Enterprise management-oriented knowledge retrieval method and system based on knowledge graph

Abstract

The present application relates to the field of data processing, in order to improve the accuracy of question and answer, provide the knowledge retrieval method and system based on knowledge graph for enterprise management, through mining the synonymy relationship, the hyponym relationship, the scene correlation relationship and so on between words, construct rich word relationship graph, based on the word relationship graph, the low confidence word combination in the question is dynamically rewritten, the accuracy of question understanding is improved;According to the user role, the retrieval priority of each knowledge base is determined, the knowledge interference of cross domain and cross department is effectively avoided, and the retrieval efficiency is higher;Combined with user portrait, personalized knowledge is generated, the pertinence and practicality of the answer are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing, specifically to a knowledge retrieval method and system based on knowledge graphs for enterprise management. Background Technology

[0002] With the rapid development of large-scale modeling technology, knowledge retrieval-based question answering systems have begun to flourish. Existing question answering systems mainly rely on keyword matching and simple semantic analysis, lacking a deep understanding of the question's intent and consideration of personalized preferences, resulting in low accuracy in responses. Summary of the Invention

[0003] To improve the accuracy of question answering, this invention provides a knowledge retrieval method and system based on knowledge graphs for enterprise management.

[0004] The technical solution adopted by the present invention to solve the above problems is:

[0005] Knowledge graph-based knowledge retrieval methods for enterprise management include:

[0006] Step 1: Build knowledge bases for multiple professional fields based on the enterprise's management needs. Upload the original data of relevant departments and fields to each knowledge base and slice the original data.

[0007] Step 2: Based on the enterprise's historical question-and-answer data and domain dictionary, mine the relationships between words to construct a word relationship graph;

[0008] Step 3: Determine the knowledge base retrieval priority based on user roles;

[0009] Step 4: Based on the prompt word template, obtain the keyword combination in the input question. Use the extracted keyword combination as the query statement to first search the knowledge base with high priority. If the confidence of the search result is lower than the first preset threshold, then rewrite the keyword combination based on the word relationship graph. If the confidence of the search result is higher than the first preset threshold but lower than the second preset threshold, then perform a global knowledge base search.

[0010] Step 5: Generate knowledge-based answers based on the search results.

[0011] Furthermore, the knowledge base covers areas including: finance, human resources, procurement, marketing, general administration, quality, occupational health, safety and environmental management, strategic management, information management, internal supervision and management, technology management and corporate governance.

[0012] Furthermore, the relationships between words include: synonymy, hierarchical relationship, and contextual relationship.

[0013] Furthermore, the relationships between words are obtained through the following methods: manual annotation of the company's historical Q&A data; and direct extraction of word relationships based on a domain dictionary, which includes: a glossary of professional terms, a table of abbreviations, a product terminology database, and standardized data on department / position names.

[0014] Furthermore, the prompt word template consists of three parts: task description, output format constraints, and example demonstration. The task description explicitly requires the large model to semantically decompose the user-input question and identify the core entities, operations, and limiting conditions. The output format constraints stipulate that the large model should return the keyword combination and its part-of-speech tag in JSON format. The example demonstration provides decomposition examples of 2-3 typical questions.

[0015] Furthermore, step 5 specifically includes:

[0016] Step 51: Build user profiles and match the optimal answer format from a predefined template library based on the user profiles;

[0017] Step 52: Generate knowledge answers using a multimodal fusion algorithm based on the search results and answer format.

[0018] Furthermore, in step 52, before employing the multimodal fusion algorithm, reasoning is performed on the retrieved knowledge fragments based on the knowledge reasoning model.

[0019] Furthermore, it also includes step 53, which involves multi-dimensional evaluation and optimization of the generated knowledge answer until it meets the requirements and then outputs the final answer.

[0020] A knowledge retrieval system based on knowledge graphs for enterprise management includes:

[0021] The knowledge base module builds multiple professional knowledge bases based on enterprise management needs. Each knowledge base uploads the original data from relevant departments and fields and performs data slicing on the original data.

[0022] The word relationship graph construction module, based on enterprise historical question and answer data and domain dictionary, mines the relationships between words to construct a word relationship graph;

[0023] The retrieval priority determination module determines the knowledge base retrieval priority based on the user's role.

