Pipeline type high robustness medical knowledge graph standardization construction method

Abstract

The application provides a pipeline type high-robustness standardized medical knowledge graph construction method and device, the method comprises the following steps: enumerating all candidate entity word spans in a medical text to perform floating marking, attaching the floating marking of all candidate entity word spans to the medical text as a training example, grouping the floating marking of the entity word spans, using a pre-trained language model to obtain contextualized representation of the entity word spans in each group, and using a neural network structure to output entity type information according to the contextualized representation; obtaining the relationship category between the entity category information according to the contextualized representation and the entity type information coding, and outputting the relationship information; and aligning the entity type information and the relationship information with a medical terminology standard to construct a medical knowledge graph.

Description

Technical Field

[0001] This application relates to the field of medical information knowledge graph technology, and in particular to a pipeline-based, highly robust, standardized construction method for medical knowledge graphs. Background Technology

[0002] Clinical text entity extraction is an important natural language processing technique for medical texts, used to extract a wealth of professional knowledge and useful information. Related clinical text entity extraction methods mainly include rule-based methods, traditional machine learning-based methods, and deep learning-based methods.

[0003] In related technologies, rule-based medical entity extraction methods, when the domain dictionary is sufficiently comprehensive and the established rules accurately describe the features of the target domain text, can outperform other methods. However, developing domain rule templates and maintaining the domain dictionary are time-consuming, labor-intensive, and require a high level of expertise. Different domains have different rules and dictionaries, resulting in poor transferability of this method.

[0004] Traditional machine learning methods require large amounts of manually labeled datasets to train the model parameters, but large-scale medical datasets are relatively scarce. Furthermore, machine learning methods necessitate experts manually selecting various features that influence named entity recognition tasks, which translates to expensive annotation costs.

[0005] Deep learning-based medical named entity recognition models generally include an input layer, an embedding layer, an encoding layer, a decoding layer, and an output layer. The language model in the embedding layer learns semantic and grammatical knowledge from the input medical text and vectorizes it, primarily through word-based representations, character-based representations, and hybrid representations based on character and word information. The encoding layer uses network structures such as CNN, RNN, and Transformer to extract and encode features from the information input from the embedding layer. The decoding layer uses a decoder to predict labels based on the output of the encoding layer, ultimately outputting the optimal label sequence.

[0006] In related technologies, even if floating markers are used to replace fixed markers, it is simply a replacement to achieve efficient batch processing calculations, but at the expense of the performance of the entity extraction model. Summary of the Invention

[0007] In view of this, the purpose of this application is to propose a pipeline-based, highly robust medical knowledge graph standardization construction method and device.

[0008] Based on the above objectives, in a first aspect, this application provides a pipeline-based, highly robust medical knowledge graph standardization construction method, which includes:

[0009] All candidate entity word spans in the medical text are enumerated and tagged with floating symbols. The floating symbols of all candidate entity word spans are appended to the medical text as a training instance. The floating symbols of the entity word spans are grouped. The contextualized representation of the entity word spans in each group is obtained using a pre-trained language model. Based on the contextualized representation, entity type information is output using a neural network structure.

[0010] Based on the contextualized representation and the entity type information encoding, the relationship categories between the entity category information are obtained, and the relationship information is output.

[0011] By aligning the entity type information and relationship information with medical terminology standards, a medical knowledge graph is constructed.

[0012] Optionally, the candidate entity words include: a start word and an end word;

[0013] The span of all candidate entity words in the enumerated medical text is marked with a floating tag, including:

[0014] Each character in the medical text is taken as the beginning word, and the beginning word and the characters following the beginning word are taken as the ending words.

[0015] An attention mask matrix is ​​applied to suspend candidate entity words with a span from the start word to the end word in sequence;

[0016] In this context, each of the floating markers is visible to another floating marker within the same span in the attention mask matrix, but not visible to floating markers in other spans.

[0017] Optionally, the floating markers for the entity word span are grouped, including:

[0018] The suspended markers are sorted sequentially;

[0019] The floating markers are divided into groups of size K, and adjacent spans are preferentially grouped into the same group, where K is the maximum number of spans in a group;

[0020] The floating markers for each group are processed in parallel.

[0021] Optionally, based on the contextualized representation and the entity type information encoding, the relationship categories between the entity category information are obtained, and the relationship information is output, including:

[0022] Obtain the contextualized representation vector of the entity word span in each group obtained using a pre-trained language model;

[0023] The entity category information is obtained and encoded into a vector of fixed dimensions;

[0024] By concatenating the contextualized representation vector and the entity category information into a fixed-dimensional vector, the category of each relationship between the main entity and the object entity in the entity category information is predicted.

