A method and system for constructing a knowledge graph based on metabolic diseases

Abstract

The present invention relates to a method and system for acquiring a knowledge graph based on metabolic diseases. First, data related to metabolic diseases are acquired, trigger words and their corresponding event types in the text of metabolic disease related data are identified, and then corresponding questions are constructed for each event trigger word and event type. The questions are spliced ​​with the context and input into the model to extract other components of the event, obtain a complete event description, and then determine the relationship between each complete event. Finally, a knowledge graph of metabolic diseases is constructed based on the extracted information. The above method of the present application makes full use of clinical information to construct a knowledge graph and constructs a more complete knowledge graph of metabolic diseases. It can help scientific researchers or medical practitioners to quickly and conveniently query relevant knowledge, assist clinicians in making scientific decisions, and on the other hand, it can also provide data support for various intelligent services related to metabolic diseases.

Description

Technical Field

[0001] The present invention relates to the field of constructing a knowledge graph of facts and principles, and in particular to a method and system for constructing a knowledge graph of facts and principles based on metabolic diseases. Background Art

[0002] With the rapid development of knowledge graphs, medical knowledge graphs are also a research hotspot. Most traditional knowledge graphs are entity-centered, and as another important element to describe the real world, events and the relationship between events, there are few systematic definitions in traditional knowledge graphs. In this context, some scholars have proposed the concept of event graphs. Event graphs are centered on events and focus on describing the relationship between events (causal relationships, transitional relationships, sequential relationships, conditional relationships, and other relationships). Unlike traditional knowledge graphs that focus on describing relatively static and slowly changing deterministic knowledge between entities, the knowledge described in the event graph is dynamic and has a certain transfer probability. It is also very important to describe the laws of event development for knowledge reasoning. However, there are few medical event knowledge graphs at present. Fully mining events in the medical field can help analyze the causes of diseases and formulate relevant prevention and control measures. Based on this, the present invention proposes a method and system for constructing a event knowledge graph based on metabolic diseases in the medical field. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for constructing a knowledge graph based on metabolic diseases, so as to fully explore events in the medical field, thereby helping to analyze the causes of the disease and formulate relevant prevention and treatment measures.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A method for acquiring a knowledge graph based on metabolic diseases, comprising:

[0006] Acquiring metabolic disease related data, wherein the metabolic disease related data includes structured data and unstructured data;

[0007] Extracting trigger words and event types for each event in the unstructured data using a sequence labeling model;

[0008] According to each of the extracted trigger words and the corresponding event type, a question about the role of event parameters is generated, and according to the question, the event parameters are obtained using a model based on MRC; the event parameters are event parameters of the event corresponding to the trigger word;

[0009] Obtaining a complete description of the event according to the trigger word, the event type, the event parameter role, and the event parameter of each event, which is recorded as the first complete event;

[0010] According to the plurality of the first complete events, the relationship between the plurality of the first complete events is obtained by using an event relationship classification model, and the plurality of the first complete events are associated according to the relationship between the events to obtain a first-level metabolic disease knowledge graph;

[0011] Extracting the structured data using regular expressions to obtain a complete description of each event in the structured data, recording it as a second complete event, and associating multiple second complete events to obtain a second-level metabolic disease knowledge graph;

[0012] Obtaining a primary metabolic disease theory knowledge graph according to the first-level metabolic disease theory knowledge graph and the second-level metabolic disease theory knowledge graph;

[0013] The primary metabolic disease knowledge graph is integrated into the knowledge of a third-party database to obtain a final metabolic disease knowledge graph, wherein the third-party database includes relevant medical knowledge graphs and a knowledge base.

[0014] The present invention also provides a system for acquiring a knowledge graph based on metabolic diseases, comprising:

[0015] A metabolic disease related data acquisition module, used to acquire metabolic disease related data, wherein the metabolic disease related data includes structured data and unstructured data;

[0016] A trigger word and event type extraction module, used to extract the trigger word and event type of each event in the unstructured data using a sequence annotation model;

[0017] An event parameter acquisition module, used to generate questions about the role of event parameters according to each of the extracted trigger words and the corresponding event type, and obtain the event parameters according to the questions using a model based on MRC; the event parameters are event parameters of the event corresponding to the trigger word;

[0018] A first complete event acquisition module, used for obtaining a complete description of an event according to the trigger word, the event type, the event parameter role and the event parameter of each event, recorded as a first complete event;

[0019] A first-level metabolic disease knowledge graph construction module is used to obtain the relationship between the multiple first complete events based on the multiple first complete events using the event relationship classification model, and associate the multiple first complete events according to the relationship between the events to obtain the first-level metabolic disease knowledge graph;

[0020] A second-level metabolic disease knowledge graph construction module is used to extract the structured data using regular expressions to obtain a complete description of each event in the structured data, record it as a second complete event, and associate multiple second complete events to obtain a second-level metabolic disease knowledge graph;

[0021] A primary metabolic disease theory knowledge graph construction module is used to obtain a primary metabolic disease theory knowledge graph according to the first-level metabolic disease theory knowledge graph and the second-level metabolic disease theory knowledge graph;

[0022] The final metabolic disease knowledge graph construction module is used to integrate the primary metabolic disease knowledge graph into the knowledge of a third-party database to obtain the final metabolic disease knowledge graph, and the third-party database includes relevant medical knowledge graphs and knowledge bases.