[0024] The knowledge retrieval module obtains keyword combinations from the input question based on the prompt word template. The extracted keyword combinations are used as query statements to first search the knowledge base with high priority. If the confidence of the search results is lower than the first preset threshold, the keyword combinations are rewritten based on the word relationship graph. If the confidence of the search results is higher than the first preset threshold but lower than the second preset threshold, a global knowledge base search is performed.

[0025] The knowledge answer generation module generates knowledge answers based on the search results.

[0026] Furthermore, the knowledge answer generation module includes:

[0027] The generator produces answers based on the search results;

[0028] The discriminator evaluates the answers generated by the generator from multiple dimensions.

[0029] The advantages of this invention compared to existing technologies are as follows: by mining synonym relationships, hierarchical relationships, and scene association relationships between words, a rich word relationship graph is constructed. Based on the word relationship graph, low-confidence word combinations in the question are dynamically rewritten, improving the accuracy of question understanding; the retrieval priority of each knowledge base is determined according to the user role, effectively avoiding cross-domain and cross-departmental knowledge interference, resulting in higher retrieval efficiency; and personalized knowledge generation is combined with user profiles, improving the relevance and practicality of the answers. Attached Figure Description

[0030] Figure 1 This is a flowchart of a knowledge retrieval method based on knowledge graphs for enterprise management. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] In terms of knowledge base construction, this invention builds knowledge bases in multiple professional fields based on enterprise management needs. During knowledge retrieval, the retrieval priority of each knowledge base is determined by user roles, effectively avoiding cross-domain and cross-departmental knowledge interference and improving retrieval efficiency. Regarding question intent understanding, a rich word relationship graph is constructed by mining synonymous, hierarchical, and scenario-related relationships between words. Based on this graph, low-confidence word combinations in the question are dynamically rewritten, improving the accuracy of question understanding. In terms of knowledge generation, personalized knowledge generation is performed based on user profiles, improving the relevance and practicality of the answers. Through these three improvements, the accuracy and relevance of question answering are effectively enhanced.

[0033] like Figure 1 As shown, a knowledge graph-based knowledge retrieval method for enterprise management includes:

[0034] Step 1: Build knowledge bases for multiple professional fields based on the enterprise's management needs. Upload the original data of relevant departments and fields to each knowledge base and slice the original data.

[0035] The constructed knowledge base includes, but is not limited to, finance, human resources, procurement, marketing, general administration, quality, occupational health, safety and environmental management, strategic management, information management, internal supervision and management, technology management, and corporate governance. Each knowledge base uploads structured or semi-structured data such as relevant departmental and field policy documents, operating procedures, and historical cases as the foundation for knowledge retrieval. Uploaded documents are sliced ​​using an embedding model (such as the BGE model), dividing long documents into multiple short text fragments, each containing a relatively complete information unit, facilitating subsequent retrieval and matching. The sliced ​​document fragments are stored in their respective knowledge bases, while retaining the association information of their original documents so that the complete document content can be traced during retrieval.

[0036] Step 2: Based on the enterprise's historical question and answer data and domain dictionary, mine the relationships between words to construct a word relationship graph.

[0037] To address situations where qwen3 fails to understand questions, leading to inaccurate answers or the inability to retrieve results using rag, a rich word relationship network is formed by mining synonyms, hierarchical relationships, and contextual associations between words based on historical enterprise question-and-answer data and domain dictionaries. For example, by mining the contextual association between "time" and "situation," when a question contains "General Manager + Time + Communication," it can be rewritten as "General Manager + Situation + Communication" to improve the accuracy of question comprehension.

[0038] The specific steps for constructing a word relationship graph are as follows:

[0039] Word relationships can be obtained in two ways:

[0040] First, we manually annotated the test history of the problem to mine word relationships in different contexts, created a Neo4j database, and set three types of word relationships: synonym relationship, hyponymy relationship, and scene association relationship.

[0041] Second, word relationships are directly extracted based on the domain dictionary, which includes: a glossary of professional terms, a table of abbreviations, a product terminology database, and standardized data on department / position names.

