural exponential function. For the user's search content and the first The matching similarity of each entity node.

10. A business information data retrieval system based on graph-enhanced retrieval, applied to the business information data retrieval method based on graph-enhanced retrieval as described in any one of claims 1-9, characterized in that, include: The enterprise knowledge graph construction module is used to construct an enterprise knowledge graph based on entity attribute structure and entity similarity association mechanism through entity recognition and relation extraction. The enterprise knowledge graph includes entity nodes, attribute nodes, relation edges, similarity edges, and dependent edges. Entity nodes storing entities are connected to the corresponding attribute nodes storing entity attributes through dependent edges. Entity nodes are connected to each other through relation edges and similarity edges. The similarity edges are marked with the similarity between entity nodes. The preliminary retrieval module is used to obtain the user's search content, calculate the matching similarity between the user's search content and each entity node, and search the enterprise knowledge graph to obtain at least one entity node as a candidate matching result. The retrieval optimization module is used to optimize candidate matching results based on similarity edges in the enterprise knowledge graph, obtain the final matching result, and extract the answer corresponding to the target question of the user's search content from the entity node corresponding to the final matching result as the retrieval result.

Images