[0023] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0024] The present invention provides a method and system for acquiring a knowledge graph based on metabolic diseases. First, the data related to metabolic diseases are acquired, trigger words and their corresponding event types in the text of the data related to metabolic diseases are identified, and then corresponding questions are constructed for each event trigger word and event type. The questions are spliced ​​with the context and input into the model to extract other components of the event, and a complete event description is obtained. Then, the relationship between each complete event is determined, and finally, a knowledge graph of metabolic diseases is constructed based on the extracted information. The above method of the present application makes full use of medical text information, medical literature, guidelines, etc. to construct a knowledge graph, and constructs a more complete knowledge graph of metabolic diseases, which can help scientific researchers or medical practitioners to quickly and conveniently query relevant knowledge, assist clinicians in making scientific decisions, and on the other hand, can also provide data support for various intelligent services related to metabolic diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only 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.

[0026] Figure 1 A flow chart of a method for acquiring a knowledge graph based on metabolic diseases provided in Example 1 of the present invention;

[0027] Figure 2 A schematic diagram of a complete event description including event type, trigger and event parameters provided in Example 1 of the present invention;

[0028] Figure 3A schematic diagram of the overall steps of event extraction provided in Example 1 of the present invention;

[0029] Figure 4 The overall schematic diagram of the event detection module provided in Example 1 of the present invention, the input is text, and the output is the tag (BIO) corresponding to each word;

[0030] Figure 5 A schematic diagram of the output of the trigger word extraction module provided in Example 1 of the present invention;

[0031] Figure 6 A schematic diagram of event parameter extraction based on the MRC model provided in Example 1 of the present invention;

[0032] Figure 7 Schematic diagram of two methods for using questions in extracting generation event parameters provided in Example 1 of the present invention;

[0033] Figure 8 A schematic diagram of extracting event relations based on a deep neural network provided in Example 1 of the present invention;

[0034] Fig. 9 A schematic diagram of a preliminary metabolic disease knowledge graph provided in Example 1 of the present invention;

[0035] Fig.10 A schematic diagram of semantic relationship classification when performing knowledge fusion using a Bert deep neural network model provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0037] The purpose of the present invention is to provide a method and system for constructing a knowledge graph based on metabolic diseases, so as to fully explore events in the medical field, thereby helping to analyze the causes of the disease and formulate relevant prevention and treatment measures.

[0038] Metabolic diseases involve a wide range of diseases, and there is currently no knowledge graph in the relevant field. At the same time, with the full development of my country's medical and health informationization process and the rapid development of hospital management system technology, Chinese electronic medical records have rapidly become popular, and major hospitals have accumulated a wealth of electronic medical record data. However, most of the current medical knowledge graphs do not incorporate clinical medical information, which undoubtedly causes a lot of waste of resources. Making full use of clinical information to build knowledge graphs and building a more complete metabolic disease knowledge graph can help scientific researchers or medical practitioners quickly and conveniently query relevant knowledge, assist clinicians in making scientific decisions, and on the other hand, provide data support for various intelligent services related to metabolic diseases.

[0039] At present, most knowledge graphs are entity-centric and focus on describing the static relationship between entities. However, events, another important element in describing the real world, and the relationship between events rarely have a systematic definition in traditional knowledge graphs. The event graph is event-centric and focuses on describing the relationship between events (sequence, co-reference, causality). Unlike traditional knowledge graphs that focus on describing relatively static and slowly changing deterministic knowledge between entities, the knowledge described in the event graph is dynamic and has a certain transfer probability, describing the laws of event development. Constructing a knowledge graph of metabolic diseases can fully explore the laws of patients' illness, provide support for disease prevention, and relieve the pressure on the country, society, and family.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] Example 1

[0042] This embodiment provides a method for constructing a knowledge graph based on metabolic diseases. Figure 1 , the method comprising:

[0043] S1. Acquire metabolic disease related data, wherein the metabolic disease related data includes structured data and unstructured data;

[0044] Specifically, we can crawl metabolic disease-related knowledge from the Internet and obtain clinical metabolic disease-related data. The form of knowledge includes but is not limited to electronic medical records, relevant papers, and patient complaint records. The above knowledge is used to construct a metabolic disease knowledge graph.

[0045] After obtaining the data related to metabolic diseases, event extraction and event relationship extraction are required. Since the above data / knowledge includes structured data and unstructured data, different extraction strategies are adopted for the two different data types. The following first introduces the extraction method for unstructured data.