[0042] When constructing word relationships, a property graph model is used to store the word relationship graph. This is achieved through four steps: node construction, edge construction, graph model storage, and association adaptation. This ensures that the graph is scalable, searchable, and compatible with subsequent knowledge retrieval processes. The specific implementation is as follows:

[0043] The first step is the construction and storage of nodes (Vertex): Nodes are constructed based on enterprise domain terms, and the node label is uniformly set to "Word"; the attributes of each node are stored in a structured manner according to the enterprise management knowledge retrieval requirements, as follows:

[0044] name: Stores the word text in string format, taken from manually annotated historical Q&A data of the enterprise and domain dictionary, as the unique core identifier of the node to avoid duplicate storage; Example: The word text "performance appraisal" is stored as the name attribute value, corresponding to a unique node.

[0045] pos: Part of speech (optional), using an enumeration format, automatically tagged by an NLP part-of-speech tagging tool, stored synchronously with the word text, used for part-of-speech matching in semantic decomposition and rewriting of keyword combinations; Example: "performance appraisal" is tagged as a noun, and "approval" is tagged as a verb.

[0046] frequency: Stores the frequency of a word in the relevant corpus in integer format. It is obtained through statistics and is used to adjust the weight of word relationships and assist in judging the retrieval priority. Example: "performance appraisal" appears 120 times in the company's historical Q&A data, so the frequency attribute value is set to 120.

[0047] domain: Stores the domain tags of words in string format. These tags are determined through manual annotation and automatic matching to link them with the corresponding professional domain knowledge base and support the determination of search priority. Example: Set the domain tag for "performance appraisal" to "human resources" and link it to the human resources domain knowledge base.

[0048] Embedding: Word vectors (optional, used for semantic retrieval), generated by encoding word text through an embedding model, stored in floating-point array format, supporting keyword combination rewriting and retrieval matching; Example: "performance appraisal" is encoded to generate [0.12, 0.35, ..., 0.78] (fixed-dimensional floating-point array).

[0049] The second step is the construction and storage of edges: edges are constructed based on the discovered relationships between words. Each edge corresponds to a relationship between two word nodes. The specific storage method is as follows:

[0050] Relationship types: Stored in an enumeration format, divided into SYNONYM_OF (synonym relationship), HYPERNYM_OF (hyper-hypernym relationship, hypernym node points to hyponym node), and SCENE_RELATED (scenario association relationship), corresponding to the three types of word association relationships mined in this invention; Example: "performance appraisal" and "performance evaluation" are SYNONYM_OF relationships, "performance indicators" and "KPI" are HYPERNYM_OF relationships, and "approval" and "process" are SCENE_RELATED relationships.

[0051] weight: Stored in floating-point format within the range of [0,1]. Calculated based on the closeness of word association. An initial value is set according to the source of the relationship. It can be dynamically updated based on user feedback and new data. It is used for candidate screening of keyword combination rewriting. Example: The association weight between "performance appraisal" and "performance evaluation" is set to 0.9, and the association weight between "approval" and "process" is set to 0.7.

[0052] `source`: Stores the relationship acquisition channel in string format for easy relationship tracing and manual review; Example: Set the relationship source of "performance appraisal" and "performance evaluation" to "dictionary", and set the relationship source of "approval" and "process" to "manual annotation".

[0053] last_updated: Stores the creation and update time of relationships in a time format to support iterative optimization of word relationship graphs; Example: The relationship creation time can be a certain year, month, and day, and after subsequent updates, last_updated is set to a certain date after that time, with the specific hour, minute, and second set as needed or ignored.

[0054] confidence: Stored in floating-point format within the range of [0,1], used for manual review and identification. Initial values ​​are set according to the relationship acquisition method to facilitate the selection of high-confidence relationships for knowledge retrieval. Example: The confidence of the manually labeled relationship between "performance appraisal" and "performance evaluation" is set to 0.95, and the confidence of the automatically extracted relationship is set to 0.7.

[0055] The third step is graph model storage: the attribute graph model is stored using the Neo4j database. The node and edge data are formatted according to the database specifications, and node and edge indexes are created to improve retrieval efficiency and adapt to the subsequent keyword rewriting process.

