Knowledge graph construction and query method, device, electronic device and storage medium

Abstract

The present invention provides a knowledge graph construction and query method, device, electronic device and storage medium, wherein the method includes: obtaining a target document set, the target document set includes multiple basic research documents; performing entity recognition on sentences in each basic research document to obtain entity pairs contained in each basic research document, the entity pairs include entities and entity relationships between entities; based on the types of entities contained in each basic research document, determining the causal strength of the entity relationship; based on the document citation information of each basic research document, determining the heat of the entity pair; based on entities, entity relationships, causal strength and heat, constructing a knowledge graph. The knowledge graph construction and query method, device, electronic device and storage medium provided by the present invention can efficiently, comprehensively and accurately sort out the information in massive biomedical data and construct a knowledge graph, helping users quickly learn the root cause of the disease.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a knowledge graph construction and query method, device, electronic device and storage medium. Background Art

[0002] With the development of information technology, the information resources on the Internet are becoming more and more abundant, the scale of information data is becoming larger and larger, and the forms of expression are becoming more and more diverse. How to realize information exchange between humans and machines and intelligently filter and process massive amounts of data is the key to technological breakthroughs in the field of artificial intelligence.

[0003] In today's biomedical research, more and more biomedical data are generated and accumulated, and eventually stored in the form of literature, conference abstracts, experimental data, etc. The rapid accumulation and continuous expansion of these data have posed a huge challenge to researchers. How to extract truly transformative information and knowledge from these complex and heterogeneous data to better understand the root causes of diseases and their treatments has become an important task in current biomedical research. Summary of the invention

[0004] The present invention provides a knowledge graph construction and query method, device, electronic device and storage medium to solve the problems in the prior art that effective information extracted from massive data is incomplete and inaccurate, and the efficiency of information mining and extraction is low.

[0005] The present invention provides a knowledge graph construction method, comprising:

[0006] Acquire a target document set, wherein the target document set includes a plurality of basic research documents;

[0007] Performing entity recognition on sentences in each basic research document to obtain entity pairs contained in each basic research document, wherein the entity pairs include entities and entity relationships between the entities;

[0008] Based on the types of entities included in each basic research document, determining the causal strength of the entity relationship between the entities included in each basic research document;

[0009] Based on the literature citation information of each basic research literature, determining the popularity of the entity pairs contained in each basic research literature;

[0010] A knowledge graph is constructed based on the entities, the entity relationships, the causal strengths, and the heat.

[0011] According to a knowledge graph construction method provided by the present invention, the step of obtaining a target document set includes:

[0012] Get the initial document collection;

[0013] Based on the publication type of each initial document in the initial document set, secondary processed documents and clinical research documents are screened out from the initial document set to obtain a candidate document set;

[0014] Basic research documents are screened from the candidate document set to construct the target document set.

[0015] According to a knowledge graph construction method provided by the present invention, based on the publication type of each initial document in the initial document set, secondary processing documents and clinical research documents are screened out from the initial document set to obtain a candidate document set, including:

[0016] Based on the publication type of each initial document in the initial document collection, secondary processed documents and clinical research documents are screened out from the initial document collection to obtain a first document collection;

[0017] Based on the publication type and publication time of each first document in the first document collection, selecting a second document from the first document collection, where the second document is the first document of the publication type to be updated;

[0018] Inputting the second document into a document classifier to obtain the document type of the second document output by the document classifier;

[0019] The second document whose document type is non-basic research document is deleted from the first document set to obtain the candidate document set.

[0020] According to a knowledge graph construction method provided by the present invention, the entity recognition is performed on the sentences in each basic research document to obtain the entity pairs contained in each basic research document, including:

[0021] Obtain the title text and abstract text of each basic research document;

[0022] Inputting the title text and the summary text into a sentence classifier to obtain the sentence type of each sentence in the title text and the summary text output by the sentence classifier;

[0023] Entity recognition is performed on the sentences of the sentence type to be recognized to obtain entity pairs contained in each basic research document.

[0024] According to a knowledge graph construction method provided by the present invention, the entity recognition is performed to obtain entity pairs contained in each basic research document, including:

[0025] Performing entity recognition on the sentence to obtain entities and pronouns in the sentence;

[0026] Inputting the entity and the pronoun into a reference relationship classifier, and obtaining a reference relationship between the entity and the pronoun output by the reference relationship classifier;

[0027] In the case where the reference relationship is yes, the pronouns in the sentence are replaced with corresponding entities to obtain an optimized sentence;

[0028] The optimization statement and entities in the optimization statement are input into an entity relationship classifier, and entity relationships between the entities output by the entity relationship classifier are obtained.

[0029] According to a knowledge graph construction method provided by the present invention, the entity recognition is performed to obtain entity pairs contained in each basic research document, and further includes:

[0030] Filling the optimization statement and entities in the optimization statement into a relational query template to obtain a relational query statement;

[0031] The relationship query statement is input into a question-answering language model to obtain the entity relationship output by the question-answering language model.

[0032] According to a knowledge graph construction method provided by the present invention, the causal strength of the entity relationship between the entities contained in each basic research document is determined based on the type of the entity contained in the basic research document, including:

[0033] If the types of entities contained in the basic research literature include intervention experimental technology entities, the causal strength of the entity relationship in the basic research literature is determined to be a strong correlation;

[0034] Otherwise, the causal strength of the entity relationship in the basic research literature is determined to be weakly correlated.

[0035] The present invention also provides a query method, comprising:

[0036] Get the target entity to be queried;

[0037] Based on the knowledge graph, determine the relevant knowledge information of the target entity;

[0038] The knowledge graph is determined based on any of the knowledge graph construction methods described above.

[0039] The present invention also provides a knowledge graph construction device, comprising:

[0040] A document acquisition unit, used to acquire a target document set, wherein the target document set includes a plurality of basic research documents;

[0041] An entity acquisition unit, used to perform entity recognition on sentences in each basic research document to obtain entity pairs contained in each basic research document, wherein the entity pairs include entities and entity relationships between the entities;

[0042] A strength determination unit, configured to determine the causal strength of entity relationships between entities included in each basic research document based on the types of entities included in each basic research document;

[0043] A popularity determination unit, used to determine the popularity of entity pairs contained in each basic research document based on the document citation information of each basic research document;

[0044] The graph construction unit constructs a knowledge graph based on the entities, the entity relationships, the causal strengths and the heat.

[0045] The present invention also provides a query device, comprising:

[0046] An acquisition unit, used for acquiring a target entity to be queried;

[0047] A query unit is used to determine relevant knowledge information of the target entity based on a knowledge graph, wherein the knowledge graph is determined based on a knowledge graph construction method as described in any one of the above items.

[0048] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the knowledge graph construction method or query method described above are implemented.

[0049] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described knowledge graph construction methods or query methods.