[0046] S2, extracting the trigger words and event types of each event in the unstructured data using a sequence labeling model;

[0047] Event extraction is a key task in building a knowledge graph of metabolic diseases. A complete event description includes event type, trigger, event parameter role and event parameter, which can be represented as a table-like structure, as shown in the attached figure. Figure 2 It should be noted that different event types have different structures, and an event also has multiple semantic modes. Story extraction is still a challenging task.

[0048] Event extraction is to extract the elements of an event from a text. The structured text information extracted by event extraction is clearer and can be widely used in tasks such as information retrieval, intelligent question answering, and knowledge graph construction. Medical event extraction (MEE) refers to the automatic extraction of structured medical events from unstructured medical text. Among them, medical events refer to events that contain entity and attribute information related to diagnosis and treatment.

[0049] Event extraction mainly consists of two subtasks: (1) event detection, identifying text event trigger words and classifying them into predefined event types; (2) event parameter extraction, identifying the entities that make up each event and classifying them according to the role they play in the event. The overall steps are shown in the attached figure. Figure 3 shown.

[0050] For the event detection task, this embodiment converts it into a sequence labeling problem, extracting event triggers and classifying them at the same time. That is, a sentence is represented as a set of label sequences, and the trigger words contained in the sentence are identified by predicting the label sequence. In a set of label sequences, the label corresponds to the smallest component unit (word) of the sentence. The label symbol contains the position information of the unit in a trigger word and the event category information. Figure 4 shown.

[0051] This embodiment uses a sequence labeling model to extract the trigger words and event types of each event in the unstructured data. Specifically, the following steps are included:

[0052] S21, selecting part of the data from the unstructured data as a training set, and annotating each event text in the training set to obtain an annotated training set; the annotated content includes the trigger word of the event, the event type, the role of the event parameter, the event parameters, and the relationship between the events;

[0053] Before using the sequence labeling model to extract event trigger words and event types, the sequence labeling model must first be trained to enable it to have extraction capabilities.

[0054] S22, using the labeled training set to train the sequence labeling model to obtain a trained sequence labeling model, wherein the event text is used as input during the training process, and the trigger word and event type are used as output; the output format is as follows Figure 4 , through simple processing, we can get the trigger word and event type

[0055] S23, inputting the remaining data in the unstructured data except for some data into the trained sequence labeling model for event detection, and obtaining the trigger word and event type of each event in the remaining data.

[0056] Now let's take an example to illustrate the above-mentioned event detection process: Given a text X = {The patient reported that he had abdominal distension and edema of both lower limbs without obvious cause one month ago, ...}, first input the sequence labeling model to convert the text into a dense word vector (that is, convert the text into a language that the computer can understand, and the vector representations of words with similar semantics are also more similar), and then pass through the bidirectional long short-term memory network (Bi-directional Long Short-Term Memory, Bi-LSTM), the self-attention layer (Self-attention) (which can selectively pay attention to certain important information and ignore other irrelevant information input at the same time) and the conditional random field (CRF) using the BIO labeling scheme (where "B" is the abbreviation of Begin, indicating that the current character is at the beginning of a trigger word; "I" is the abbreviation of Inside, indicating that the current character is located in the other part of the trigger word except the beginning; "O" is the abbreviation of Outside, indicating that the current character is not part of the trigger word) to label each word. The labeling results are as follows: Figure 5 As shown (the upper part is the text input, and the lower part is the corresponding output). Finally, all the trigger words in this text and the corresponding event types T = {(trigger word = occurrence, event type = illness), ...} are obtained.

[0057] The total number of labels is: 2*event type+1, where 2 represents B and I, and 1 represents O. For example, for the illness event, there are two types of labels (B-illness and I-illness). The number of labels is equal to the number of words in the text, and each word is labeled separately.

[0058] S3, generating questions about the role of event parameters according to each of the extracted trigger words and the corresponding event type, and obtaining the event parameters according to the questions using the MRC-based model; the event parameters are the event parameters of the event corresponding to the trigger word;

[0059] The current mainstream methods for event parameter extraction focus on learning the information between parameters and given triggers, but due to the complex interdependencies between triggers and parameters, it is very expensive to obtain fine-grained labeled training data.

[0060] Inspired by the machine reading comprehension task (MRC), this embodiment converts the event parameter extraction task into an MRC task, that is, in addition to the necessary parameters (event type, trigger word), the relevant parameters (other parameter roles in the same event and corresponding event parameters (corresponding words)) can provide information for performing reasoning. In this way, long-distance dependencies can be alleviated and reasoning can be performed globally.

[0061] Step S3 specifically includes:

[0062] S31, training the MRC-based model using the annotated training set to obtain a trained MRC-based model; wherein the MRC-based model training process uses the event text and questions about the role of event parameters as inputs, and the event parameters as outputs, and the questions about the role of event parameters are obtained by the event type corresponding to the event trigger word;

[0063] S32. Generate questions about the role of event parameters according to the event type corresponding to each of the extracted trigger words, input the questions and event context text into the trained MRC-based model, and obtain event parameters.