[0056] The fourth step is association and adaptation: the word relationship graph is associated with the knowledge base module and knowledge retrieval module of this invention to ensure that the node domain attributes correspond to the knowledge bases of each subject. The relationship types and weights of the edges support the rewriting of keyword combinations, while providing semantic support for knowledge reasoning and personalized answer generation, thus achieving integration with the entire retrieval system.

[0057] The word relationship graph constructed in this invention is scalable and can be continuously enriched and improved with the addition of new data. During the retrieval process, the word relationship graph is used for query expansion and semantic understanding, effectively improving the recall and precision of the retrieval.

[0058] Step 3: Determine the knowledge base retrieval priority based on user roles.

[0059] During retrieval, priority is given to matching knowledge bases that are strongly related to the user's role (e.g., the human resources department should prioritize matching the human resources knowledge base, and the leadership should prioritize matching the strategic management and corporate governance knowledge base) to avoid cross-domain and cross-departmental knowledge interference.

[0060] Step 4: Obtain keyword combinations from the input question based on the prompt word template. Use the extracted keyword combinations as the query statement to first search the knowledge base with high priority. If the confidence of the search results is lower than the first preset threshold, the keyword combinations are rewritten based on the word relationship graph. If the confidence of the search results is higher than the first preset threshold but lower than the second preset threshold, a global knowledge base search is performed.

[0061] First, a prompt word template is constructed for a large-scale model. This template comprises three parts: a task description, output format constraints, and example demonstrations. The task description explicitly requires the large-scale model to semantically decompose the user-input question, identifying the core entities, actions, and limiting conditions. The output format constraints stipulate that the large-scale model must return keyword combinations and their part-of-speech tags in JSON format. The example demonstrations provide 2-3 typical question decomposition examples to guide the large-scale model in understanding the task requirements. By inputting this prompt word template and the user question into the large-scale model, a structured output of keyword combinations is obtained. .

[0062] Subsequently, the extracted keywords are combined into a query statement and input into a high-priority knowledge base for vector retrieval to obtain a set of candidate documents. A cross-encoder is used to re-rank the candidate documents, and a relevance score for each document to the query is calculated. The confidence score C(K) of the keyword combination is calculated based on the relevance scores after reordering. Specifically, the average relevance score of the top k documents is used. This confidence level reflects the degree of semantic matching between the current keyword combination and the normative documents in the knowledge base.

[0063] The calculated confidence score C(K) is compared with the first preset threshold θ_low. If C(K) < θ_low, the current expression is deemed non-standard and needs to be rewritten based on a pre-constructed word relationship graph. During the rewriting process, the graph is first queried to obtain synonyms and scene-related words for the keywords. Then, candidate rewriting combinations are generated through synonym replacement and scene expansion. Finally, the rewritten result that maximizes the improvement in re-ranking score is selected as the optimized query statement. Through the rewriting mechanism, a refined understanding and standardized processing of user questions is achieved, effectively improving the accuracy of subsequent searches. If C(K) ≥ θ_low, the current keyword combination is deemed standard and can be directly used for subsequent searches. C(K) is judged again: if C(K) is less than the second preset threshold, it is determined that no suitable answer can be matched in the current knowledge base, and a global knowledge base search is automatically triggered to balance precision and recall. When generating an answer, only the global knowledge base search results can be used, or the results of the two searches can be combined. The first preset threshold is less than the second preset threshold, and the specific values ​​of the two can be selected according to actual needs, without restriction.

[0064] Step 5: Generate knowledge-based answers based on the search results.

[0065] In this embodiment, to make the generated knowledge answers more targeted, user profiles are first created based on the user's department, job level, and past questions. The user profile includes multiple dimensions such as basic user information, interests, and historical behavior, used to depict the user's personalized characteristics. Then, various answer templates are defined using prompts. For example, technical personnel can describe the relevant data and operational steps of the question in detail, while managers can focus on describing the advantages and disadvantages related to the conclusion of the question. Personalized knowledge generation is achieved by combining different prompts based on the user profile and preferences.