[0050] The knowledge graph construction and query method, device, electronic device and storage medium provided by the present invention obtain multiple basic research documents, perform entity recognition on the sentences in each basic research document, comprehensively and accurately mine entities and entity relationships between entities, and construct a knowledge graph based on entities, entity relationships, causal strength, and combined with the heat obtained by mining the relevant information based on each basic research document, so that the information in the massive biomedical data can be comprehensively and accurately refined into the knowledge graph, thereby improving the efficiency and reliability of information mining and refining, and reducing the time and cost required for information mining. And the causal strength and heat of the entity relationship are marked in the knowledge graph, making the information display more intuitive. The knowledge graph thus constructed can help users quickly learn the root cause of the disease and provide new insights into its treatment method. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1 It is a schematic diagram of the process of the knowledge graph construction method provided by the present invention;

[0053] Figure 2 It is a schematic diagram of the structure of the knowledge graph provided by the present invention;

[0054] Figure 3 It is a flowchart of step 110 in the knowledge graph construction method provided by the present invention;

[0055] Figure 4 It is a flowchart of step 112 in the knowledge graph construction method provided by the present invention;

[0056] Figure 5 It is a flowchart of step 120 in the knowledge graph construction method provided by the present invention;

[0057] Figure 6 It is a flowchart of step 130 in the knowledge graph construction method provided by the present invention;

[0058] Figure 7 It is a flowchart of the query method provided by the present invention;

[0059] Figure 8 It is a schematic diagram of the structure of the knowledge graph construction device provided by the present invention;

[0060] Fig. 9 It is a structural schematic diagram of the query device provided by the present invention;

[0061] Fig.10 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] At present, with the continuous development of intelligent information service applications, machine learning, as a powerful data analysis and processing tool, has been widely used in biomedical research. Through the application of machine learning algorithms, new associations and patterns can be discovered from large-scale biomedical data, revealing the potential mechanisms and therapeutic targets of diseases, and providing new insights into the root causes of diseases and their treatment methods.

[0064] In order to help machines better understand text data, knowledge graph technology came into being. Knowledge graph is a form of knowledge organization and specification centered on natural language processing (NLP) and combining multiple technologies such as applied mathematics, graphics, and information visualization. Knowledge graph is essentially a semantic network consisting of nodes representing entities and edges representing the relationships between entities. In layman's terms, knowledge graph is a relational network obtained by connecting all different types of information together. It effectively processes, processes, and integrates complex document data into simple and clear "entity, relationship, entity" triplets, and finally aggregates a large amount of knowledge to achieve rapid response and reasoning of knowledge.

[0065] The existing medical literature retrieval database PubMed contains more than 30 million papers, describing the relationship between biological entities such as molecules and diseases. On the one hand, the relationship between entities is very complex, and unstructured text data is not easy to use. In addition, due to the complexity of scientific research and the uneven quality of papers, these entity relationships are not completely credible. On the other hand, researchers not only pay attention to the confirmed entity relationships, but also pay attention to the entity relationships that have signs but need further confirmation, so as to carry out further research and obtain scientific research results. In this regard, an embodiment of the present invention provides a knowledge graph construction method, which assists researchers in making more accurate data-driven business and research decisions by constructing a systematic, structured entity relationship knowledge graph that indicates the strength of causal relationships.

[0066] Figure 1 It is a flow chart of the knowledge graph construction method provided by the present invention, such as Figure 1 As shown, the method includes:

[0067] Step 110, obtaining a target document set, where the target document set includes a plurality of basic research documents;

[0068] Specifically, basic research is the basic scientific research on the normal physiological and pathological mechanisms of the human body. It has the advantage of high causal argument strength and can provide a theoretical basis and scientific support for clinical diagnosis, treatment, and prognosis. Basic research literature refers to literature whose content is related to the above basic research. The target literature collection can be a collection of multiple literature related to the above basic research.

[0069] The target document set may be obtained from a document retrieval database that includes a large number of medical, biological, health or nursing related documents. For example, the target document set may be obtained from document retrieval databases such as PubMed, Web of Science, and MedPeer. This embodiment of the present invention does not specifically limit this.

[0070] When obtaining the target document set, all documents in the document retrieval database can be screened based on the target features, and the target documents involving basic research can be screened out, thereby obtaining the target document set. The target features here refer to features that are related to the documents in the document retrieval database and can determine whether the documents belong to basic research documents, such as document type, publication type (publication_type) of the document, etc. The method of screening all documents in the document retrieval database can be achieved through rule matching, text binary classification, or rule matching combined with text binary classification. The embodiment of the present invention does not make specific limitations on this.

[0071] Step 120, performing entity recognition on the sentences in each basic research document to obtain entity pairs contained in each basic research document, wherein the entity pairs include entities and entity relationships between entities;

[0072] Specifically, after obtaining the target document set, entity recognition is performed on the sentences in the target basic research documents. The target basic research documents here refer to the basic research documents in the target document set, and the sentences here refer to the text information in the target basic research documents. Entities are the most basic elements in the knowledge graph. There are different relationships between different entities, namely entity relationships. When there is a relationship between two entities, the two entities and the entity relationship between them constitute an entity pair.

[0073] In order to efficiently, comprehensively and accurately mine biological entities related to disease pathogenesis and entity relationships between entities from the target basic research literature, entities can be defined in advance. For example, entities can include disease entities, molecular entities, drug entities, pathophysiological mechanism entities, signal pathway entities, organ or tissue entities, cell entities, microbial entities, baseline characteristic entities and experimental technology entities, among which disease entities, molecular entities, drug entities, pathophysiological mechanism entities, signal pathway entities, organ or tissue entities, cell entities, microbial entities and baseline characteristic entities are used as entity nodes in the knowledge graph, and experimental technology entities are used to determine the strength of entity relationships between entities.

[0074] For each type of entity predefined above, a standard dictionary can be pre-constructed for entity extraction and normalization. When pre-constructing a standard dictionary, it can be constructed based on MeSH (Medical Subject Headings), HGNC (HUGO Gene Nomenclature Committee) and internal data accumulated independently by the enterprise, and the embodiment of the present invention does not specifically limit this. For example, when obtaining relevant entities of the standard dictionary from MeSH, you can preset appropriate inclusion and exclusion rules based on target requirements and filter from the MeSH tree structure. Taking the disease entity in the pre-constructed standard dictionary as an example, MeSH contains the subject vocabulary of various diseases. When obtaining disease-related entities from MeSH, you can preset inclusion and exclusion rules that are suitable for the disease entity, filter the disease vocabulary in the MeSH tree structure, and thus construct the disease entity in the standard dictionary.

[0075] Based on the numerous entities in the pre-built standard dictionary, the sentences in the target basic research literature can be entity recognized to obtain all the entities contained in the target basic research literature. After obtaining a large number of entities, it is also necessary to identify the relationship between the entities and associate these entities to obtain the entity relationship between the entities. In order to accurately identify the relationship between entities and improve the reliability of the entity relationship finally obtained, the entity relationship can be defined in advance based on expert experience, and the entity relationship includes but is not limited to: positive (upregulation, activation, promotion), negative (downregulation, inactivation, inhibition), inclusion, treatment, etc.; for example, the entity relationship between molecule A and molecule B includes but is not limited to: molecule A promotes the function of molecule B, molecule A inhibits the function of molecule B; the entity relationship between molecule A and signaling pathway 1 includes but is not limited to: molecule A is a member of signaling pathway 1; the entity relationship between drug A and disease B includes but is not limited to: drug A treats disease B.

[0076] After extracting all entities contained in the target basic research document, the relationship between different entities can be judged based on the pre-defined entity relationship and combined with the sentences in the target basic research document to obtain the entity relationship between entities, thereby obtaining several entity pairs.

[0077] Step 130, determining the causal strength of entity relationships between entities included in each basic research document based on the types of entities included in each basic research document;

[0078] Specifically, the causal strength of entity relationships refers to the strength of the relationships between entities. After obtaining all entities contained in the target basic research literature, it is necessary to discriminate the relationships between the entities. When it is judged that there is no relationship between two entities, it indicates that the two entities are not associated, and the causal strength of the entity relationship between the two entities can be determined to be irrelevant; when there is a relationship between two entities, it indicates that the two entities are associated, and the causal strength of the entity relationship between the two entities can be determined to be relevant.