[0025] Optionally, the step of concatenating the contextualized representation encoding vector and the entity category information into a fixed-dimensional vector to predict the category of each relationship between the subject entity and the object entity in the entity category information includes:

[0026] The concatenated contextualized representation encoding vector and the entity category information are encoded into fixed-dimensional vectors and passed in parallel through FFN-head and FFN-tail based on feedforward neural networks to obtain FFN-head output vector and FFN-tail output vector respectively.

[0027] The FFN-head output vector and the FFN-tail output vector are passed through a classification network to predict the category of each relationship between the main entity and the guest entity.

[0028] Optionally, aligning the entity type information and the relationship information with medical terminology standards includes:

[0029] By using frame matching, combined with element-level and structure-level matching, the string similarity s is calculated. str Similarity s of attribute sets attr Similarity of relational structures rel The similarity between the entity type information and the medical terminology standard is calculated comprehensively;

[0030] If the similarity between the entity type information and the medical terminology standard is greater than a threshold, then the medical entity in the entity type information and the medical terminology standard are considered to be the same medical entity.

[0031] If the similarity of the entity type information to the medical terminology standard is less than or equal to a threshold, then the medical entity in the entity type information is considered to be different from the medical terminology standard, and the original medical text is retained.

[0032] Optionally, before aligning the entity information and the relationship information, the method further includes:

[0033] The entity information and the relationship information are stored using Neo4j.

[0034] Secondly, this application provides a pipeline-type highly robust medical knowledge graph standardization construction device, comprising:

[0035] The entity extraction module is used to suspend and group all candidate entity word spans in the medical text, and extract entity type information from the candidate entity word spans;

[0036] The relationship extraction module extracts relationships from the entity type information, predicts the category of each relationship between the entity type information, and outputs the relationship information.

[0037] The medical knowledge graph construction module is used to align the entity type information, the relationship information, and the medical terminology standards to construct a medical knowledge graph.

[0038] The storage module is used to store the entity type information and the relationship information.

[0039] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0040] Fourthly, this application provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the methods described above.

[0041] As can be seen from the above, the pipeline-based, highly robust medical knowledge graph standardization construction method and apparatus provided in this application overcomes the limitations of related technologies in fully learning contextual information during the entity information extraction stage when acquiring medical text. Enumerating all candidate entity words across spans for dangling tags better identifies overlapping and nested entity information. Grouping and parallel processing of dangling tags overcomes the drawback of processing all candidate entity words one by one in related technologies, thus accelerating processing speed. Relationship extraction is designed to predict the relationships between entity type information. By aligning entity type information and relationship information, potential semantic redundancy and inconsistencies in expression are eliminated, thus resolving ambiguity in entity type information and relationship information. Finally, a medical information knowledge graph is constructed and visualized. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating a pipeline-based, highly robust medical knowledge graph standardization construction method according to an embodiment of this application.

[0044] Figure 2 This is a schematic diagram illustrating the hovering marking of medical text according to an embodiment of this application;

[0045] Figure 3 This is a schematic diagram illustrating two entity type information in an embodiment of this application;

[0046] Figure 4 This is a schematic diagram illustrating the calculation of the similarity of two entity type information using a GCN graph convolutional network, as an embodiment of this application.

[0047] Figure 5 This is a schematic diagram of a pipeline-type highly robust medical knowledge graph standardization construction device according to an embodiment of this application;

[0048] Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0050] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar words used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "throat" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0051] In the related art, when performing entity extraction, a fixed labeling method and a hanging labeling method are adopted. In the fixed labeling method, two fixed labels are inserted before and after the text to highlight the position of the span in the input text, and two pairs of labels are inserted to locate whether the span label pair is a main entity or an object entity. However, this method cannot process multiple span span label pairs at the same time. In the hanging labeling method, when performing hanging labeling on spans, the fixed labels are simply replaced with hanging labels to achieve efficient batch processing, without considering the mutual relationship between spans, which may not enable accurate identification during relation extraction, thus sacrificing the performance of the entity extraction model.

