A medical knowledge graph construction method and device

Abstract

Embodiments of the present specification provide a medical knowledge graph construction method and device. In the method, the pages in the electronic file of the first medical literature are input into a large model, entity data of entities in the pages are recognized, the recognized entity data is merged into the constructed knowledge graph through semantic similarity matching, and the merging includes taking the entities as nodes, taking knowledge points as attributes of the nodes, and constructing untyped connection edges between the entities and their associated entities. When the entity data corresponding to all pages of the electronic file is merged into the constructed knowledge graph, a medical knowledge graph of the first medical literature is obtained. The entity data contains entity names, entity types, associated entities, and knowledge points related to the entities. The knowledge point text is the original sentence or slightly modified text of the page text content, which contains complete and rich association relationship information between the associated entities. The above construction process needs to protect the privacy data.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of artificial intelligence medical application technology, and in particular to a method and apparatus for constructing a medical knowledge graph. Background Technology

[0002] With the rapid development of Artificial Intelligence (AI) technology, its application in the medical field is becoming increasingly widespread. However, the inherent illusion problem of AI systems prevents them from being directly used for medical guidance. AI illusion refers to the phenomenon where AI systems generate seemingly reasonable but actually erroneous or false information, which is an inherent limitation of current AI technology. In medical applications, AI illusion may lead to incorrect diagnostic recommendations or treatment plans. To ensure the reliability and safety of AI medical applications, a comprehensive evaluation system is needed to verify the accuracy and credibility of AI systems. Medical knowledge graphs are data models that use graph structures to represent knowledge, consisting of entities (nodes) and relationships (edges), capable of describing concepts, entities, and their interrelationships in the medical knowledge world. Evaluation sets can be quickly constructed based on medical knowledge graphs, thus meeting the needs for large-scale and comprehensive evaluation of AI medical applications. Medical knowledge graphs contain private data, which needs to be protected. Whether a medical knowledge graph can accurately reflect the semantic knowledge in medical literature significantly affects the quality of the evaluation set, thereby impacting the performance of AI medical applications.

[0003] Therefore, we hope to find an improved solution that can construct a knowledge graph that can more accurately describe the medical knowledge in the original medical literature. Summary of the Invention

[0004] This specification describes one or more embodiments of a method and apparatus for constructing a medical knowledge graph, in order to build a knowledge graph that can more accurately reproduce medical knowledge. The specific technical solution is as follows.

[0005] Firstly, the embodiments provide a method for constructing a medical knowledge graph, including:

[0006] Several pages from the electronic file of the first medical document are input into a large model. The large model identifies entity data of several entities within the pages. Each page contains text content, and each entity data includes an entity name, entity type, related entities, and related knowledge points. The entity type includes only disease entity types and treatment method entity types. The knowledge points include first knowledge point text, which is determined based on the corresponding text content and is used to describe the relationship information between the entity and related entities. The relationship information includes the type of relationship and at least one of the following: conditions for establishment, reasons, and methods of establishment.

[0007] The identified entities are semantically similar to several entities in the constructed knowledge graph.

[0008] Based on the semantic similarity matching results, the entity data of the identified entities are merged into the constructed knowledge graph; wherein, the merging includes using entities as nodes and knowledge points as attributes of nodes to construct typeless connection edges between entities and their associated entities.

[0009] When the entity data corresponding to all pages of the electronic document are merged into the constructed knowledge graph, the constructed knowledge graph is used as the medical knowledge graph of the first medical document.

[0010] In one implementation, the step of inputting several pages into a large model includes:

[0011] Several pages and the constructed knowledge graph are input into the large model. Using the large model and the entities in the constructed knowledge graph as references, entity data of several entities in the several pages are identified.

[0012] In one implementation, the step of inputting several pages and the constructed knowledge graph into the large model includes:

[0013] Input the entity names and relationships between entities from several pages of the constructed knowledge graph into the large model.

[0014] In one implementation, the step of inputting several pages into a large model includes:

[0015] The current identification unit contains several pages, which are then input into the large model; wherein, all pages contained in the electronic document are divided into multiple identification units;

[0016] After merging entity data of several entities identified from the current identification unit into the constructed knowledge graph, the method further includes:

[0017] Update the next recognition unit to the current recognition unit, and return to the step of inputting the several pages contained in the current recognition unit into the large model.

[0018] In one implementation, the step of performing semantic similarity matching between the identified entities and several entities in the constructed knowledge graph includes:

[0019] The entity names of the identified entities are semantically similar to the entity names of the entities in the constructed knowledge graph.

[0020] In one implementation, the step of merging the entity data of the identified entities into the constructed knowledge graph includes:

[0021] When it is determined that the first entity among the identified entities is a similar entity to the second entity in the constructed knowledge graph, the associated entities and knowledge points in the entity data of the first entity are merged into the node data of the second entity.

[0022] When it is determined that the third entity among the identified entities is not a similar entity to any entity in the constructed knowledge graph, the third entity is directly added as a new node to the constructed knowledge graph.

[0023] In one implementation, the step of directly adding the third entity as a new node to the constructed knowledge graph includes:

[0024] Based on the knowledge point text and corresponding knowledge point number stored in the shared knowledge point pool, determine the knowledge point number corresponding to the knowledge point text of the third entity, and determine that the knowledge point text of the third entity is stored in the shared knowledge point pool;

[0025] The identified knowledge point number is added to the attributes of the new node corresponding to the third entity in the constructed knowledge graph.

[0026] In one implementation, the steps of determining the knowledge point number corresponding to the knowledge point text of the third entity and determining that the knowledge point text of the third entity is stored in the shared knowledge point pool include:

[0027] When there is a knowledge point text in the shared knowledge point pool that is the same as the knowledge point text of the third entity, the number of the knowledge point text in the shared knowledge point pool is used as the corresponding number of the knowledge point text of the third entity.

[0028] When there is no knowledge point text in the shared knowledge point pool that is identical to the knowledge point text of the third entity, a number for the knowledge point text of the third entity is generated based on the existing knowledge point numbers in the shared knowledge point pool, and the number and the knowledge point text of the third entity are stored in the shared knowledge point pool accordingly.

[0029] In one implementation, when obtaining the medical knowledge graph corresponding to the first medical document, the method further includes:

[0030] Based on the query subgraph to be queried, a query is performed in the medical knowledge graph to obtain the query results;

[0031] Based on the knowledge points corresponding to the entities and their associated entities in the query results, the required questions and / or answers are generated.