[0064] In step S2, all trigger words and their event types in the text have been obtained T = {(t1, type1), ..., (t m , type m First, for each trigger word (that is, each event), the component parameters to be extracted generate the corresponding question Q = {{q 1.1 ,q 1.2 , ...}, {q 2.1 ,q 2.2 , ...}, ..., {q m.1 ,q m.2 ,...}}(Note: different events have different roles and quantities of constituent parameters), then pass the question and context text (text to be extracted) into the MRC-based model (i.e. the trained MRC model) using the format [CLS]question[SEP]context, where [CLS] and [SEP] are two special identifiers, indicating the beginning and end of a sentence. Through the interaction of multi-head and multi-layer attention, the question and text are deeply integrated, and finally the answer to the question is predicted to be the starting and ending position in the text, and the answer to the question (the event parameter to be extracted) is output ( Figure 6 For example, for the trigger word t1, a question {q1.1 ,q 1.2 , …}, By iteratively inputting the generated questions into the model, we can get the various parameters of the event, and after integration, we can get a complete description of the event.

[0065] Among them, regarding the generation of the question of event component parameter roles, this embodiment adopts a template-based generation strategy to obtain. The template-based method can directly use the event parameter role + event type as the question, or add other parameters as additional information. Take the medical consultation event as an example, as shown in the attached figure. Figure 7 As shown:

[0066] Although other relevant parameters in the same event can be used as auxiliary information for reasoning during the training phase, this information is noisy during the testing phase, and may even be completely irrelevant parameters. This difference in training and testing may lead to a decline in model performance. In order to bridge this gap, the present application adopts the method of knowledge distillation to train a model that can be reasoned without additional information, namely the student model. The teacher model (the model originally trained using auxiliary information) and the student model have the same architecture but different parameters. During the knowledge distillation process, the parameters of the teacher model are fixed, and only the parameters of the student model are optimized. At the same time, the model can also be expanded to the case of multiple teachers (i.e., using different degrees (proportions) of assistance).

[0067] S4, obtaining a complete description of the event according to the trigger word, the event type, the event parameter role and the event parameter of each event, and recording it as the first complete event;

[0068] S5. According to the plurality of the first complete events, using an event relationship classification model to obtain the relationship between the plurality of the first complete events, and associating the plurality of the first complete events according to the relationship between the events to obtain a first-level metabolic disease knowledge graph;

[0069] After obtaining the complete description of each complete event, it is necessary to extract the event relationship to obtain the relationship between multiple complete events. In order to obtain the relationship between events (such as time sequence, cause and effect, etc.), this application uses a method based on a deep neural network model to extract the relationship between events, as shown in the attached figure. Figure 8 As shown. The model uses a pre-trained language model as the encoding layer and takes the two event mentions and the sentences between the events as input. For event 1 and event 2 (the event on the left and the event on the right in a text), average pooling is used on the corresponding context vectors to obtain vectors V1 and V2 as vector representations of the two events. V1, V2, and |V1-V2| are then connected to generate a pair representation V = (V1, V2, |V1-V2|). Finally, V is passed to the linear layer and the SoftMax layer for relationship prediction.

[0070] After obtaining the relationship between multiple complete events, the multiple complete events are associated according to the relationship between the above-mentioned predicted events to obtain a preliminary unstructured metabolic disease knowledge graph, that is, the above-mentioned first-level metabolic disease knowledge graph.

[0071] So far, the information extraction process of unstructured data has been completed. The following is an introduction to the extraction method of structured data.

[0072] S6. Extracting the structured data using regular expressions to obtain a complete description of each event in the structured data, recording the complete description as a second complete event, and associating multiple of the second complete events to obtain a second-level metabolic disease knowledge graph;

[0073] S7, obtaining a primary metabolic disease theory knowledge graph according to the first-level metabolic disease theory knowledge graph and the second-level metabolic disease theory knowledge graph;

[0074] After information extraction, the text is converted into a structured event representation (including event-related entities and attributes, as well as the relationships between events, etc.).

[0075] Each medical event contains several related medical entities and other event elements. These entities are connected to each other through the event relationship formed by the medical event, and at this time, a preliminary metabolic disease knowledge graph can be formed, as shown in the attached figure. Fig. 9 shown.

[0076] Fig. 9 In the figure, the two nodes in the middle (rounded rectangles, with the node label indicating the event type) represent a medical event, and the rectangular nodes connected to them are entities or attributes related to the event (i.e., the constituent parameters of each event), such as patients, events, symptoms, etc. in a medical event. From the above figure, it can be seen that an event subject is associated with multiple event subjects through event nodes, that is, it can be determined that there is a relationship between the two event subjects. Some nodes may have null values ​​of NA, which are generally non-critical event elements and generally do not affect the construction of the overall graph.

[0077] S8. Integrate the primary metabolic disease knowledge graph into the third-party database knowledge to obtain the final metabolic disease knowledge graph.