[0079] In order to better distinguish the causal strength of entity relationships, so that users can quickly learn about the strength of relationships between entities and better understand the root causes of diseases, when the causal strength of entity relationships is related, the causal strength can be further divided into strong correlation and weak correlation. In order to more clearly and accurately obtain the causal strength of entity relationships between entities, it can be judged based on the types of entities contained in the target basic research literature. Here, the type of entity refers to a type of entity that can be used to determine the strength of relationships between entities. For example, the type of entity can be the experimental technology entity in the above-mentioned predefined entities. Experimental technology entities refer to entities related to the main means and main methods and methods adopted in the entire experimental process. It includes many technical entities. Therefore, based on the specific content of experimental technology entities, it can be judged that the causal strength of entity relationships between entities contained in the target basic research literature is strongly correlated or weakly correlated, thereby obtaining accurate causal strength of entity relationships.

[0080] When judging the causal strength of the entity relationship between entities based on the experimental technology entity, the judgment and determination can be made based on the preset judgment rules. The preset judgment rule here can be that the experimental technology entity includes the preset experimental technology entity. Since a relationship has been formed between the two entities, when the experimental technology entity includes the preset experimental technology entity, the causal strength of the entity relationship between the two entities can be determined to be strongly correlated; when the experimental technology entity does not include the preset experimental technology entity, the causal strength of the entity relationship between the two entities can be determined to be weakly correlated. The preset experimental technology entity here is a type of entity in the experimental technology entity that can be used to judge that the causal strength of the entity relationship between two entities is strongly correlated. For example, the preset experimental technology entity can be an intervention experimental technology entity. It should be noted that when judging the causal strength of the entity relationship between entities based on the experimental technology entity, the experimental technology entity and the entity to be judged are extracted based on the same target basic research literature.

[0081] Step 140, based on the document citation information of each basic research document, determine the popularity of the entity pairs contained in each basic research document;

[0082] Specifically, literature citation information refers to relevant information and data about the target basic research literature cited, which can be used to evaluate the influence and usefulness of the target basic research literature. Literature citation information includes but is not limited to at least one of the following: impact factor, JCR (Journal Citation Reports) division, Chinese Academy of Sciences division, citation count, monthly average citation count, etc. When determining the popularity of an entity pair, each target basic research literature can be scored based on the literature citation information to obtain the influence score of each target basic research literature, and then the popularity of the entity pair can be obtained based on the influence scores of all target basic research literature in which the entity pair appears.

[0083] Here, when obtaining the influence score of each target basic research document, we first obtain the literature citation information of each target basic research document, namely, indicators such as impact factor, JCR classification, CAS classification, number of citations and monthly average number of citations. Then, we can score the target basic research document after comprehensively considering all the above indicators. The score can be set from 1 to 5 points, and the score obtained is the influence score of the target basic research document. It is also possible to score each target basic research document based on the above-mentioned indicators to obtain the score corresponding to each indicator. For example, when scoring the target basic research document based on the impact factor, the impact factor can be divided into five levels: 0-20%, 21%-40%, 41%-60%, 61%-80%, and 81%-100%, and each level corresponds to 1 point, 2 points, 3 points, 4 points, and 5 points. When the impact factor of the target basic research document is 57%, the target basic research document can be scored 3 points. After obtaining the score corresponding to each indicator, the sum of the scores corresponding to all indicators is used as the influence score of the target basic research document. The embodiment of the present invention does not specifically limit the method of obtaining the influence score of the target basic research document.

[0084] After obtaining the influence score of each target basic research document, a secondary weighted score can be performed on all target basic research documents where the entity pair appears, thereby obtaining the popularity of the entity pair. A knowledge graph based on the popularity of the entity pair can help users quickly query relevant knowledge information of the entity from the knowledge graph.

[0085] Step 150, construct a knowledge graph based on entities, entity relationships, causal strength and heat.

[0086] Specifically, the knowledge graph is essentially a semantic network consisting of nodes representing entities and edges representing the relationships between entities. After obtaining entities and entity relationships based on the basic research documents in the target document collection, a knowledge graph can be constructed based on entities and entity relationships. Each entity constitutes a node in the knowledge graph, and each entity relationship constitutes an edge connection between entities in the knowledge graph. The specific relationship between entities (i.e., the specific content of the entity relationship) constitutes the attribute of the edge. In order to make the constructed knowledge graph clearer, more comprehensive, and more accurate, a knowledge graph can be constructed based on the causal strength of entity relationships and the heat of entity pairs on the basis of entity-entity relationships. For example, the heat of the entity pair can be represented by the thickness of the edge, and different causal strengths can be represented by different colors, so that users can quickly obtain relevant knowledge information from the knowledge graph.

[0087] Figure 2 is a schematic diagram of the structure of the knowledge graph provided by an embodiment of the present invention, such as Figure 2 As shown in the figure, the entities included in the knowledge graph are disease 1, disease 2, molecule 1, molecule 2, molecule 3, cell 1 and cell 2. Each entity constitutes a node, and the entity pairs included are (molecule 1, relationship 1, cell 1), (cell 1, relationship 2, cell 2), (molecule 1, relationship 3, disease 1), (cell 2, relationship 3, molecule 2), (disease 1, relationship 4, molecule 3) and (molecule 2, relationship 5, disease 2). Each entity relationship constitutes an edge connection, and the specific content of the entity relationship constitutes the attribute of the edge; the thicker the edge, the higher the popularity of the entity pair, such as Figure 2 The heat of the three entity pairs (molecule 1, relation 1, cell 1), (cell 1, relation 2, cell 2), and (cell 2, relation 3, molecule 2) shown in the figure is relatively high; entity pairs with different causal strengths can be distinguished by different colors, such as Figure 2 The causal strength of the entity relationship between molecule 1 and cell 1 shown in is a strong correlation. In addition, the number of basic research papers involving the entity can also be represented by the node size, such as Figure 2 The molecule 1 shown in the figure has the largest node, indicating that the number of basic research papers involving molecule 1 is the largest.

[0088] The knowledge graph construction method provided by the embodiment of the present invention obtains multiple basic research documents, performs entity recognition on the sentences in each basic research document, comprehensively and accurately mines entities and entity relationships between entities, and constructs a knowledge graph based on entities, entity relationships, causal strength, and combined with the heat obtained by mining the relevant information based on each basic research document, so that the information obtained in the massive biomedical data can be comprehensively and accurately refined into the knowledge graph, thereby improving the efficiency and reliability of information mining and refining, and reducing the time and cost required for information mining. The causal strength and heat of the entity relationship are marked in the knowledge graph, making the information display more intuitive.

[0089] The knowledge graph constructed in this way can help users quickly understand the root cause of the disease and provide new insights into its treatment methods.

[0090] Based on the above embodiments, Figure 3 is a flow chart of step 110 in the knowledge graph construction method provided by the present invention, such as Figure 3 As shown, step 110 includes:

[0091] Step 111, obtaining an initial document set;

[0092] Specifically, the initial document set refers to a collection of all documents in a document retrieval database. The initial document set can be obtained from document retrieval databases from different sources, such as PubMed, Web of Science, MedPeer, etc., which is not specifically limited in the embodiment of the present invention.

[0093] It should be understood that the initial document collections obtained from different sources can be merged and then the target document collection can be screened; the initial document collections obtained from different sources can also be screened first, a knowledge graph can be constructed based on the target document collection, and then the knowledge graphs can be combined. For example, after obtaining the initial document collection from PubMed, the initial document collection is screened to obtain the target document collection, and a first knowledge graph can be constructed based on the target document collection; after obtaining the initial document collection from Web of Science, the initial document collection is screened to obtain the target document collection, and a second knowledge graph can be constructed based on the target document collection. The first knowledge graph can be combined with the second knowledge graph to obtain a knowledge graph.