[0032] In one implementation, when obtaining the medical knowledge graph of the first medical document, the method further includes:

[0033] The medical knowledge graph is stored in a different storage space than the medical knowledge graph of the second medical document;

[0034] When a search is required, a first result is obtained based on the medical knowledge graph of the first medical document, and a second result is obtained based on the medical knowledge graph of the second medical document;

[0035] The first result, the second result, and their corresponding source information are used as search results; wherein, the source information includes the name of the corresponding medical literature.

[0036] Secondly, the embodiments provide an apparatus for constructing a medical knowledge graph, comprising:

[0037] The entity recognition module is configured to input several pages from the electronic file of the first medical document into a large model, and through the large model, identify entity data of several entities in the pages; wherein, each page contains text content, and any entity data includes entity name, entity type, related entities that are associated with the entity, and knowledge points related to the entity; the entity type only includes disease entity type and treatment method entity type, and the knowledge points include first knowledge point text, which is determined based on the corresponding text content and is used to describe the relationship information between the entity and related entities, the relationship information including the type of relationship, and at least one of the following: establishment conditions, reasons, and establishment methods;

[0038] The data matching module is configured to perform semantic similarity matching between the identified entities and several entities in the constructed knowledge graph;

[0039] The data merging module is configured to merge the entity data of the identified entities into the constructed knowledge graph based on the semantic similarity matching results; wherein, the merging includes constructing typeless connection edges between entities and their associated entities, using entities as nodes and knowledge points as attributes of the nodes.

[0040] The knowledge graph construction module is configured to use the constructed knowledge graph as the medical knowledge graph of the first medical document when the entity data corresponding to all pages of the electronic document are merged into the constructed knowledge graph.

[0041] Thirdly, the embodiments provide a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described in any one of the first aspects.

[0042] Fourthly, an embodiment provides a computing device including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method described in any one of the first aspects.

[0043] The method and apparatus provided in the embodiments of this specification identify entities and entity data in medical literature through a large model. The entity data contains only two entity types, and the type between entities is represented by the first knowledge point text. When constructing a knowledge graph based on the identified entity data, knowledge points are used as attributes of nodes. The first knowledge point text is the original text or slightly modified text of the page content, which contains complete and rich information on the relationships between related entities (including the type, conditions, reasons, and methods of establishing the relationship). This allows the constructed medical knowledge graph to largely contain the complete knowledge in the medical literature, and such a medical knowledge graph can more accurately restore the original meaning of medical knowledge. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0045] Figure 1 This is a schematic diagram illustrating an implementation scenario of one embodiment disclosed in this application;

[0046] Figure 2 A flowchart illustrating a method for constructing a medical knowledge graph, provided as an example;

[0047] Figure 3 A schematic diagram illustrating the principle of an existing knowledge graph and shared knowledge point pool;

[0048] Figure 4 A schematic diagram of an overall architecture provided for an embodiment;

[0049] Figure 5 This is a schematic block diagram of a medical knowledge graph construction apparatus provided for an embodiment. Detailed Implementation

[0050] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0051] Figure 1 This is a schematic diagram illustrating an implementation scenario of one embodiment disclosed in this application. It includes a large model device, a computing device, and a storage device for storing a medical knowledge graph. The large model device deploys a large model. The computing device can acquire the electronic file of the medical document and send the electronic file to the large model device. The large model is used to perform entity recognition on the text content of the input medical document, obtaining entity data for multiple entities, and sending the entity data to the computing device. The computing device constructs a knowledge graph based on the entity data and stores the constructed medical knowledge graph in the graph database of the storage device. The medical knowledge graph can be used to generate question-and-answer pairs (question text and corresponding answer text) about medical knowledge, and can also be used for graph retrieval enhancement generation.

[0052] The entity data identified by the large model includes entity names, entity types, related entities, and knowledge points. Knowledge points are derived from text content and contain complete and rich information about the relationships between related entities. Entity types include only two categories: disease entity types and treatment method entity types. When constructing a knowledge graph based on entity data, the computing device uses entities as nodes and knowledge points as attributes of the nodes, constructing typeless connections between related entities. That is, the knowledge graph schema is a dual-entity + binary tuple design with only two entity types and typeless connections. This schema design reduces the complexity of the knowledge graph, and the complete and rich relationship information contained in the knowledge points can more accurately reproduce the original meaning of knowledge in medical literature, thus enabling the knowledge graph to provide more accurate medical information.

[0053] In this specification, the knowledge graph refers to a medical knowledge graph. A knowledge graph is a data model that uses a graph structure to represent knowledge, consisting of entities (nodes) and relationships (edges), and is a semantic network capable of describing concepts, entities, and their interrelationships in the real world. In this specification, the medical knowledge graph is used to store and organize medical knowledge such as diseases and treatment methods from medical literature.

[0054] The full name of a large language model is Large Language Model (LLM). A large language model is a natural language processing model based on deep learning techniques, typically with billions to hundreds of billions or even more parameters, possessing powerful language understanding and generation capabilities. Large language models can employ the Transformer architecture or its variants (such as GPT, BERT, etc.), which utilizes the attention mechanism to globally model sequential data, efficiently handling long-distance dependencies and thus performing exceptionally well in natural language tasks. Large language models learn the statistical features and semantic relationships of language through pre-training on large-scale corpora, giving them outstanding generalization capabilities. The core capabilities of large language models include, but are not limited to: understanding contextual semantics, generating coherent and grammatically correct text, performing logical reasoning, and handling multi-task scenarios. Their usage typically includes two modes: direct inference and fine-tuning. In direct inference mode, the user guides the large language model to generate specific outputs by designing prompts. Prompts can be task descriptions or instructions in text form, used to stimulate the large language model's semantic understanding and generation capabilities. In fine-tuning mode, large language models are further trained on small-scale datasets within a specific domain to optimize their performance on specific tasks. The powerful generalization capabilities and flexibility of large language models make them an important tool in the field of artificial intelligence, providing efficient and accurate solutions for automated text generation and understanding.

[0055] In some embodiments, large language models can also understand and generate data from other modalities (such as visual and audio). In this case, large language models can also be called multimodal large language models (MLLMs). MLLMs provide a richer and more natural interactive experience by integrating multiple types of input and output, such as text, images, and sound. The core advantage of MLLMs lies in their ability to process and understand information from different modalities and fuse this information to accomplish complex tasks. For example, MLLMs can analyze an image and generate descriptive text, or generate a corresponding image based on a text description. This cross-modal understanding and generation capability makes MLLMs widely applicable across multiple fields.

[0056] It should be noted that the key technologies of large language models can be found in the detailed description in the paper "A Survey of Large Language Models" (paper number: arXiv:2303.18223v16, published on March 11, 2025), and will not be repeated here.