[0066] In knowledge generation, the generator incorporates multimodal data, such as images, charts, and audio. Through multimodal fusion technology, it unifies and integrates information from different modalities, enabling the generator to produce richer and more vivid knowledge-based answers. For example, for questions involving product descriptions, the generator can combine product images and text descriptions to generate a more intuitive answer.

[0067] Furthermore, this embodiment incorporates a knowledge reasoning module within the generator, utilizing a pre-trained knowledge reasoning model (such as CommonsenseQA) to reason and expand upon the retrieved knowledge fragments. Through reasoning, the generator can uncover the implicit information within the knowledge fragments, generating more in-depth and logical answers. For example, for the question "How to improve project efficiency," the generator can deduce a series of specific measures and suggestions based on retrieved project management knowledge.

[0068] The discriminator not only receives knowledge-based answers generated by the generator and real user feedback as input, but also introduces multi-dimensional evaluation metrics, such as the accuracy, completeness, fluency, and readability of the answers. Through a comprehensive evaluation of these metrics, the discriminator can more fully determine whether the generated knowledge-based answers are authentic and meet user needs. For example, for the generated answers, the discriminator can check whether there are factual errors or whether they cover all the key points of the user's question.

[0069] The discriminator can also perform in-depth analysis of real user feedback to uncover the underlying reasons for user satisfaction with answers. For example, by analyzing user ratings and comments on answers, it can understand user preferences regarding answer content, format, and language style, and feed this information back to the generator to guide it in further optimizing its generation strategy.

[0070] During the training of the generator and discriminator, this embodiment introduces a reinforcement learning mechanism. The generator is treated as an intelligent agent, and the feedback from the discriminator is used as a reward signal. The generator's policy is optimized through reinforcement learning algorithms (such as Deep Q-Network (DQN), policy gradient algorithm, etc.). The generator continuously adjusts its generation method based on the reward signal to obtain higher rewards, that is, to generate knowledge answers that better meet the user's needs.

[0071] To improve training efficiency and stability, this embodiment employs a distributed adversarial training framework. The generator and discriminator are deployed across multiple computing nodes, accelerating the training process through parallel computation. Simultaneously, the fault tolerance mechanism of the distributed framework ensures that training can continue even if some nodes fail, thus improving training reliability.

[0072] Furthermore, this embodiment also performs post-processing on the generated knowledge answers, including grammar checking, spelling correction, and format optimization. This ensures that the answers are accurate in language and format, improving their readability and professionalism. Simultaneously, relevant summaries or key points are automatically generated based on the answer content, facilitating users' quick browsing and understanding of the core content.

[0073] During interaction, diverse interaction methods are provided, such as voice interaction and graphical interaction. The Wisper and Cosyvoice models are introduced to realize voice interaction, which meets the usage habits and needs of different users.

[0074] During the interaction, the system dynamically adjusts the content and manner of the response based on real-time user feedback. For example, if a user has a question about a part of the response, the system can provide a more detailed explanation; if the user is not satisfied with the response, the system can regenerate the response or offer other relevant knowledge points for the user to choose from.

[0075] In addition, a user feedback loop mechanism is established to encourage users to evaluate and provide suggestions on knowledge answers, and the entire knowledge retrieval optimization system is continuously optimized based on user feedback.

[0076] Correspondingly, this invention also provides a knowledge retrieval system based on knowledge graphs for enterprise management, including:

[0077] The knowledge base module builds multiple professional knowledge bases based on enterprise management needs. Each knowledge base uploads the original data from relevant departments and fields and performs data slicing on the original data.

[0078] The word relationship graph construction module, based on enterprise historical question and answer data and domain dictionary, mines the relationships between words to construct a word relationship graph;

[0079] The retrieval priority determination module determines the knowledge base retrieval priority based on the user's role.

[0080] The knowledge retrieval module obtains keyword combinations from the input question based on the prompt word template. The extracted keyword combinations are used as query statements to first search the knowledge base with high priority. If the confidence of the search results is lower than the first preset threshold, the keyword combinations are rewritten based on the word relationship graph. If the confidence of the search results is higher than the first preset threshold but lower than the second preset threshold, a global knowledge base search is performed.

[0081] The knowledge answer generation module generates knowledge answers based on the search results.