[0094] Step 112, based on the publication type of each initial document in the initial document set, screen out secondary processed documents and clinical research documents from the initial document set to obtain a candidate document set;

[0095] Specifically, primary documents are documents produced directly by people based on their own practical experience in production, scientific research, social activities, etc., also known as original documents; secondary processed documents are the products of primary documents after processing and sorting, that is, the external features of disordered primary documents such as titles, authors, sources, etc. are recorded, or their contents are compressed into indirect, abstracts or digests, and then ordered into document forms such as catalogs, reviews, guidelines, etc. Clinical research is a scientific research activity organized and implemented by multidisciplinary personnel with the diagnosis, treatment, prognosis, etiology and prevention of diseases as the main research content, patients as the main research objects, and medical service institutions as the main research base. Clinical research documents refer to documents whose contents are related to the above scientific research activities. Compared with secondary processed literature and clinical research literature, the content in basic research literature is related to the basic scientific research on the normal physiological and pathological mechanisms of the human body, and has the advantage of high causal argumentation strength. Therefore, when screening the initial literature collection, a preliminary screening can be performed based on the publication type to filter out secondary processed literature and clinical research literature from the initial literature collection, thereby obtaining the target literature collection involving basic research, that is, the candidate literature collection.

[0096] When the initial document set is screened based on the publication type, it can be achieved through rule matching. Each initial document in the initial document set corresponds to at least one publication type. Therefore, the specific rule matching method can be: if the publication type of the initial document includes at least one of the first publication types and does not include any of the second publication types, then the initial document is considered to be a document involving basic research; otherwise, the initial document is considered to be a secondary processed document or a clinical research document, and it is deleted from the initial document set. The first publication type here may include "Journal Article", "Letter"; the second publication type may include "Randomized Controlled Trial", "Controlled Clinical Trial", "Clinical Trial, Phase I", "Clinical Trial, Phase II", "Clinical Trial, Phase III", "Clinical Trial, Phase IV", "Randomized Controlled Trial, Veterinary", "Systematic Review", "Meta-Analysis", "Practice Guideline", and "Guideline".

[0097] When the initial document set is screened based on the publication type, it can also be implemented through a text binary classification model. The text binary classification model is a commonly used text classification model. All initial documents in the initial document set can be divided into two categories according to the publication type, namely basic research documents and non-basic research documents. Each initial document in the initial document set is input into the text binary classification model, and the text binary classification model outputs "yes" or "no" of the initial document being a basic research document. If the initial document is not a basic research document, it will be deleted from the initial document set.

[0098] When screening the initial document collection based on publication type, it can also be achieved by combining rule matching with text binary classification, which will not be described in detail here.

[0099] Since the publication types provided in the literature retrieval database have a certain time lag, in order to ensure the accurate screening of the latest published basic research literature, after the initial literature set is preliminarily screened to obtain the candidate literature set, the candidate literature set can be accurately screened to ensure that the documents in the target literature set are all basic research documents, which is conducive to the subsequent efficient and accurate extraction of entities and entity relationships from the basic research documents in the target literature set, thereby improving the efficiency and accuracy of knowledge graph construction.

[0100] Step 113, screening basic research documents from the candidate document set to construct a target document set.

[0101] Specifically, when accurately screening the candidate document set, the screening can be based on a pre-trained model plus a fine-tuned document classifier. First, the text data of each candidate document in the candidate document set is preprocessed and converted into a form that can be input into the model. The converted text data is input into the document classifier, and the document classifier is used for classification to determine whether the candidate document belongs to a basic research document. If the candidate document does not belong to a basic research document, it will be deleted from the candidate document set, thereby constructing a target document set based on the final candidate document set.

[0102] Exemplarily, when the initial document collection is obtained from PubMed, the above pre-trained model can use the PubMedBert pre-trained model, which is a BERT (Bidirectional Encoder Representation from Transformers) model trained in the medical field. Its pre-training data comes from the abstracts and full-text articles of PubMed, so the model can understand the professional terms and language in the medical field. When obtaining the pre-trained model, you can download and import the PubMedBert pre-trained model from the model library of the Hugging Face website, which includes pre-trained weights and corresponding configuration files. When importing the model, you need to specify the model and configuration file to be used.

[0103] After obtaining the PubMedBert pre-trained model, the pre-trained model can be used as a basis for fine-tuning on the target dataset. The target dataset here is the training data automatically constructed based on the above rule matching method. The literature data that meets the above rules is added as positive examples, and those that do not meet the rules are added as negative examples. There is no need to manually annotate the training data, thereby saving a lot of time and resources. In order to enable the model to classify the training samples and output whether the training samples belong to basic research literature, a fully connected output layer can be added on top of the pre-trained layer of the model to classify the training samples and determine whether the training samples belong to basic research literature. The training samples here refer to the sample documents in the training data. During the fine-tuning training, the data in the target dataset is first pre-processed and converted into a form that can be input into the model to obtain the converted text data, and then the text data is input into the pre-trained model. The text data is processed by the input layer and Transformer encoder of the pre-trained model, and features are extracted on the pre-trained layer, and then the extracted features are input into the fully connected layer for classification and judgment. The weights and biases of the fully connected layer are updated according to the training data to minimize the loss function on the target dataset, thereby obtaining a document classifier based on the pre-trained model plus fine-tuning. It is understandable that during fine-tuning training, corresponding training strategies and hyperparameters can be used to optimize model performance.

[0104] The embodiment of the present invention first screens the initial document set based on the publication type to obtain a candidate document set, and then accurately screens the candidate document set through a document classifier based on a pre-trained model and fine-tuning, thereby ensuring that the documents in the obtained target document set are all basic research documents, which is conducive to the subsequent efficient and accurate extraction of entities and entity relationships from each basic research document in the target document set, thereby improving the efficiency and accuracy of knowledge graph construction.

[0105] Based on the above embodiments, Figure 4 is a flow chart of step 112 in the knowledge graph construction method provided by the present invention, such as Figure 4 As shown, step 112 includes:

[0106] Step 1121, based on the publication type of each initial document in the initial document set, secondary processing documents and clinical research documents are screened out from the initial document set to obtain a first document set;

[0107] Specifically, secondary processed documents are the product of processing and organizing primary documents, and their content is not as comprehensive as that of primary documents. If entity recognition and extraction are performed based on secondary processed documents, it is easy to cause the final constructed knowledge graph to be incomplete. The content in clinical research documents is irrelevant to the basic scientific research on the normal physiological and pathological mechanisms of the human body, and has little effect on constructing the disease knowledge graph. Therefore, when screening the initial documents, it is necessary to delete the secondary processed documents and clinical research documents from the initial document set, so as to obtain the first document set involving basic research.

[0108] When the initial document set is screened to obtain the first document set, it can be implemented based on rule matching, or it can be implemented by text binary classification, or it can be implemented by rule matching combined with text binary classification. The embodiment of the present invention does not make specific limitations on this. The specific implementation process can refer to the embodiment of step 112 above, which will not be repeated here.

[0109] Since the publication types provided in the literature retrieval database have a certain time lag, for example, for a document whose publication type provided by PubMed is only "Journal Article" and whose publication time is less than 12 months, it is generally considered that the publication type of the document is still to be updated. Therefore, in order to avoid errors in screening documents of this type, the document classifier in the above embodiment can be directly used to screen documents of this type, thereby ensuring that the documents finally screened are all basic research documents.