[0057] In this specification, large model devices, computing devices, and storage devices can be implemented using any device, equipment, platform, device cluster, etc., that has computing and processing capabilities.

[0058] Figure 1 This is just one of many implementation scenarios for the medical knowledge graph construction method provided in this application. In other implementation scenarios, the large model can also be deployed on a computing device, or the medical knowledge graph can be stored in a graph database deployed on a computing device. This application does not limit the specific implementation scenario.

[0059] It's understandable that medical knowledge graphs have extremely high requirements for information accuracy, unlike other types of knowledge graphs. In the construction of some existing knowledge graphs, complex medical relationships are described using ordinary triples. For example, a typical knowledge graph contains entity relationships like these:

[0060] {Children over three years old} — Suffering from → {Cold} ← Treatment — {xx Cold Remedy} (1)

[0061] In this context, {} represents entities, and arrows represent the directions of edges. When the subgraph “{Cold}←Treatment—{xxColdRelief}” is extracted from the above entity relationships, the question-answer pairs generated using this subgraph might be...

[0062] Q: What treatment should I use for a cold?

[0063] A: xx cold medicine.

[0064] The original description of the above entity relationship (1) in medical literature is: When a child over three years old has a cold, xx cold medicine can be used for treatment. However, the question-and-answer pair generated by the ordinary knowledge graph only expresses the meaning of "a cold can be treated with xx cold medicine", but ignores the limiting condition of "a child over three years old", which leads to a deviation between the generated question and answer and fails to restore the original meaning of the knowledge in the medical literature.

[0065] To more accurately reproduce the original meaning of knowledge in medical literature, this embodiment sets the text of knowledge points containing complete and rich information on the relationships between related entities as node attributes when constructing the knowledge graph. The connecting edges are designed to be undirected and untyped, and the nodes employ a dual entity type design. This design improves the knowledge graph's ability to express medical knowledge, thus enabling it to provide more accurate medical information. The node-edge design of the knowledge graph does not become more complex due to its enhanced expressive power; on the contrary, it becomes simpler. The following will combine... Figure 2 The methods provided in the embodiments will be described in detail.

[0066] Figure 2 This is a flowchart illustrating a method for constructing a medical knowledge graph, provided as an example. The method is executed via a computing device and specifically includes the following steps.

[0067] Step S210: Input several pages from the electronic file of the first medical document B1 into the large model. Through the large model N1, identify the entity data of several entities in the several pages. For example, identify the entity data of m entities.

[0068] In this context, the first medical document (B1) can be any medical document, which may include books, journal articles, academic papers, medical records, and medical reports. Books may include medical guides or textbooks. Medical documents typically consist of multiple pages, with each page containing text content related to medical knowledge. The purpose of constructing a knowledge graph is to organize the knowledge within medical documents using a graph structure, thereby enabling computers to more easily retrieve and understand the knowledge within these documents.

[0069] The large model N1 can be a commonly used large model, and it is obtained after training. The large model N1 is used to perform entity recognition on the input text content to obtain entity data of the entities.

[0070] Entities are usually represented by entity words, which are nouns in language that represent the names or concepts of things. Their historical categories correspond to nouns and pronouns in modern grammar. As a subclass of entity words, entity words also include nouns (including locative words, temporal words, and spatial words), numerals, classifiers, and some pronouns.

[0071] Entity recognition is a key technology in natural language processing, used to identify and extract meaningful entities from text content, such as names of people, places, and organizations. In the medical field, entity recognition is mainly used to extract medical concepts such as disease names, drug names, and treatment methods.

[0072] Electronic files of medical literature can be, but are not limited to, PDF or image formats such as JPG. The computing device can pre-parse the electronic file page by page, extracting the structured content and obtaining the text content of each page. This pre-prepared text content is then input into the large model N1.

[0073] A medical document typically contains many pages, especially books, which can reach hundreds of pages. To improve the ability of large models to recognize text content, one implementation approach is to progressively identify entity data and construct a knowledge graph. This progressive construction is an iterative system building method that gradually improves system functionality by processing data in stages and steps.

[0074] In practical implementation, all pages of the electronic document can be divided into multiple recognition units, and the pages contained in the current recognition unit are input into the large model N1. Each recognition unit can contain one or more pages. For example, one page can be considered as a recognition unit, or two pages can be considered as a recognition unit. By processing medical literature page by page, each page processed is analyzed in conjunction with existing knowledge, achieving cumulative knowledge growth.

[0075] After completing the recognition of the current recognition unit, the next recognition unit can be updated to the current recognition unit, and the steps of inputting the several pages contained in the current recognition unit into the large model can be repeated.

[0076] To improve the consistency of the large model in recognizing text content across multiple pages, several pages and the constructed knowledge graph G1 can be input into the large model. The large model N1, using entities in the constructed knowledge graph as references, can then recognize entity data of several entities across the several pages.

[0077] The "several pages" here can refer to several pages within the current recognition unit, specifically one page or a few pages.

[0078] The input to the large model can be entity names and relationships between entities from the constructed knowledge graph G1. Initially, the constructed knowledge graph G1 has no nodes or edges, so only pages can be input to the large model N1. As the iteration process progresses, the data in the constructed knowledge graph G1 gradually increases. At this point, entity names and related entities can be input to the large model N1 as a reference. In other words, in each iteration, the pages in the current recognition unit and the constructed knowledge graph G1 can be input to the large model N1.

[0079] For example, the built knowledge graph G1 input to a large model can include the following data:

[0080] Cold - xx Cold Remedy;

[0081] Children over three years old - have a cold;

[0082] Common cold - xx cold capsules.

[0083] In this context, "cold" and "xx cold medicine" are entity names, representing related entities. The hyphen "-" indicates a connection between the two entities.

[0084] In this embodiment, processing the document in small blocks (i.e., recognition units) rather than the entire electronic document allows the large model N1 to more carefully identify the specified text content. This results in a more detailed knowledge graph generated at each recognition unit in the electronic document. Feeding the large model N1 with all or a large number of pages of the entire electronic document directly would lead to the loss of significant details and facts, which is related to the attention mechanism of the large model N1 itself and the training corpus. The large model N1 tends to perform recognition tasks in a relatively uncomplicated manner.

[0085] In the progressive construction approach, the already constructed knowledge graph G1 is input into the large model N1 as a reference, ensuring the contextual consistency of knowledge extraction and the cumulative learning effect.

[0086] To improve the accuracy of entity data recognition by the large model, corresponding prompts can be set and input into the large model N1. These prompts can include the page's text content, entity names and relationships from the constructed knowledge graph G1, and instances of knowledge points.