[0078] As of step S7, the construction of the preliminary strength knowledge map has been completed, but the construction process of the knowledge map is a process of continuous iteration and correction, and the knowledge is time-effective. As the knowledge related to metabolic diseases is constantly updated, the previously constructed knowledge map may have a lot of ambiguity, redundancy, omissions, obsolescence and erroneous information, and third-party databases (such as relevant medical knowledge maps, knowledge bases, etc.) are also constantly updated. Therefore, by integrating the knowledge of third-party databases (i.e., knowledge fusion), it is particularly important to repeatedly iterate and update knowledge. Through continuous iteration, correction and expansion, the accuracy of the knowledge map data can be ensured, providing guarantee for the effective implementation of the application layer.

[0079] The knowledge fusion task includes synonym relationship matching and hyponymy relationship matching, which consists of three stages: candidate entity screening, data sampling, and semantic relationship classification.

[0080] 1. Candidate entity screening

[0081] Obtaining a set of entities S of the same type as the entity m in the metabolic disease-related data from the third-party database, and calculating the similarity of character features between m and all entities in S, and selecting the top k entities with the highest similarity as candidate entities of m;

[0082] 2. Data Sampling

[0083] Selecting part of entity data from the candidate entities by adopting a stratified proportional sampling method; annotating the relationship between the part of entity data to obtain annotated data; the relationship between the part of entity data includes a synonym relationship and a hyponym relationship;

[0084] After obtaining the candidate entity pairs, in order to train the semantic relationship classification model, the entity pairs need to be labeled. Considering the large number of entity pairs, the cost of labeling all of them is high, so a stratified proportional sampling method is used to label part of the data. Specifically, the average similarity score of the candidate entity set of each entity to be matched is calculated, and the candidate entity set is divided into several layers at equal intervals according to the average score. In each layer, samples are randomly selected according to the proportion of the number of entities to be matched to the whole.

[0085] 3. Semantic relationship classification

[0086] This stage is divided into two steps: model training and model application. First, the BERT-based semantic relationship classification model is trained using the labeled data to obtain the optimal model (the labeled data is used as input and the category is used as output during the model training process). Then the optimal model is used to predict the semantic relationship of the remaining unlabeled entity pairs.

[0087] This embodiment uses a pre-trained language model to encode entity pairs, obtains the semantic relationship representation of entity pairs, and then uses two fully connected layers to classify the semantic relationship and calculate the category of the semantic relationship, where [CLS] and [SEP] are two special identifiers that indicate the beginning and end of a sentence. Fig.10 shown.

[0088] Matching the remaining entity data according to the relationships between the remaining entity data to obtain sorted data;

[0089] The sorted data is used to expand the primary metabolic disease knowledge graph to obtain the final metabolic disease knowledge graph.

[0090] It should be noted that, usually, the medical knowledge graph consists of two parts: the ontology layer and the instance layer. The ontology layer includes entity types such as patients, diseases, symptoms, examinations, and treatments; each entity in the instance layer contains its own attributes and attribute values, and there are relationships between entities. First, the initial model of the metabolic disease ontology is defined, and then the ontology is defined as the conceptual framework to perform information extraction (structuring unstructured data and then combining it with the acquired structured data to form a preliminary knowledge representation) and knowledge fusion to construct a metabolic disease knowledge graph.

[0091] Therefore, as an optional implementation, the method further includes, before step S1, constructing an ontology.

[0092] The design methods of ontology can be divided into two types: bottom-up and top-down. Most knowledge graphs such as Dbpedia and YAGO adopt a bottom-up construction method to ensure the coverage of the ontology. However, in the medical field, the accuracy of the ontology is strict and the data model is relatively fixed, so it is more effective to adopt a top-down design method. The ontology of the metabolic disease knowledge graph includes: medical entities, medical events and the relationship between them, especially the temporal relationship between medical events. The ontology representation of the knowledge graph mainly includes:

[0093] 1. Entity, such as "XX Hospital", "XX Department", etc.;

[0094] 2. Events, such as "consultation", "treatment", "discharge", etc.;

[0095] 3. Relevant attributes of entities and events, such as “time of medical consultation event”, “age of patient”, etc.;

[0096] 4. The logical evolutionary relationship between events mainly includes temporal relationship, cause-and-effect relationship, etc.

[0097] The specific construction steps are as follows:

[0098] 1. Summarize the important concepts in the field of metabolic diseases, define the class names, and define the relationships between classes. The most important concepts in the field of metabolic diseases are initially defined as causes, symptoms, diagnosis, and treatment.

[0099] 2. Define the object attributes of the class, that is, determine the relationship between class concepts and sort out the hierarchical relationship of the class. From a semantic point of view, the domain ontology basically contains four relationships: part of, kind of, instance of, and attribute of. Other relationships between classes can also be created according to actual needs, such as Cause of, Result in, etc.

[0100] 3. Define the data attributes of the class. For example, the name of a disease, its attribute value is string type, and it has an inverse functional property constraint.