[0110] Step 1122, based on the publication type and publication time of each first document in the first document collection, select a second document from the first document collection, the second document being the first document whose publication type is to be updated;

[0111] Specifically, the first document refers to each basic research document included in the first document set. After the initial document set is screened to obtain the first document set, since there may be first documents in the first document set whose publication type is to be updated, the publication type of such first documents is not accurate. Therefore, in order to accurately screen such first documents, all first documents in the first document set can be screened based on the publication type and publication time, and the first documents whose publication type and publication time meet the preset conditions are screened out to obtain the second document. The first document whose publication type and publication time meet the preset conditions here is the first document whose publication type is to be updated, that is, the above-mentioned second document. It should be noted that the preset condition here can be that the publication type is only "Journal Article" and the publication time is within 12 months.

[0112] When the first document set is screened based on publication time and publication type, the publication time and publication type of each first document in the first document set can be obtained, and the publication time and publication type can be compared with preset conditions. If the preset conditions are met, it is considered that the publication type of the first document needs to be updated, and the first document can be used as the second document.

[0113] Step 1123, input the second document into the document classifier, and obtain the document type of the second document output by the document classifier;

[0114] Specifically, the document classifier here can be the document classifier based on the pre-trained model and fine-tuning in the above embodiment. After the second document is screened, the text data of the second document can be pre-processed first, converted into a form that can be input into the model, and the converted text data is input into the document classifier, and classified by the document classifier to obtain the document type of the second document output by the document classifier. The document type here can be used to determine whether the second document belongs to a basic research document. The document type can include several types, and the document classifier can classify the second document based on the content of the second document, and output the document type of the second document, which may be one or more document types.

[0115] Step 1124, delete the second document whose document type is non-basic research document from the first document set to obtain a candidate document set.

[0116] Specifically, when judging whether the second document is a basic research document or a non-basic research document, the judgment can be made based on the document type of the second document output by the document classifier. When the document type of the second document includes at least one of the first document types and does not include any of the second document types, the second document is considered to be a basic research document; otherwise, the second document is considered to be a non-basic research document. When the second document is determined to be a non-basic research document, the second document is deleted from the first document set, thereby obtaining a candidate document set. The first document type here can include "Journal Article", "Letter"; the second document type can include "Randomized Controlled Trial", "Controlled Clinical Trial", "Clinical Trial, Phase I", "Clinical Trial, Phase II", "Clinical Trial, Phase III", "Clinical Trial, Phase IV", "RandomizedControlled Trial, Veterinary", "Systematic Review", "Meta-Analysis", "PracticeGuideline", "Guideline".

[0117] The embodiment of the present invention first screens the initial document set based on the publication type, deletes the secondary processed documents and clinical research documents in the initial document set, obtains the first document set, and then screens the first document set based on the publication type and publication time, thereby obtaining the first document of the publication type to be updated, and uses the first document as the second document. Finally, the second document is classified and judged by a document classifier, which can avoid screening errors in the documents of the publication type to be updated, avoid missing the basic research documents therein, and avoid mistaking the non-basic research documents therein as the target basic research documents, thereby ensuring that the entities and entity relationships subsequently extracted from the basic research documents of the target document set are more accurate and comprehensive.

[0118] Based on any of the above embodiments, Figure 5 is a flow chart of step 120 in the knowledge graph construction method provided by the present invention, such as Figure 5 As shown, step 120 includes:

[0119] Step 121, obtaining the title text and abstract text of each basic research document;

[0120] Specifically, when extracting entities and entity relationships from the text of each basic research document in the target document set, since the overall text of each basic research document is very long and most of it is redundant information, in order to improve the efficiency of entity recognition extraction, the text extraction range can be set to the title text and abstract text. After obtaining the text information of each basic research document, the title information and abstract information can be extracted based on the text information to obtain the title text and abstract text.

[0121] In order to further improve the efficiency of entity mining extraction, the following steps 122 and 123 can be executed to classify and judge the title text and the summary text, distinguish the main information from the title text and the summary text, retain the sentences containing the main information in the title text and the summary text, and discard the sentences containing the secondary information in the title text and the summary text, so as to perform entity recognition on the sentences in the retained title text and the summary text, thereby further improving the efficiency of entity extraction.

[0122] Step 122, input the title text and the abstract text into a sentence classifier, and obtain the sentence type of each sentence in the title text and the abstract text output by the sentence classifier;

[0123] Specifically, the sentence classifier is obtained by training the pre-trained model based on manually annotated training data, and is used to determine the sentence type of each sentence in the title text and the summary text. The sentence types here include sentences to be identified and sentences that do not need to be identified, wherein the sentences to be identified refer to the sentences containing the main information in the title text and the summary text, and the sentences that do not need to be identified refer to the sentences that do not contain the main information in the title text and the summary text, and the above-mentioned main information refers to the text information related to the entities and entity relationships to be extracted.

[0124] The obtained title text and abstract text of the target basic research document are input into the sentence classifier, and classified by the sentence classifier to determine the sentence type of each sentence in the title text and abstract text of the target basic research document. If the sentence contains main information, the sentence is retained and the sentence type of the sentence is output as the sentence to be identified; if the sentence does not contain main information, the sentence is discarded.

[0125] Before executing step 122, a sentence classifier can be trained. Specifically, the sentence classifier can be trained in the following manner: obtain a pre-trained model, such as a PubMedBert pre-trained model, collect a large number of sample title texts and abstract texts, manually annotate and identify the sentence type of each sentence in the sample title texts and abstract texts, and train the pre-trained model based on the sample title texts and abstract texts and the manually annotated sentence types to obtain a sentence classifier.

[0126] Step 123, performing entity recognition on the sentence type to be recognized to obtain entity pairs contained in each basic research document.

[0127] Specifically, the sentences to be identified refer to the sentences containing the main information in the title text and the abstract text that are retained after classification and judgment by the sentence classifier. When performing entity recognition on this type of sentence, firstly, based on the pre-built standard dictionary, the rule matching scheme can be used to extract entities, and all entities that have appeared in the pre-built standard dictionary can be extracted to ensure the accuracy of the entity extraction based on the standard dictionary; secondly, since there are still many entities or entity aliases that may not be included in the pre-built standard dictionary, the named entity recognition model can be used to further perform entity recognition extraction on the sentences to be identified, and the semantics in the sentences can be parsed by the named entity recognition model, so as to extract the entities required for constructing the knowledge graph, and ensure that the entities finally extracted are more comprehensive and accurate.

[0128] It should be noted that the named entity recognition model here is a basic information extraction task model in the field of natural language processing, which is used to identify named entities from a given text and classify the entities. Since there are many types of entities to be extracted, including disease entities, molecular entities, drug entities, pathophysiological mechanism entities, signal pathway entities, organ or tissue entities, cell entities, microbial entities, baseline feature entities, and experimental technology entities, in order to achieve the extraction of multiple entities, a multi-task architecture can be used on the named entity recognition model. The input of the named entity recognition model is the sentence to be recognized in the title text and the abstract text, and its output is the result of multiple tasks, each task is the extraction of an entity, such as disease entity extraction task, molecular entity extraction task, etc.

[0129] The embodiment of the present invention obtains the title text and abstract text of each basic research document and sets the text extraction scope to the title text and abstract text, thereby improving the efficiency of entity recognition and entity relationship extraction. By classifying and judging the title text and abstract text through a sentence classifier, the scope of text extraction can be further narrowed, thereby improving the efficiency of entity and entity relationship extraction, and further improving the efficiency of knowledge graph construction.