[0087] Furthermore, in the data input to the large model, in addition to the page of the current recognition unit to be identified, the previous pages can also be input as reference information. Through the context-aware capability of the large model N1, entity data can be identified more accurately.

[0088] In practice, the page number of the page to be identified, the constructed knowledge graph G1, and the complete electronic file can be input into the large model N1 in the form of a prompt, and the large model N1 can be instructed to perceive the context to obtain the recognition result.

[0089] Context awareness refers to the system's ability to understand and utilize relevant background information when processing current information. In this embodiment, when processing each page of medical literature, the system combines the constructed knowledge graph G1 with complete document information for analysis, improving the accuracy and consistency of identification.

[0090] The entity data will be explained in detail below.

[0091] Each entity data point includes the entity name, entity type, related entities, and related knowledge points. The entity type only includes disease entity types and treatment method entity types. That is, when identifying entities, the large model N1 determines the entity as a disease entity type or a treatment method entity type based on semantics, and does not classify the entity as other types. Table 1 shows an example of an entity identified by the large model N1.

[0092] Table 1

[0093]

[0094] The entity data includes the entity name, entity type, related entities, and knowledge associated with the entity. Larger models can identify multiple entities; Table 1 is just one example of an entity.

[0095] Knowledge points exist in the form of knowledge point text. An entity's entity data can contain at least one knowledge point text, or multiple knowledge point texts. The entity data shown in Table 1 contains two knowledge points: Symptoms of a cold: sneezing, runny nose, headache, fever; Children over three years old can generally use xx cold medicine to treat colds.

[0096] A knowledge point is a relatively independent, complete, and communicable unit of information, typically revolving around a concept, fact, principle, or skill. It is not a few isolated words or a general passage of text, but rather a bounded, structured, and transferable piece of text. Bounded means it has a clear starting and ending point, answering a specific question. Structured means it usually includes at least one of the three levels: "what," "why," and "how to use." Transferable means it can be invoked in different contexts, such as generating questions and answers. Knowledge points are characterized by completeness, explainability, and applicability.

[0097] Knowledge points are derived from the text content on a page. A knowledge point can be the original text within the page's content, or it can be obtained by modifying the original text in a larger model. These modifications can include removing typos, adjusting word order, and other operations that do not alter the semantics of the original text.

[0098] Taking any entity A as an example, its entity data contains a first knowledge point text. This first knowledge point text describes the relationship information between entity A and its associated entities. This relationship information includes the type of relationship, and at least one of the following: the conditions for the relationship to exist, the reason for its existence, the method of its existence, and its effect. The first knowledge point contains the entity names of entity A and its associated entities. Larger models do not recognize the type of relationship between entities with existing relationships; instead, they represent this relationship type through the first knowledge point text.

[0099] The knowledge points of entity A may also include a second knowledge point text, which is not used to describe the above-mentioned relationship information, but is used to describe entity A.

[0100] For example, taking the two knowledge points in Table 1 as examples, the first knowledge point, "Symptoms of a cold: sneezing, runny nose, headache, and fever," belongs to the second knowledge text and is used to describe the entity "cold." The second knowledge point, "Children over three years old can generally use xx cold medicine to treat colds," belongs to the first knowledge text and is used to describe the relationship between the entities "cold" and "xx cold medicine." The relationship information includes the relationship type, namely "treatment," and also includes the condition "children over three years old."

[0101] As can be seen from the description of the first knowledge point text, the entity data of two related entities contain at least one common knowledge point. For example, for the related entities "cold" and "xx cold medicine", the knowledge point "children over three years old can generally use xx cold medicine to treat colds" will appear in the knowledge points of both entities at least simultaneously.

[0102] The first knowledge point text contains complete information related to several entities involved, and may also contain other limiting conditions. That is to say, there are at least two situations: one is that the first knowledge point text contains information on the relationships between multiple entities, and the other is that the first knowledge point text contains other limiting conditions in addition to the relationship information, and these limiting conditions do not include the identified entities.

[0103] Step S220 involves performing semantic similarity matching between the identified m entities and several entities in the constructed knowledge graph G1. For example, the current constructed knowledge graph G1 contains m entities.

[0104] When using a progressive construction mechanism, semantic similarity matching can be performed between the m entities identified from the current identification unit and the n entities in the constructed knowledge graph G1. Initially, if the current identification unit is the first identification unit, this step can be skipped. If the current identification unit is not the first identification unit, and several entities already exist in the constructed knowledge graph G1, semantic similarity matching can be performed.

[0105] When performing semantic similarity matching, the entity names of the identified m entities can be matched with the entity names of the m entities in the constructed knowledge graph G1. Entity names are represented by entity words. Specifically, when performing semantic similarity matching on entity names, the entity names can be converted into vectors and similarity matching can be performed between vectors; alternatively, the entity names can be input into the large model N1, which performs textual meaning matching and outputs similarity values.

[0106] When the semantic similarity between two entity names is greater than a preset threshold, the two entity names are considered similar entities; when the semantic similarity between two entity names is not greater than the preset threshold, the two entity names are considered not similar entities. Semantic similarity is an indicator that measures the degree to which two words, phrases, or texts are similar in meaning.

[0107] Specifically, when m entities are identified, and the constructed knowledge graph G1 contains n entities, the m entities can be semantically matched with the n entities respectively, i.e., m*n similarity matches can be performed. The final result may be that no two entities belong to the same entity group, or it may be that at least one entity group belongs to the same entity group.

[0108] This embodiment identifies different terms expressing the same medical concept by calculating semantic similarity, thereby achieving intelligent entity fusion.

[0109] Step S230: Based on the semantic similarity matching results, merge the entity data of several identified entities into the constructed knowledge graph G1.

[0110] The aforementioned merging process involves treating entities as nodes and knowledge points as attributes of those nodes, constructing typeless connections between entities and their associated entities. For each identification unit, the graph update operation in this step can be performed, merging the entity data identified from the new identification unit into the already constructed knowledge graph G1, and dynamically modifying and improving the knowledge graph structure. Image updates include operations such as adding new entities, establishing new relationships, and merging duplicate entities.

[0111] Based on the semantic similarity matching results, when it is determined that, for example, the first entity A1 among the m identified entities is similar to the second entity A2 in the constructed knowledge graph G1, the associated entities and knowledge points in the entity data of the first entity A1 are merged into the node data of the second entity A2. This achieves the merging of similar entities without needing to add the first entity A1 as a new node to the constructed knowledge graph G1, avoiding duplicate nodes and simplifying the structure of the knowledge graph.

[0112] For example, the entity shown in Table 1 is the first entity A1, and the entity shown in Table 2 below is the second entity A2 in the constructed knowledge graph G1.