[0101] 4. Define attribute facets. Each attribute consists of multiple facets, including attribute value type, domain, value range, etc. By defining attribute facets, the application scope of the attribute is limited, which plays an important role in subsequent semantic reasoning.

[0102] Knowledge graphs are gradually being applied to all walks of life and have become the cornerstone of artificial intelligence. A simple static entity knowledge graph obviously can no longer meet the needs of practical applications, while the event knowledge graph depicts dynamic events and has a stronger application value. The knowledge described by the event graph is dynamic, describing the laws of event development and facilitating reasoning about knowledge. Fully exploring events in the medical field can help analyze the causes of diseases and formulate relevant prevention and control measures. In order to demonstrate the practical application value of the event graph, the following briefly gives examples from five research aspects: patient statistics, etiology, diagnosis, treatment, and prognosis, as shown in Table 1:

[0103] Table 1 Examples of practical application value of the event map

[0104] Example of a problem Patient statistics Number of people aged 30 to 50 with diabetes Causes What complications are likely to occur with diabetes? diagnosis The patient has edema of both lower limbs, vomiting, and dizziness. What diseases may he have? treat Symptoms and proportion of patients experiencing discomfort after taking specific drugs Prognosis The proportion of gout patients who relapse after treatment

[0105] Example 2

[0106] This embodiment provides a system for acquiring a knowledge graph based on metabolic diseases, including:

[0107] A metabolic disease related data acquisition module M1 is used to acquire metabolic disease related data, wherein the metabolic disease related data includes structured data and unstructured data;

[0108] A trigger word and event type extraction module M2, used to extract the trigger word and event type of each event in the unstructured data using a sequence annotation model;

[0109] The event parameter acquisition module M3 is used to generate questions about the role of event parameters according to each of the extracted trigger words and the corresponding event type, and obtain the event parameters according to the questions using the MRC-based model; the event parameters are the event parameters of the event corresponding to the trigger word;

[0110] A first complete event acquisition module M4, configured to obtain a complete description of an event according to the trigger word, the event type, the event parameter role and the event parameter of each event, recorded as a first complete event;

[0111] A first-level metabolic disease knowledge graph construction module M5 is used to obtain the relationship between the multiple first complete events based on the multiple first complete events using the event relationship classification model, and associate the multiple first complete events based on the relationship between the events to obtain the first-level metabolic disease knowledge graph;

[0112] The second-level metabolic disease knowledge graph construction module M6 is used to extract the structured data using regular expressions to obtain a complete description of each event in the structured data, record it as a second complete event, and associate multiple second complete events to obtain a second-level metabolic disease knowledge graph;

[0113] A primary metabolic disease theory knowledge graph construction module M7 is used to obtain a primary metabolic disease theory knowledge graph according to the first-level metabolic disease theory knowledge graph and the second-level metabolic disease theory knowledge graph;

[0114] The final metabolic disease knowledge graph construction module M8 is used to integrate the primary metabolic disease knowledge graph into the third-party database knowledge to obtain the final metabolic disease knowledge graph.

[0115] Optionally, the trigger word and event type extraction module specifically includes:

[0116] The labeled training set acquisition submodule is used to select part of the data from the unstructured data as the training set, and to label each event text in the training set to obtain the labeled training set; the labeled content includes the trigger word of the event, the event type, the role of the event parameter, the event parameters, and the relationship between the events;

[0117] A sequence annotation model training submodule is used to train the sequence standard model using the annotated training set to obtain a trained sequence annotation model, wherein the event text is used as input and the trigger word and event type are used as output during the training process;

[0118] The event detection submodule is used to input the remaining data in the unstructured data except for some data into the trained sequence annotation model for event detection, and obtain the trigger word and event type of each event in the remaining data.

[0119] Optionally, the event parameter acquisition module specifically includes:

[0120] The MRC-based model training submodule is used to train the MRC-based model using the labeled training set to obtain a trained MRC-based model; wherein the MRC-based model training process uses event text and questions about the role of event parameters as input, and event parameters as output; the questions about the role of event parameters are obtained through the event type corresponding to the event trigger word.

[0121] The event parameter acquisition submodule is used to generate questions about the role of event parameters according to the event type corresponding to each of the extracted trigger words, and input the questions into the trained MRC-based model to obtain event parameters.