[0130] Based on the above embodiment, since the title text and abstract text contain a large number of pronouns, these pronouns may also represent certain entities. In order to further ensure the comprehensiveness and accuracy of entity and entity relationship recognition extraction, a new task can be added to the named entity recognition model in the multi-task architecture, that is, extracting pronouns in the title text and abstract text. To this end, entity recognition is performed in step 120 or step 123 to obtain entity pairs contained in each basic research document, including:

[0131] Perform entity recognition on the sentence to obtain the entities and pronouns in the sentence;

[0132] Input the entity and the pronoun into the reference relationship classifier, and obtain the reference relationship between the entity and the pronoun output by the reference relationship classifier;

[0133] When the reference relationship is yes, the pronouns in the sentence are replaced with the corresponding entities to obtain the optimized sentence;

[0134] The optimized statement and entities in the optimized statement are input into the entity relationship classifier to obtain the entity relationship between the entities output by the entity relationship classifier.

[0135] Specifically, when performing entity recognition on sentences in the title text and abstract text, entities can be extracted based on a pre-built standard dictionary using rule matching; then, a named entity recognition model with a multi-task architecture can be used for extraction to obtain all entities and pronouns contained in the sentence. After obtaining a large number of entities and related pronouns, it is necessary to associate entities with entities and entities with pronouns. First, for all pronouns and entities, a reference relationship classifier can be used to determine whether each pronoun and each entity have a reference relationship. The identified entities and pronouns as well as the title text and abstract text are input into the reference relationship classifier. Here, the reference relationship classifier uses a binary classification method to classify the relationship between pronouns and entities, and determines whether the pronoun and entity have a reference relationship. If the reference relationship is yes, it indicates that the pronoun refers to the entity, that is, the pronoun is equivalent to the entity, and the pronoun can be replaced by the entity; if the reference relationship is no, it indicates that the pronoun is not equivalent to the entity, and the pronoun can be deleted to obtain an optimized sentence.

[0136] Secondly, after associating entities with pronouns, entities need to be associated with entities to obtain entity relationships between entities. Since there may be many different representations of the same entity in the same target basic research literature, such as the full name and abbreviation of an entity, after obtaining the optimized statement, the relationship between different entities in the same entity category can be judged based on the entity relationship classifier. For example, the optimized statement and entity A and entity B in the optimized statement are input into the entity relationship classifier, and the entity relationship between entity A and entity B output by the entity relationship classifier can be obtained. The entity relationship here can be any one of the following entity relationships: entity A is equal to entity B, entity A contains entity B, entity B contains entity A, and entity A and entity B are unrelated. Since there are many types of relationships between different entities in the same entity category, an entity relationship classifier based on a multi-classification model can be used to classify and judge this type of entity relationship.

[0137] Finally, after associating entities with pronouns and entities with different representations in the same category, it is necessary to discriminate the relationships between all entities to obtain the entity relationships between entities. Here, the relationships between all entities are discriminated, including the entity relationship discrimination between entities of the same entity category, such as disease entity 1 and disease entity 2; it also includes the entity relationship discrimination between entities of different entity categories, such as an entity in the disease and an entity in the drug. After obtaining the optimized statement, the optimized statement and the two entities of different entity categories in the optimized statement are input into the entity relationship classifier, and the entity relationship between the two entities output by the entity relationship classifier can be obtained. In order to judge the entity relationships between all entities, an entity relationship classifier based on the token-classification model can be used for implementation.

[0138] In an embodiment of the present invention, by first processing the pronouns in the sentence, and when the reference relationship between the pronouns and the entities is yes, replacing the pronouns in the sentence with corresponding entities, the semantics of the obtained optimized sentence can be made clearer, and it is convenient to subsequently identify the entity relationship between entities based on the optimized sentence, so that the entity relationship can be identified and mined more efficiently, comprehensively and accurately.

[0139] Based on the above embodiment, entity recognition is performed in step 120 or step 123 to obtain entity pairs contained in each basic research document, and further includes:

[0140] Fill the optimized statement and the entities in the optimized statement into the relational query template to obtain the relational query statement;

[0141] The relational query statement is input into the question-answering language model to obtain the entity relationship output by the question-answering language model.

[0142] Specifically, when performing entity recognition and mining to obtain entities and entity relationships, the question-answering language model can also be used to predict. The question-answering language model of an embodiment of the present invention can be an LLM (Large Language Model) language model. After obtaining the optimized statement, the optimized statement and the entities extracted from the optimized statement are filled into the relational query template, that is, according to the optimized statement and the entities extracted from the optimized statement, the relevant content in the relational query template is replaced accordingly to obtain a relational query statement, and the relational query statement is used as the input of the question-answering language model, and is parsed by the question-answering language model based on a given format to obtain the entity relationship output by the question-answering language model.

[0143] Exemplarily, the relationship query template may be:

[0144] "Below is a biomedical section and the entities it contains. Please determine the possible binary relationships between entities based on the text semantics. Input [text content] input [entity content], possible entity relationships: [entity relationship description]. Please output the results line by line in the following format "(entity 1, entity 2, relationship)"."

[0145] In the above relational query template, the content in the brackets needs to be replaced accordingly according to the optimized statement and the extracted entity to obtain the relational query statement.

[0146] In order to further obtain more comprehensive and accurate entity relationships, the entity relationships output by the question-answering language model may be merged with the entity relationships obtained based on the entity relationship classifier in the above embodiment.

[0147] Based on any of the above embodiments, Figure 6 is a flow chart of step 130 in the knowledge graph construction method provided by the present invention, such as Figure 6 As shown, step 130 includes:

[0148] Step 131, if the type of entity contained in the basic research document includes an intervention experiment technology entity, the causal strength of the entity relationship in the basic research document is determined to be a strong correlation;

[0149] Step 132, otherwise the causal strength of the entity relationship in the basic research literature is determined to be weakly correlated.

[0150] Specifically, the causal strength of entity relationships refers to the strength of relationships between entities. The above-mentioned entity type refers to a type of entity that can be used to determine the strength of relationships between entities. In an embodiment of the present invention, the type of entity can be an experimental technology entity in a predefined entity. When there is an entity relationship between entities, the causal strength of the entity relationship between the entities can be determined to be related. In this case, the causal strength of the entity relationship can be further divided into strong correlation and weak correlation based on the experimental technology entity, thereby helping users quickly learn about the strength of relationships between entities and better understand the root cause of the disease based on the knowledge graph.

[0151] The above-mentioned intervention experiment technical entity refers to a type of experimental technical entity that can be used to determine that the causal strength of the entity relationship between two entities is strongly correlated. The intervention experiment technical entity includes but is not limited to: CRISPR / Cas9 (gene therapy), RNA interference, plasmid transfection, and viral vector-mediated overexpression.

[0152] It is understandable that when acquiring entities and entity relationships, identification and extraction are performed for each target basic research document. Therefore, when judging the causal strength of entity relationships, the entity relationships between entities contained in the same target basic research document are also judged. When judging the causal strength of entity relationships based on experimental technology entities including intervention experimental technology entities, if an entity relationship is formed between two entities in the same target basic research document, and the experimental technology entity includes at least one of the intervention experimental technology entities, the causal strength of the entity relationship between the two entities is determined to be strongly correlated; if an entity relationship is formed between two entities in the same target basic research document, and the experimental technology entity does not include any of the intervention experimental technology entities, the causal strength of the entity relationship between the two entities is determined to be weakly correlated; if no entity relationship is formed between two entities in the same target basic research document, the causal strength of the entity relationship between the two entities is determined to be irrelevant.