[0113] Table 2

[0114]

[0115] The "common cold" in Table 2 and the "cold" in Table 1 are similar entities. The related entities and knowledge points in the entities shown in Table 1 can be merged into the entities shown in Table 2 to obtain the node data shown in Table 3.

[0116] Table 3

[0117]

[0118] The underlined parts in Table 3 represent new content, including newly added related entities and knowledge points.

[0119] In this embodiment, semantic fusion was performed on the first entity A1 and the second entity A2. Semantic fusion is a technique that merges entities that express the same concept but use different words by understanding the semantic meaning of the text. For example, "cold" and "common cold" are identified as the same medical concept, and "hypertension" and "elevated blood pressure" are identified as the same medical concept and merged.

[0120] Based on the semantic similarity matching results, when it is determined that, for example, the third entity A3 among the m identified entities, does not belong to any similar entity to any entity in the constructed knowledge graph G1, the third entity A3 is directly added as a new node to the constructed knowledge graph G1.

[0121] Specifically, the node data of a new node can be determined based on the entity data of the third entity A3. The node data includes the entity identifier (i.e., node ID), entity name, entity type, associated nodes, and knowledge points. The data in the associated nodes is the identifier data of the associated entities.

[0122] In this embodiment, the knowledge graph construction process employs dual-entity modeling and an undirected, untyped graph mechanism. The constructed knowledge graph G1 contains only two types of entity nodes: disease entity type nodes and treatment method entity type nodes, and does not contain nodes of other types. Entity nodes belonging to the disease entity type include diseases, symptoms, and pathological manifestations, such as headaches and colds. Entity nodes belonging to the disease entity type include drug treatments, surgical plans, and rehabilitation measures.

[0123] The nodes in the constructed knowledge graph G1 represent entities. The node data is obtained from the entity data and includes the entity name, entity type, related entities that are associated with the entity, and knowledge points related to the entity.

[0124] The edges connecting nodes are undirected and typeless. An undirected graph is a graph structure in graph theory where edges have no directionality, representing symmetric relationships between nodes. In this scheme, edges are used to connect disease and treatment entities, representing their correlation without emphasizing directionality. Furthermore, edges lack type and attribute information. In the knowledge graph schema, there can be edges connecting disease and treatment entities, disease entities with each other, and treatment entities with each other. Therefore, in the triples of the knowledge graph in this embodiment, the edges have no descriptive information, simplifying the triples to binary tuples.

[0125] The innovative dual-entity modeling approach simplifies complex medical knowledge into a model containing only two types of entities: disease symptoms and treatment methods. The relationships between these entities are expressed through undirected graph connections.

[0126] Step S240: When the entity data corresponding to all pages of the electronic document are merged into the constructed knowledge graph G1, the constructed knowledge graph G1 is used as the medical knowledge graph of the first medical document B1.

[0127] Different medical knowledge graphs can be constructed for different medical literature, without forcibly merging the medical knowledge graphs of different medical literature, thus fully preserving the characteristics and differences of different medical literature.

[0128] In this embodiment, a progressive iterative graph construction mechanism is adopted, gradually constructing the knowledge graph by iterating through recognition units (e.g., by page). Steps S210 to S230 described above are the process of constructing a knowledge graph for a single recognition unit. This process is executed iteratively, continuously updating the recognition units until the recognition and graph construction process for all recognition units is completed.

[0129] When medical literature has few pages, such as a patient's medical record, it typically consists of only one or a few pages. In this case, step S210 can also input all pages of the medical literature into a large model, identify the entity data of each entity, and construct a knowledge graph from the entity data to obtain the medical knowledge graph of the medical literature. The entity data is described in the same way as in step S210. Constructing the knowledge graph from the entity data includes treating entities as nodes, treating knowledge points as attributes of nodes, constructing typeless connections between entities and their associated entities, and directly using entity names and entity types as node names and node types.

[0130] In this embodiment, knowledge points related to entities are identified through a large model, and these knowledge points are added to the attributes of entity nodes. Furthermore, each entity can be associated with multiple detailed knowledge points, which retain the original literature information completely during the extraction process. This makes the medical knowledge graph more expressive and accurate.

[0131] Compared to entity relations (1), the entity relations of the knowledge graph in this embodiment are as follows:

[0132] {Children over three years old}—{Cold}—{xx Cold Remedy} (2)

[0133] In this diagram, {} represents entities, and horizontal lines represent edges, which are undirected and typeless. When the subgraph “{cold}—{xx cold medicine}” is extracted from the above entity relationship, a question-and-answer pair can be generated by combining it with the knowledge point “When children over three years old have a cold, xx cold medicine can be used for treatment” from the node data. The knowledge point contains complete and rich information related to the entities, and the question-and-answer pair generated based on this knowledge point might be:

[0134] Q: What treatment should be used for a child over three years old who has a cold?

[0135] A: xx cold medicine.

[0136] It is evident that the generated document pairs are based on knowledge points, resulting in higher accuracy and scientific rigor.

[0137] To reduce the data volume of the medical knowledge graph, in another embodiment of this application, a shared knowledge point pool S can be constructed. A unique number is assigned to each knowledge point of all entities, and these numbers are stored in the shared knowledge point pool S. The knowledge point number is also set in the node attributes. Since several entities with related relationships are associated with at least the same knowledge point simultaneously, storing the knowledge point in the shared knowledge point pool S avoids duplicate storage of the same knowledge point, thereby reducing the data volume of the medical knowledge graph. This embodiment will be described in detail below.

[0138] In step S230, taking the step of directly adding the third entity A3 as a new node to the constructed knowledge graph G1 as an example, the usage of shared knowledge points and the construction of the graph are explained, specifically including the following steps 1 and 2.

[0139] Step 1: Based on the knowledge point text and corresponding knowledge point number stored in the shared knowledge point pool S, determine the knowledge point number corresponding to the knowledge point text in the entity data of the third entity A3, and confirm that the knowledge point text of the third entity A3 has been stored in the shared knowledge point pool S.

[0140] The shared knowledge point pool S stores the knowledge point texts and their numbers. The entity data of the third entity A3 can contain one or more knowledge point texts. The following explanation uses a single knowledge point text as an example; each knowledge point text is processed in the same way.

[0141] When there are no knowledge points in the shared knowledge point pool S, the knowledge point text number of the third entity A3 can be generated directly, and the knowledge point text of the third entity A3 and its number can be stored in the shared knowledge point pool S accordingly. For example, the number can be generated sequentially starting from 1.