[0082] Specifically, the knowledge answer generation module includes: a generator, which generates answers based on search results; and a discriminator, which evaluates the answers generated by the generator from multiple dimensions.

Claims

1. A knowledge retrieval method based on knowledge graphs for enterprise management, characterized in that: include: Step 1: Build knowledge bases for multiple professional fields based on the enterprise's management needs. Upload the original data of relevant departments and fields to each knowledge base and slice the original data. Step 2: Based on the enterprise's historical question-and-answer data and domain dictionary, mine the relationships between words to construct a word relationship graph; Step 3: Determine the knowledge base retrieval priority based on user roles; Step 4: Obtain keyword combinations from the input question based on the prompt word template. Use the extracted keyword combinations as the query statement to first search the knowledge base with high priority. If the confidence of the search results is lower than the first preset threshold, rewrite the keyword combinations based on the word relationship graph. If the confidence level of the search result is higher than the first preset threshold but lower than the second preset threshold, then a global knowledge base search is performed. Step 5: Generate knowledge-based answers based on the search results.

2. The knowledge retrieval method based on knowledge graphs for enterprise management according to claim 1, characterized in that, The knowledge base covers areas including: finance, human resources, procurement, marketing, general administration, quality, occupational health, safety and environmental management, strategic management, information management, internal supervision and management, technology management and corporate governance.

3. The knowledge retrieval method based on knowledge graphs for enterprise management according to claim 1, characterized in that, The relationships between words include: synonymy, hierarchical relationship, and contextual relationship.

4. The knowledge retrieval method based on knowledge graphs for enterprise management according to claim 3, characterized in that, The relationships between words were obtained through manual annotation of the company's historical Q&A data; Based on the domain dictionary, word relationships are directly extracted. The domain dictionary includes: a glossary of professional terms, a table of abbreviations, a product terminology database, and standardized data on department / job titles.

5. The knowledge retrieval method based on knowledge graphs for enterprise management according to claim 3, characterized in that, The prompt word template consists of three parts: task description, output format constraints, and example demonstration. The task description clearly requires the large model to semantically decompose the user input question and identify the core entities, operations, and limiting conditions. The output format constraints stipulate that the large model should return the combination of keywords and their part-of-speech tags in JSON format. The example demonstration provides decomposition examples of 2-3 typical questions.

6. The knowledge retrieval method based on knowledge graphs for enterprise management according to claim 3, characterized in that, Step 5 specifically involves: Step 51: Build user profiles and match the optimal answer format from a predefined template library based on the user profiles; Step 52: Generate knowledge answers using a multimodal fusion algorithm based on the search results and answer format.

7. The knowledge retrieval method based on knowledge graphs for enterprise management according to claim 6, characterized in that, In step 52, before employing the multimodal fusion algorithm, the retrieved knowledge fragments are inferred based on a knowledge reasoning model.

8. The knowledge retrieval method based on knowledge graphs for enterprise management according to claim 6, characterized in that, It also includes step 53, which involves multi-dimensional evaluation and optimization of the generated knowledge answer until it meets the requirements and then outputs the final answer.

9. A knowledge retrieval system based on knowledge graphs for enterprise management, characterized in that: include: The knowledge base module builds multiple professional knowledge bases based on enterprise management needs. Each knowledge base uploads the original data from relevant departments and fields and performs data slicing on the original data. The word relationship graph construction module, based on enterprise historical question and answer data and domain dictionary, mines the relationships between words to construct a word relationship graph; The retrieval priority determination module determines the knowledge base retrieval priority based on the user's role. The knowledge retrieval module obtains keyword combinations from the input question based on the prompt word template. The extracted keyword combinations are used as query statements to first search the knowledge base with high priority. If the confidence of the search results is lower than the first preset threshold, the keyword combinations are rewritten based on the word relationship graph. If the confidence level of the search result is higher than the first preset threshold but lower than the second preset threshold, then a global knowledge base search is performed. The knowledge answer generation module generates knowledge answers based on the search results.

10. The knowledge retrieval system based on knowledge graphs for enterprise management according to claim 9, characterized in that, The knowledge answer generation module includes: The generator produces answers based on the search results; The discriminator evaluates the answers generated by the generator from multiple dimensions.

Images