[0153] Based on any of the above embodiments, Figure 7 It is a flowchart of the query method provided by the present invention, such as Figure 7 As shown, the method includes:

[0154] Step 710, obtaining the target entity to be queried;

[0155] Step 720, based on the knowledge graph, determine the relevant knowledge information of the target entity, and the knowledge graph is determined based on the above-mentioned knowledge graph construction method.

[0156] Specifically, for basic research literature from various sources, the high-throughput information on the pathogenesis of diseases can be efficiently, comprehensively and accurately sorted out through the methods provided in the above embodiments, and a causal argumentation evaluation system can be constructed based on experimental technology entities, thereby constructing a disease knowledge graph to help users quickly understand the root cause of the disease and provide new insights into its treatment methods. On this basis, a query system can be constructed to enable users to quickly find disease-related knowledge information.

[0157] The user can input the target entity to be queried through a user terminal in the form of a mobile phone, computer, tablet computer, etc., and send the target entity to be queried to the server of the query system. The target entity to be queried here refers to an entity related to the normal physiological mechanism and pathological mechanism of the human body, which can be any one or more of a disease entity, a molecular entity, a drug entity, a pathophysiological mechanism entity, a signal pathway entity, an organ or tissue entity, a cell entity, a microbial entity, and a baseline characteristic entity. The embodiment of the present invention does not specifically limit this. The knowledge graph contains the above-mentioned target entity to be queried.

[0158] After receiving the target entity to be queried, a search is performed based on the constructed knowledge graph to locate the node corresponding to the target entity and other nodes or sub-nodes at all levels that are connected to the node, and a local graph containing the nodes of the above target entity and their connection relationships is intercepted from the knowledge graph as the relevant knowledge information of the target entity and returned to the user terminal for the user to view.

[0159] The query method provided in the embodiment of the present invention can help users quickly understand the root cause of the disease and provide new insights into disease treatment methods by realizing rapid retrieval and query of relevant knowledge information of target entities related to the normal physiological mechanisms and pathological mechanisms of the human body based on a comprehensive and accurate knowledge graph, so as to enable users to make more accurate data-driven business and research decisions.

[0160] Based on any of the above embodiments, Figure 8 is a schematic diagram of the structure of the knowledge graph construction device provided by the present invention, such as Figure 8 As shown, the device comprises:

[0161] A document acquisition unit 810 is used to acquire a target document set, where the target document set includes a plurality of basic research documents;

[0162] An entity acquisition unit 820 is used to perform entity recognition on sentences in each basic research document to obtain entity pairs contained in each basic research document, where the entity pairs include entities and entity relationships between entities;

[0163] A strength determination unit 830, for determining the causal strength of entity relationships between entities included in each basic research document based on the types of entities included in each basic research document;

[0164] A popularity determination unit 840 is used to determine the popularity of the entity pairs contained in each basic research document based on the document citation information of each basic research document;

[0165] The graph construction unit 850 constructs a knowledge graph based on entities, entity relationships, causal strength and heat.

[0166] The knowledge graph construction device provided by the embodiment of the present invention obtains multiple basic research documents, performs entity recognition on the sentences in each basic research document, comprehensively and accurately mines entities and entity relationships between entities, and constructs a knowledge graph based on entities, entity relationships, causal strength, and combined with the heat obtained by mining the relevant information based on each basic research document, so that the information obtained in the massive biomedical data can be comprehensively and accurately refined into the knowledge graph, thereby improving the efficiency and reliability of information mining and refining, and reducing the time and cost required for information mining. The causal strength and heat of the entity relationship are marked in the knowledge graph, making the information display more intuitive.

[0167] The knowledge graph constructed in this way can help users quickly understand the root cause of the disease and provide new insights into its treatment methods.

[0168] Based on any of the above embodiments, the document acquisition unit 810 includes:

[0169] An initial document acquisition subunit is used to acquire an initial document set;

[0170] A candidate document acquisition subunit is used to screen out secondary processed documents and clinical research documents from the initial document collection based on the publication type of each initial document in the initial document collection to obtain a candidate document collection;

[0171] The target document construction subunit is used to screen basic research documents from the candidate document collection to construct the target document collection.

[0172] Based on any of the above embodiments, the candidate document acquisition subunit is used for:

[0173] Based on the publication type of each initial document in the initial document collection, secondary processed documents and clinical research documents are screened out from the initial document collection to obtain a first document collection;

[0174] Based on the publication type and publication time of each first document in the first document collection, a second document is selected from the first document collection, where the second document is the first document whose publication type is to be updated;

[0175] Inputting the second document into the document classifier, and obtaining the document type of the second document output by the document classifier;

[0176] The second document whose document type is non-basic research document is deleted from the first document set to obtain a candidate document set.

[0177] Based on any of the above embodiments, the entity acquisition unit 820 includes:

[0178] Obtain subunits, obtain the title text and abstract text of each basic research document;

[0179] The classification subunit inputs the title text and the abstract text into a sentence classifier to obtain the sentence type of each sentence in the title text and the abstract text output by the sentence classifier;

[0180] The recognition subunit is used to perform entity recognition on the sentences of the sentence type to be recognized, and obtain the entity pairs contained in each basic research document.

[0181] Based on any of the above embodiments, the entity acquisition unit 820 or the identification subunit is used to:

[0182] Perform entity recognition on the sentence to obtain the entities and pronouns in the sentence;

[0183] Input the entity and the pronoun into the reference relationship classifier, and obtain the reference relationship between the entity and the pronoun output by the reference relationship classifier;

[0184] When the reference relationship is yes, the pronouns in the sentence are replaced with the corresponding entities to obtain the optimized sentence;

[0185] The optimized statement and entities in the optimized statement are input into the entity relationship classifier to obtain the entity relationship between the entities output by the entity relationship classifier.

[0186] Based on any of the above embodiments, the entity acquisition unit 820 or the identification subunit is further configured to:

[0187] Fill the optimized statement and the entities in the optimized statement into the relational query template to obtain the relational query statement;

[0188] The relational query statement is input into the question-answering language model to obtain the entity relationship output by the question-answering language model.

[0189] Based on any of the above embodiments, the strength determination unit 830 is used to:

[0190] If the types of entities contained in the basic research literature include intervention experimental technology entities, the causal strength of the entity relationship in the basic research literature is determined to be a strong correlation;

[0191] Otherwise, the causal strength of the entity relationship in the basic research literature is determined to be weakly correlated.

[0192] Based on any of the above embodiments, Fig. 9 is a schematic diagram of the structure of the query device provided by the present invention, such as Fig. 9 As shown, the device comprises:

[0193] An acquisition unit 910 is used to acquire a target entity to be queried;

[0194] The query unit 920 is used to determine the relevant knowledge information of the target entity based on the knowledge graph, and the knowledge graph is determined based on the above-mentioned knowledge graph construction method.

[0195] The query device provided in the embodiment of the present invention can help users quickly learn the root cause of the disease and provide new insights into disease treatment methods by realizing rapid retrieval and query of relevant knowledge information of target entities related to the normal physiological mechanisms and pathological mechanisms of the human body based on a comprehensive and accurate knowledge graph, so as to enable users to make more accurate data-driven business and research decisions.