[0142] When the shared knowledge point pool S already stores knowledge point text, the knowledge point text of the third entity A3 can be matched with the knowledge point text in the shared knowledge point pool S.

[0143] When a knowledge point text identical to the knowledge point text of the third entity A3 exists in the shared knowledge point pool S, the number of that knowledge point text in the shared knowledge point pool S shall be used as the corresponding number of the knowledge point text of the third entity A3.

[0144] When there is no knowledge point text in the shared knowledge point pool S that is identical to the knowledge point text of the third entity A3, a number corresponding to the knowledge point text of the third entity A3 is generated based on the existing knowledge point numbers in the shared knowledge point pool S, and the number and the knowledge point text of the third entity A3 are stored in the shared knowledge point pool S accordingly.

[0145] Identical knowledge point texts mean that two knowledge points are completely identical. This can be determined by calculating the MD5 (Mean Digest) values ​​of the two knowledge point texts and comparing whether the MD5 values ​​are the same.

[0146] When generating the ID corresponding to the knowledge point text of the third entity A3, the ID can be assigned sequentially based on existing IDs in the shared knowledge point pool S, or it can be assigned out of order, as long as the generated ID is unique within the shared knowledge point pool S. The ID is a unique identifier for the knowledge point, and the two correspond one-to-one.

[0147] Step 2: Add the determined knowledge point number to the attributes of the new node corresponding to the third entity A3 in the constructed knowledge graph G1.

[0148] For example, if node 1 and node 2 are related, and both their entity data contain knowledge point 'a', and node 2 and node 3 are related, and both their entity data contain knowledge point 'b', then knowledge points 'a' and 'b' can be stored in a shared knowledge point pool S, and assigned the numbers 'a' and 'b' respectively. The number 'a' can be set in the attributes of node 1, the numbers 'a' and 'b' can be set in the attributes of node 2, and the number 'b' can be set in the attributes of node 3.

[0149] Figure 3 This is a schematic diagram illustrating the principle of a constructed knowledge graph and a shared knowledge point pool. The left side shows an exemplary knowledge graph node relationship diagram, where each node's data displays the entity ID, entity name, entity type, and knowledge point number. Ellipses indicate omitted information. Black circles represent disease entities, and white circles represent treatment method entities. The shared knowledge point pool lists each knowledge point number and its corresponding text; ellipses indicate omitted specific knowledge point text. The knowledge point numbers in node 1 include 'a'. 123 b1 and c1, where a 123 These are knowledge point numbers shared by nodes 1, 2, and 3. b1 and c1 are knowledge point numbers unique to node 1. The knowledge point numbers in node 2 include a... 123 And d2, where d2 is a unique knowledge point number for node 2. The knowledge point numbers in node 3 include a 123 f3 and e 34 Where f3 is the unique knowledge point number of node 3, e 34 This is a knowledge point number shared by nodes 3 and 4. The knowledge point numbers in node 4 include e. 34 .

[0150] As can be seen, knowledge points shared by multiple nodes only need to be stored once in the shared knowledge point pool, while the node data only needs to store the knowledge point number, without needing to store the knowledge point repeatedly. The shared knowledge point pool can be stored in the same storage space as the corresponding medical knowledge graph.

[0151] In this embodiment, knowledge point sharing is an information organization mechanism designed in this embodiment. It allows the same detailed medical knowledge point to be associated and referenced by multiple different entities, avoiding duplicate storage of information and improving the efficiency of knowledge utilization.

[0152] When the medical knowledge graph corresponding to the first medical document B1 is obtained, a query can be performed on this medical knowledge graph based on the query subgraph to be queried, and the query results can be obtained. Based on the entities in the query results and the knowledge points corresponding to their related entities, the required questions and / or answers can be generated. The process of generating questions and / or answers can be executed through a large model.

[0153] The medical knowledge graph in this embodiment can be used to generate an evaluation set, which contains multiple question-answer pairs. The evaluation set is a standardized dataset used to test and evaluate the performance of the AI ​​system, containing input samples and their corresponding correct answers. The medical AI evaluation set is used to verify the accuracy and reliability of the AI ​​system in understanding and applying medical knowledge.

[0154] Medical knowledge graphs can also be used to provide retrieval services. In practical applications, the medical knowledge graph of the second medical document B2 can also be obtained in the manner described above, and stored in a different storage space than the medical knowledge graph of the second medical document B2.

[0155] Specifically, multiple medical knowledge graphs can be stored in the search engine Elasticsearch, and an inverted index can be built for each medical knowledge graph.

[0156] Elasticsearch is a distributed search and analytics engine based on the full-text search engine (Lucene), providing full-text search, structured search, and analytics capabilities. This embodiment uses Elasticsearch to store and retrieve multiple independent medical knowledge graphs.

[0157] An inverted index is an index structure that maps each word in a document to a list of documents containing that word. In this embodiment, the inverted index is used to establish a fast retrieval mechanism from medical terms (entity names) to their corresponding locations in the medical knowledge graph, supporting efficient cross-graph information retrieval.

[0158] When a search is required, a first result is obtained based on the medical knowledge graph of the first medical document B1, and a second result is obtained based on the medical knowledge graph of the second medical document B2. The first result, the second result, and their corresponding source information are used as the search results. The source information includes the name of the corresponding medical document.

[0159] This embodiment employs cross-graph retrieval, a technique that retrieves information across multiple independent medical knowledge graphs. This technique enables the simultaneous search and integration of relevant information from graphs of different sources, achieving knowledge complementarity and verification.

[0160] Figure 4 This is a schematic diagram of an overall architecture provided for an embodiment. It includes a document parsing processor, a knowledge graph builder, and a cross-graph retrieval system. The document parsing processor is responsible for parsing PDF-format medical documents page by page, extracting structured text content, and inputting the text content into the knowledge graph builder. The knowledge graph builder uses an AI analysis engine for entity recognition and knowledge point extraction, and constructs and maintains the graph structure through similar entity recognition components and a graph updater. The above components can construct knowledge graphs for different medical documents separately; for example, they can obtain graph 1 for guideline 1, graph 2 for guideline 2, and graph k for guideline k, etc. Constructing an independent graph for each document ensures the self-consistency of the content.

[0161] The cross-graph retrieval system is used to achieve semantic retrieval and similar content matching between multiple graphs based on an inverted index builder and a semantic retrieval engine. Within the same graph, retrieval can be performed through nodes or through inverted index retrieval.

[0162] The above embodiments employ a progressive graph growth mechanism and a page-by-page iterative processing approach. When processing new content on each page, analysis can be performed in conjunction with the already constructed knowledge graph, avoiding the inefficiency of multiple full-scale processing and significantly improving construction efficiency.