[0122] Optionally, the final metabolic disease knowledge graph construction module specifically includes:

[0123] A candidate entity acquisition submodule is used to obtain a set S of entities similar to the entity m in the metabolic disease-related data from the third-party database, calculate the similarity of character features between m and all entities in S, and select the top k entities with the highest similarity as candidate entities of m;

[0124] The candidate entity annotation submodule is used to select part of the entity data from the candidate entity by adopting a stratified proportional sampling method; annotate the relationship between the part of the entity data to obtain annotated data; the relationship between the part of the entity data includes a synonym relationship and a hyponym relationship;

[0125] A semantic relationship classification model training submodule is used to train a semantic relationship classification model based on a deep neural network using the annotated data to obtain an optimal semantic relationship classification model. During the training process of the semantic relationship classification model, the annotated data is used as input and the relationship between entities is used as output;

[0126] The remaining entity data relationship acquisition submodule is used to input the remaining entity data into the optimal semantic relationship classification model to obtain the relationship between the remaining entity data; the remaining entity data is the data in the candidate entity except the partial entity data;

[0127] A remaining entity data matching submodule, used for matching the remaining entity data according to the relationships between the remaining entity data to obtain sorted data;

[0128] The data expansion submodule is used to expand the primary metabolic disease knowledge graph using the sorted data to obtain the final metabolic disease knowledge graph.

[0129] As for the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0130] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for acquiring a knowledge graph based on metabolic diseases, characterized in that: include: Acquiring metabolic disease related data, wherein the metabolic disease related data includes structured data and unstructured data; Extracting trigger words and event types for each event in the unstructured data using a sequence labeling model; Generate questions about the role of event parameters according to each of the extracted trigger words and the corresponding event type, and obtain the event parameters using the MRC-based model according to the questions; The event parameter is the event parameter of the event corresponding to the trigger word; Obtaining a complete description of the event according to the trigger word, the event type, the event parameter role, and the event parameter of each event, which is recorded as the first complete event; According to the plurality of the first complete events, the relationship between the plurality of the first complete events is obtained by using an event relationship classification model, and the plurality of the first complete events are associated according to the relationship between the events to obtain a first-level metabolic disease knowledge graph; Extracting the structured data using regular expressions to obtain a complete description of each event in the structured data, recording it as a second complete event, and associating multiple second complete events to obtain a second-level metabolic disease knowledge graph; Obtaining a primary metabolic disease theory knowledge graph according to the first-level metabolic disease theory knowledge graph and the second-level metabolic disease theory knowledge graph; Integrate the primary metabolic disease knowledge graph into the knowledge of a third-party database to obtain a final metabolic disease knowledge graph, wherein the third-party database includes a relevant medical knowledge graph and a knowledge base; The step of integrating the primary metabolic disease knowledge graph into the third-party database knowledge to obtain the final metabolic disease knowledge graph specifically includes: Obtaining a set of entities S of the same type as the entity m in the metabolic disease-related data from the third-party database, and calculating the similarity of character features between m and all entities in S, and selecting the top k entities with the highest similarity as candidate entities of m; Selecting part of entity data from the candidate entities by adopting a stratified proportional sampling method; annotating the relationship between the part of entity data to obtain annotated data; the relationship includes a synonym relationship and a hyponym relationship; Using the annotated data to train a semantic relationship classification model based on a deep neural network to obtain an optimal semantic relationship classification model, wherein the annotated data is used as input and the relationship between entities is used as output during the training of the semantic relationship classification model; Inputting the remaining entity data into the optimal semantic relationship classification model to obtain the relationship between the remaining entity data; the remaining entity data is the data in the candidate entity except the partial entity data; Matching the remaining entity data according to the relationships between the remaining entity data to obtain sorted data; The sorted data is used to expand the primary metabolic disease knowledge graph to obtain the final metabolic disease knowledge graph.

2. The method according to claim 1, characterized in that The extracting of the trigger words and event types of each event in the unstructured data by using the sequence labeling model specifically includes: Selecting part of the unstructured data as a training set, annotating each event text in the training set to obtain an annotated training set; the annotated content includes the trigger word of the event, the event type, the role of the event parameter, the event parameters, and the relationship between the events; Using the labeled training set to train the sequence labeling model to obtain a trained sequence labeling model, wherein the event text is used as input and the trigger word and event type are used as output during the training process; The remaining data except for some data in the unstructured data is input into the trained sequence labeling model for event detection to obtain the trigger word and event type of each event in the remaining data.

3. The method according to claim 2, characterized in that The step of generating questions about the role of event parameters according to the event type corresponding to each of the extracted trigger words, and obtaining the event parameters using the MRC-based model according to the questions, specifically includes: The MRC-based model is trained using the labeled training set to obtain a trained MRC-based model; wherein the MRC-based model training process uses event text and questions about event parameter roles as inputs, and event parameters as outputs, and the questions about event parameter roles are obtained through event types corresponding to event trigger words; According to the event type corresponding to each of the extracted trigger words, a question about the role of event parameters is generated, and the question and event context text are input into the trained MRC-based model to obtain event parameters.

4. The method according to claim 1, characterized in that Before obtaining the metabolic disease related data, the method further includes constructing an initial model of the metabolic disease ontology: Summarize the concepts in the field of metabolic diseases, classify the concepts, and obtain concept types; the concept types include causes, symptoms, diagnosis, and treatment; determining the relationships between the various types of concepts; Defining the attributes of the data contained in each of the concept types to obtain an initial model of the metabolic disease ontology; The final metabolic disease knowledge graph is constructed based on the initial model of the metabolic disease ontology.