[0196] Fig.10 An example of a physical structure diagram of an electronic device is shown in FIG. Fig.10 As shown, the electronic device may include: a processor 1010, a communication interface 1020, a memory 1030 and a communication bus 1040, wherein the processor 1010, the communication interface 1020 and the memory 1030 communicate with each other through the communication bus 1040. The processor 1010 may call the logic instructions in the memory 1030 to execute the knowledge graph construction method, which includes: obtaining a target document set, the target document set includes a plurality of basic research documents; performing entity recognition on the sentences in each basic research document to obtain the entity pairs contained in each basic research document, the entity pairs including entities and entity relationships between entities; determining the causal strength of the entity relationships between entities contained in each basic research document based on the types of entities contained in each basic research document; determining the heat of the entity pairs contained in each basic research document based on the document citation information of each basic research document; and constructing a knowledge graph based on entities, entity relationships, causal strength and heat.

[0197] In addition, the processor 1010 can call the logic instructions in the memory 1030 to execute the query method, which includes: obtaining the target entity to be queried; based on the knowledge graph, determining the relevant knowledge information of the target entity, and the knowledge graph is determined based on the above-mentioned knowledge graph construction method.

[0198] In addition, the logic instructions in the above-mentioned memory 1030 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk.

[0199] On the other hand, the present invention also provides a computer program product, which includes a computer program, and the computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the knowledge graph construction method provided by the above-mentioned methods, and the method includes: obtaining a target document set, the target document set includes multiple basic research documents; performing entity recognition on sentences in each basic research document to obtain entity pairs contained in each basic research document, and the entity pairs include entities and entity relationships between entities; based on the types of entities contained in each basic research document, determining the causal strength of the entity relationships between entities contained in each basic research document; based on the document citation information of each basic research document, determining the popularity of the entity pairs contained in each basic research document; and constructing a knowledge graph based on entities, entity relationships, causal strength and popularity.

[0200] In addition, the computer can also execute the query method provided by the above methods, which includes: obtaining the target entity to be queried; determining the relevant knowledge information of the target entity based on the knowledge graph, and the knowledge graph is determined based on the above knowledge graph construction method.

[0201] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the knowledge graph construction method provided by the above-mentioned methods, the method comprising: obtaining a target document set, the target document set comprising multiple basic research documents; performing entity recognition on sentences in each basic research document to obtain entity pairs contained in each basic research document, the entity pairs comprising entities and entity relationships between entities; determining the causal strength of the entity relationships between entities contained in each basic research document based on the types of entities contained in each basic research document; determining the heat of the entity pairs contained in each basic research document based on the document citation information of each basic research document; and constructing a knowledge graph based on entities, entity relationships, causal strength and heat.

[0202] When the computer program is executed by the processor, it is implemented to execute the query method provided by the above methods, which includes: obtaining the target entity to be queried; based on the knowledge graph, determining the relevant knowledge information of the target entity, and the knowledge graph is determined based on the above knowledge graph construction method.

[0203] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0204] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0205] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A knowledge graph construction method, characterized in that: include: Acquire a target document set, wherein the target document set includes a plurality of basic research documents; Performing entity recognition on sentences in each basic research document to obtain entity pairs contained in each basic research document, wherein the entity pairs include entities and entity relationships between the entities; Based on the types of entities included in each basic research document, determining the causal strength of the entity relationship between the entities included in each basic research document; Based on the literature citation information of each basic research literature, determining the popularity of the entity pairs contained in each basic research literature; Constructing a knowledge graph based on the entities, the entity relationships, the causal strengths, and the heat; The step of obtaining a target document set includes: Get the initial document collection; Based on the publication type of each initial document in the initial document set, secondary processed documents and clinical research documents are screened out from the initial document set to obtain a candidate document set; Basic research documents are screened from the candidate document set to construct the target document set.

2. The knowledge graph construction method according to claim 1, characterized in that: Based on the publication type of each initial document in the initial document set, secondary processed documents and clinical research documents are screened out from the initial document set to obtain a candidate document set, including: Based on the publication type of each initial document in the initial document collection, secondary processed documents and clinical research documents are screened out from the initial document collection to obtain a first document collection; Based on the publication type and publication time of each first document in the first document collection, selecting a second document from the first document collection, where the second document is the first document of the publication type to be updated; Inputting the second document into a document classifier to obtain the document type of the second document output by the document classifier; The second document whose document type is non-basic research document is deleted from the first document set to obtain the candidate document set.

3. The knowledge graph construction method according to claim 1, characterized in that: The entity recognition is performed on the sentences in each basic research document to obtain the entity pairs contained in each basic research document, including: Obtain the title text and abstract text of each basic research document; Inputting the title text and the summary text into a sentence classifier to obtain the sentence type of each sentence in the title text and the summary text output by the sentence classifier; Entity recognition is performed on the sentences of the sentence type to be recognized to obtain entity pairs contained in each basic research document.

4. The knowledge graph construction method according to claim 1 or 3, characterized in that: The entity recognition is performed to obtain entity pairs contained in each basic research document, including: Performing entity recognition on the sentence to obtain entities and pronouns in the sentence; Inputting the entity and the pronoun into a reference relationship classifier, and obtaining a reference relationship between the entity and the pronoun output by the reference relationship classifier; In the case where the reference relationship is yes, the pronouns in the sentence are replaced with corresponding entities to obtain an optimized sentence; The optimization statement and entities in the optimization statement are input into an entity relationship classifier, and entity relationships between the entities output by the entity relationship classifier are obtained.

5. The knowledge graph construction method according to claim 4, characterized in that: The entity recognition is performed to obtain entity pairs contained in each basic research document, and further includes: Filling the optimization statement and entities in the optimization statement into a relational query template to obtain a relational query statement; The relationship query statement is input into a question-answering language model to obtain the entity relationship output by the question-answering language model.

6. The knowledge graph construction method according to claim 1, characterized in that: The determining, based on the types of entities included in each basic research document, the causal strength of the entity relationship between the entities included in each basic research document comprises: If the types of entities contained in the basic research literature include intervention experimental technology entities, the causal strength of the entity relationship in the basic research literature is determined to be a strong correlation; Otherwise, the causal strength of the entity relationship in the basic research literature is determined to be weakly correlated.

7. A query method, characterized in that: include: Get the target entity to be queried; Based on the knowledge graph, determine the relevant knowledge information of the target entity; The knowledge graph is determined based on the knowledge graph construction method as described in any one of claims 1 to 6.

8. A knowledge graph construction device, characterized in that: include: A document acquisition unit, used to acquire a target document set, wherein the target document set includes a plurality of basic research documents; An entity acquisition unit, used to perform entity recognition on sentences in each basic research document to obtain entity pairs contained in each basic research document, wherein the entity pairs include entities and entity relationships between the entities; A strength determination unit, configured to determine the causal strength of entity relationships between entities included in each basic research document based on the types of entities included in each basic research document; A popularity determination unit, used to determine the popularity of entity pairs contained in each basic research document based on the document citation information of each basic research document; A graph construction unit, constructing a knowledge graph based on the entity, the entity relationship, the causal strength and the heat; The document acquisition unit is specifically used for: Get the initial document collection; Based on the publication type of each initial document in the initial document set, secondary processed documents and clinical research documents are screened out from the initial document set to obtain a candidate document set; Basic research documents are screened from the candidate document set to construct the target document set.

9. A query device, characterized in that: include: An acquisition unit, used for acquiring a target entity to be queried; A query unit, used to determine relevant knowledge information of the target entity based on a knowledge graph, wherein the knowledge graph is determined based on the knowledge graph construction method as described in any one of claims 1 to 6.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, it implements the knowledge graph construction method according to any one of claims 1 to 6 or the query method according to claim 7.

Images