[0163] To address the issues of content conflicts and confusion among medical literature from different sources, this embodiment innovatively establishes an independent knowledge graph for each authoritative document, much like creating a dedicated knowledge map for each book. This ensures the consistency and reliability of the content in each graph, while enabling information exchange through a unified cross-graph retrieval system. This avoids information conflicts and achieves complementary verification of knowledge.

[0164] To address the issues of complex structures and difficult maintenance in traditional knowledge graphs, this embodiment employs a dual-entity simplified modeling mechanism. This mechanism simplifies complex medical knowledge into two core entities: disease symptoms and treatment methods. Relationships are established through undirected, typeless connections, significantly reducing system complexity. Simultaneously, a rich-information knowledge point sharing mechanism ensures the integrity of medical information and improves data storage efficiency.

[0165] To address the issue of insufficient accuracy in knowledge extraction, this embodiment employs a semantic fusion update mechanism that can intelligently identify the same medical concepts (such as hypertension and elevated blood pressure) expressed in different ways, automatically merge and update them, significantly improving the accuracy of knowledge extraction and the quality of the knowledge map.

[0166] The above embodiments focus on constructing dual-entity nodes and undirected, typeless connecting edges. Based on this, knowledge points are added to attributes to reflect rich information between entities. In another embodiment, in a knowledge graph containing two or more entity types and directed edges, adding knowledge points reflecting rich information between entities to node attributes can also improve the accuracy of the knowledge graph's representation of medical literature.

[0167] In this specification, the terms "first" in the first medical literature, the first knowledge point text, the first entity, and the first result, as well as the corresponding "second" (if present) in the text, are used merely for the convenience of distinction and description and do not have any limiting meaning.

[0168] The foregoing description describes specific embodiments of this specification; other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than those shown in the embodiments, and the desired result may still be achieved. Furthermore, the processes depicted in the drawings do not necessarily need to follow the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0169] Figure 5 This is a schematic block diagram of a medical knowledge graph construction apparatus provided for an embodiment. This apparatus embodiment is related to... Figure 2 The method embodiment shown corresponds to this. The device 500 is deployed in a computing device and includes:

[0170] The entity recognition module 510 is configured to input several pages from the electronic file of the first medical document into a large model, and through the large model, identify entity data of several entities in the several pages; wherein, each page contains text content, and any entity data includes entity name, entity type, related entities that are associated with the entity, and knowledge points related to the entity; the entity type only includes disease entity type and treatment method entity type, and the knowledge points include first knowledge point text, which is determined based on the corresponding text content and is used to describe the relationship information between the entity and related entities. The relationship information includes the type of relationship, and also includes at least one of the following: establishment conditions, reasons, and establishment methods;

[0171] The data matching module 520 is configured to perform semantic similarity matching between several identified entities and several entities in the constructed knowledge graph;

[0172] The data merging module 530 is configured to merge the entity data of several identified entities into the constructed knowledge graph based on semantic similarity matching results. The merging process includes constructing typeless connections between entities and their associated entities, treating entities as nodes and knowledge points as attributes of those nodes.

[0173] The knowledge graph construction module 540 is configured to use the constructed knowledge graph as the medical knowledge graph of the first medical document when the entity data corresponding to all pages of the electronic document is merged into the constructed knowledge graph.

[0174] In one implementation, the entity recognition module 510 is specifically configured to: input several pages and the constructed knowledge graph into a large model, and through the large model, using the entities in the constructed knowledge graph as references, recognize the entity data of several entities in the several pages.

[0175] In one implementation, the entity recognition module 510 is specifically configured to input entity names and relationships between entities from several pages and the constructed knowledge graph into the large model.

[0176] In one implementation, the entity recognition module 510 is specifically configured to input several pages contained in the current recognition unit into a large model. Specifically, all pages contained in the electronic document are divided into multiple recognition units.

[0177] The device 500 also includes a page update module (not shown in the figure), configured to update the next identification unit to the current identification unit after merging the entity data of several entities identified from the current identification unit into the constructed knowledge graph, and return to the execution entity identification module 510.

[0178] In one implementation, the data matching module 520 is specifically configured to perform semantic similarity matching between the entity names of the identified entities and the entity names of the entities in the constructed knowledge graph.

[0179] In one implementation, the data merging module 530 includes an entity merging submodule 31 and a new node adding submodule 32. The entity merging submodule 31 is configured to merge the associated entities and knowledge points from the entity data of the first entity into the node data of the second entity when it is determined that a first entity among several identified entities is a similar entity to a second entity in the constructed knowledge graph. The new node adding submodule 32 is configured to directly add the third entity as a new node to the constructed knowledge graph when it is determined that a third entity among several identified entities is not a similar entity to any entity in the constructed knowledge graph.

[0180] In one implementation, the new node adding submodule 32 includes an ID determination unit 32_1 and an attribute determination unit 32_2. The ID determination unit 32_1 is configured to determine the knowledge point number corresponding to the knowledge point text of the third entity based on the knowledge point text and corresponding knowledge point IDs already stored in the shared knowledge point pool, and to determine that the knowledge point text of the third entity is stored in the shared knowledge point pool. The attribute determination unit 32_2 is configured to add the determined knowledge point number to the attributes of the new node corresponding to the third entity in the constructed knowledge graph.

[0181] In one implementation, the numbering determination unit 32_1 is specifically configured as follows: when there is a knowledge point text in the shared knowledge point pool that is the same as the knowledge point text of the third entity, the number of the knowledge point text in the shared knowledge point pool is used as the corresponding number of the knowledge point text of the third entity; when there is no knowledge point text in the shared knowledge point pool that is the same as the knowledge point text of the third entity, the number of the knowledge point text of the third entity is generated based on the existing knowledge point numbers in the shared knowledge point pool, and the number and the knowledge point text of the third entity are stored in the shared knowledge point pool accordingly.

[0182] In one implementation, the apparatus 500 includes a subgraph query module 550 and a text generation module 560. The subgraph query module 550 is configured to, upon obtaining a medical knowledge graph corresponding to the first medical document, perform a query within the medical knowledge graph based on the subgraph to be queried, and obtain query results. The text generation module 560 is configured to generate the required questions and / or answers based on the entities in the query results and the knowledge points corresponding to their associated entities.

[0183] In one implementation, the device 500 further includes a knowledge graph storage module 570, a knowledge graph retrieval module 580, and a result determination module 590. The knowledge graph storage module 570 is configured to store the medical knowledge graph of the first medical document in a different storage space than the medical knowledge graph of the second medical document when the medical knowledge graph of the first medical document is obtained. The knowledge graph retrieval module 580 is configured to obtain a first result based on the medical knowledge graph of the first medical document and a second result based on the medical knowledge graph of the second medical document when a retrieval is required. The result determination module 590 is configured to use the first result, the second result, and their corresponding source information as the retrieval results; wherein the source information includes the name of the corresponding medical document.