5. A knowledge graph acquisition system based on metabolic diseases, characterized in that: include: A metabolic disease related data acquisition module, used to acquire metabolic disease related data, wherein the metabolic disease related data includes structured data and unstructured data; A trigger word and event type extraction module, used to extract the trigger word and event type of each event in the unstructured data using a sequence annotation model; An event parameter acquisition module, used to generate questions about the role of event parameters according to the event type corresponding to each of the extracted trigger words, and obtain the event parameters using an MRC-based model according to the questions; The event parameter is the event parameter of the event corresponding to the trigger word; A first complete event acquisition module, used for obtaining a complete description of an event according to the trigger word, the event type, the event parameter role and the event parameter of each event, recorded as a first complete event; A first-level metabolic disease knowledge graph construction module is used to obtain the relationship between the multiple first complete events based on the multiple first complete events using the event relationship classification model, and associate the multiple first complete events based on the relationship between the events to obtain the first-level metabolic disease knowledge graph; A second-level metabolic disease knowledge graph construction module is used to extract the structured data using regular expressions to obtain a complete description of each event in the structured data, record it as a second complete event, and associate multiple second complete events to obtain a second-level metabolic disease knowledge graph; A primary metabolic disease theory knowledge graph construction module, used to obtain a primary metabolic disease theory knowledge graph according to the first-level metabolic disease theory knowledge graph and the second-level metabolic disease theory knowledge graph; A final metabolic disease knowledge graph construction module is used to integrate the primary metabolic disease knowledge graph into the knowledge of a third-party database to obtain a final metabolic disease knowledge graph, wherein the third-party database includes a relevant medical knowledge graph and a knowledge base; The final metabolic disease knowledge graph construction module specifically includes: A selected entity acquisition submodule is used to obtain a set of entities S of the same type as the entity m in the metabolic disease-related data from the third-party database, calculate the similarity of character features between m and all entities in S, and select the top k entities with the highest similarity as candidate entities of m; The candidate entity annotation submodule is used to select part of the entity data from the candidate entity by adopting a stratified proportional sampling method; annotate the relationship between the part of the entity data to obtain annotated data; the relationship between the part of the entity data includes a synonym relationship and a hyponym relationship; A semantic relationship classification model training submodule is used to train a semantic relationship classification model based on a deep neural network using the annotated data to obtain an optimal semantic relationship classification model. During the training process of the semantic relationship classification model, the annotated data is used as input and the relationship between entities is used as output; The remaining entity data relationship acquisition submodule is used to input the remaining entity data into the optimal semantic relationship classification model to obtain the relationship between the remaining entity data; the remaining entity data is the data in the candidate entity except the partial entity data; A remaining entity data matching submodule, used for matching the remaining entity data according to the relationships between the remaining entity data to obtain sorted data; The data expansion submodule is used to expand the primary metabolic disease knowledge graph using the sorted data to obtain the final metabolic disease knowledge graph.

6. The system according to claim 5, characterized in that The trigger word and event type extraction module specifically includes: The labeled training set acquisition submodule is used to select part of the data from the unstructured data as the training set, and to label each event text in the training set to obtain the labeled training set; the labeled content includes the trigger word of the event, the event type, the role of the event parameter, the event parameters, and the relationship between the events; A sequence labeling model training submodule is used to train the sequence labeling model using the labeled training set to obtain a trained sequence labeling model, wherein the event text is used as input and the trigger word and event type are used as output during the training process; The event detection submodule is used to input the remaining data in the unstructured data except for some data into the trained sequence annotation model for event detection, and obtain the trigger word and event type of each event in the remaining data.

7. The system according to claim 6, characterized in that The event parameter acquisition module specifically includes: The MRC-based model training submodule is used to train the MRC-based model using the labeled training set to obtain a trained MRC-based model; wherein the MRC-based model training process uses the event text and questions about the role of event parameters as inputs, and the event parameters as outputs, and the questions about the role of event parameters are obtained by the event type corresponding to the event trigger word; The event parameter acquisition submodule is used to generate questions about the role of event parameters according to the event type corresponding to each of the extracted trigger words, and input the questions and event context text into the trained MRC-based model to obtain event parameters.

8. The system according to claim 5, characterized in that It also includes an initial model building module of the metabolic disease ontology, which is used to build an initial model of the metabolic disease ontology before obtaining the metabolic disease related data; the initial model building module of the metabolic disease ontology specifically includes: The concept construction and classification submodule is used to summarize the concepts in the field of metabolic diseases, classify the concepts, and obtain concept types; the concept types include causes, symptoms, diagnosis, and treatment; A relationship determination submodule, used to determine the relationship between the concept types; The attribute definition submodule is used to define the attributes containing data in each of the concept types to obtain the initial model of the metabolic disease ontology; the final metabolic disease knowledge graph is constructed on the basis of the initial model of the metabolic disease ontology.

Images