[0184] The above-described apparatus embodiments correspond to the method embodiments, and detailed descriptions can be found in the description of the method embodiments section, which will not be repeated here. The apparatus embodiments are derived based on the corresponding method embodiments and have the same technical effects as the corresponding method embodiments; detailed descriptions can be found in the corresponding method embodiments.

[0185] This specification also provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform... Figures 1 to 4 Any one of the methods described.

[0186] This specification also provides a computing device, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement... Figures 1 to 4 Any one of the methods described.

[0187] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for storage media and computing devices are basically similar to the method embodiments, so they are described more simply; relevant parts can be referred to the descriptions of the method embodiments.

[0188] Those skilled in the art will recognize that the functions described in the embodiments of the present invention in one or more of the above examples can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0189] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made based on the technical solutions of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a medical knowledge graph, comprising: Several pages from the electronic file of the first medical document are input into a large model. The large model identifies entity data of several entities within the pages. Each page contains text content, and each entity data includes an entity name, entity type, related entities, and related knowledge points. The entity type includes only disease entity types and treatment method entity types. The knowledge points include first knowledge point text, which is determined based on the corresponding text content and is used to describe the relationship information between the entity and related entities. The relationship information includes the type of relationship and at least one of the following: conditions for establishment, reasons, and methods of establishment. The identified entities are semantically similar to several entities in the constructed knowledge graph. Based on the semantic similarity matching results, the entity data of the identified entities are merged into the constructed knowledge graph; wherein, the merging includes using entities as nodes and knowledge points as attributes of nodes to construct typeless connection edges between entities and their associated entities. When the entity data corresponding to all pages of the electronic document are merged into the constructed knowledge graph, the constructed knowledge graph is used as the medical knowledge graph of the first medical document.

2. The method according to claim 1, wherein the step of inputting several pages from the electronic file of the first medical document into the large model comprises: Several pages and the constructed knowledge graph are input into the large model. Using the large model and the entities in the constructed knowledge graph as references, entity data of several entities in the several pages are identified.

3. The method according to claim 2, wherein the step of inputting several pages and the constructed knowledge graph into the large model includes: Input the entity names and relationships between entities from several pages of the constructed knowledge graph into the large model.

4. The method according to claim 1, wherein the step of inputting several pages from the electronic file of the first medical document into the large model comprises: The current identification unit contains several pages, which are then input into the large model; wherein, all pages contained in the electronic document are divided into multiple identification units; After merging entity data of several entities identified from the current identification unit into the constructed knowledge graph, the method further includes: Update the next recognition unit to the current recognition unit, and return to the step of inputting the several pages contained in the current recognition unit into the large model.

5. The method according to claim 1, wherein the step of performing semantic similarity matching between the identified entities and the entities in the constructed knowledge graph includes: The entity names of the identified entities are semantically similar to the entity names of the entities in the constructed knowledge graph.

6. The method according to claim 1, wherein the step of merging the entity data of the identified plurality of entities into the constructed knowledge graph includes: When it is determined that the first entity among the identified entities is a similar entity to the second entity in the constructed knowledge graph, the associated entities and knowledge points in the entity data of the first entity are merged into the node data of the second entity. When it is determined that the third entity among the identified entities is not a similar entity to any entity in the constructed knowledge graph, the third entity is directly added as a new node to the constructed knowledge graph.

7. The method according to claim 6, wherein the step of directly adding the third entity as a new node to the constructed knowledge graph includes: Based on the knowledge point text and corresponding knowledge point number stored in the shared knowledge point pool, determine the knowledge point number corresponding to the knowledge point text of the third entity, and determine that the knowledge point text of the third entity is stored in the shared knowledge point pool; The identified knowledge point number is added to the attributes of the new node corresponding to the third entity in the constructed knowledge graph.

8. The method according to claim 7, wherein the step of determining the knowledge point number corresponding to the knowledge point text of the third entity and determining that the knowledge point text of the third entity is stored in the shared knowledge point pool includes: When there is a knowledge point text in the shared knowledge point pool that is the same as the knowledge point text of the third entity, the number of the knowledge point text in the shared knowledge point pool is used as the corresponding number of the knowledge point text of the third entity. When there is no knowledge point text in the shared knowledge point pool that is identical to the knowledge point text of the third entity, a number for the knowledge point text of the third entity is generated based on the existing knowledge point numbers in the shared knowledge point pool, and the number and the knowledge point text of the third entity are stored in the shared knowledge point pool accordingly.

9. The method according to claim 1, further comprising, when obtaining the medical knowledge graph corresponding to the first medical literature: Based on the query subgraph to be queried, a query is performed in the medical knowledge graph to obtain the query results; Based on the knowledge points corresponding to the entities and their associated entities in the query results, the required questions and / or answers are generated.

10. The method according to claim 1, further comprising, when obtaining the medical knowledge graph of the first medical literature: The medical knowledge graph is stored in a different storage space than the medical knowledge graph of the second medical document; When a search is required, a first result is obtained based on the medical knowledge graph of the first medical document, and a second result is obtained based on the medical knowledge graph of the second medical document; The first result, the second result, and their corresponding source information are used as search results; wherein, the source information includes the name of the corresponding medical literature.

11. A device for constructing a medical knowledge graph, comprising: The entity recognition module is configured to input several pages from the electronic file of the first medical document into a large model, and through the large model, identify entity data of several entities in the pages; wherein, each page contains text content, and any entity data includes entity name, entity type, related entities that are associated with the entity, and knowledge points related to the entity; the entity type only includes disease entity type and treatment method entity type, and the knowledge points include first knowledge point text, which is determined based on the corresponding text content and is used to describe the relationship information between the entity and related entities, the relationship information including the type of relationship, and at least one of the following: establishment conditions, reasons, and establishment methods; The data matching module is configured to perform semantic similarity matching between the identified entities and several entities in the constructed knowledge graph; The data merging module is configured to merge the entity data of the identified entities into the constructed knowledge graph based on the semantic similarity matching results; wherein, the merging includes constructing typeless connection edges between entities and their associated entities, using entities as nodes and knowledge points as attributes of the nodes. The knowledge graph construction module is configured to use the constructed knowledge graph as the medical knowledge graph of the first medical document when the entity data corresponding to all pages of the electronic document are merged into the constructed knowledge graph.

12. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-10.

13. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-10.

Images