[0052] in conjunction with Figure 1 which shows a flowchart of a pipelined highly robust standardized construction method for a medical knowledge graph, the present application provides a pipelined highly robust standardized construction method for a medical knowledge graph, comprising:

[0053] S101: Enumerate all candidate entity word spans in the medical text for hanging labeling, append the hanging labels of all candidate entity word spans to the medical text to form one training instance, group the hanging labels of the entity word spans, obtain the contextual representation of the entity word spans in each group using a pre-trained language model, and output entity type information through a neural network structure according to the contextual representation.

[0054] Specifically, one pair of hanging labels is associated with one span. First, all possible candidate entity words in each sentence of the medical text are enumerated, the first character of the medical text is taken as the start word of a span, and the first character is also taken as the end word of the span, with subsequent characters sequentially taken as the end word of the span, so that all spans related to the first character are selected and classified into a first group. Then, the second character is taken as the start word of a span, the second character is taken as the end word of the span, subsequent characters are sequentially taken as the end word of the span, with reference to Figure 2 , for the medical text "the throat tends to be dry and sore", first take the character "throat" as the start word of the span, and sequentially take "throat", "mouth", "tend", "to", "dry", "sore" as the end word of the span. Then take "mouth" as the start word of the span, and sequentially take "mouth", "tend", "to", "dry", "sore" as the end word of the span, and so on, until the hanging labeling ends when "sore" is taken as both the start word and the end word of the span.

[0055] It can be understood that a candidate entity word at least includes a start word and an end word, and may also include intermediate words. For example, for the candidate entity word "throat", both the start word and the end word can be regarded as "throat" with no intermediate word, while for the candidate entity word "throat mouth tend", the start word is "throat", the end word is "tend", and the intermediate word is "mouth".

[0056] However, to enable the specification of multiple pairs of dangling tokens in parallel, a directional attention mask matrix is ​​applied, sequentially dangling all candidate entity words spanning from the start word to the end word. Specifically, each dangling token is visible to another dangling token of the same pair within the attention mask matrix, but invisible to other dangling tokens.

[0057] Appending all candidate entity word span dangling tags to the medical text creates a training instance. This training instance, containing all candidate entity word span dangling tags, becomes very long, making data processing slow. To speed up the processing of all candidate entity word spans, the spans are grouped for parallel processing. Specifically, first, all the aforementioned candidate entity word span dangling tags are sorted, and then the tags are divided into groups of size K, where K is the maximum span number within each group. For example, refer to... Figure 2 For example, the phrase "throat is easily dry and sore" has a K value of 6. Then, "throat", "throat", "throat capacity", "throat is easily dry", "throat is easily dry and sore" are grouped into the first group. Next, "child", "child capacity", "child is easily dry", "child is easily dry and sore", and "capacity" are grouped into the second group, and so on. Adjacent spans are prioritized for grouping into the same group; for instance, "capacity" is also grouped into the second group. Then, the floating markers for the spans of each entity word group are processed in parallel to improve the processing speed of the floating markers for each entity word span.

[0058] Then, a pre-trained language model is used to perform a contextualized representation of each training instance. That is, the contextualized representation of the span of candidate entity words in each group is performed by the pre-trained language model. Based on the contextualized representation of the span of candidate entity words in each group, the contextualized representation is input into the neural network structure to output the entity category information of the entity words.

[0059] It should be noted that the pre-trained language model here can be BERT, and the neural network structure can be a bidirectional long short-term memory network or a multilayer perceptron.

[0060] Enumerating all candidate entity word spans for dangling tagging can avoid ambiguity caused by overlap and nesting during entity word extraction.

[0061] S102. Based on the contextualized representation and entity type information encoding, obtain the relationship categories between entity category information and output the relationship information.

[0062] Specifically, relation extraction is divided into two parts: extracting entity type information and extracting the contextualized representation encoding vector of entity word span obtained using a pre-trained language model. The entity type information is encoded into a fixed-dimensional vector, and then the contextualized representation encoding vector of entity word span is concatenated with the fixed-dimensional vector as input.

[0063]

[0064] Where x represents the input, The encoding vector for the contextualized representation of entity word span. For the input medical text, Entity type information is encoded into a fixed-dimensional vector.

[0065] Furthermore, to ensure the directionality of the encoding, the concatenated contextualized representation encoding vector and entity category information are encoded into fixed-dimensional vectors and input in parallel through a feedforward neural network-based FFN-head and FFN-tail, resulting in two vectors: the FFN-head output vector and the FFN-tail output vector.

[0066]

[0067]

[0068] We can approximate the vector representations of the principal entity and the object entity.

[0069] Subsequently, these two vectors are passed through the Biaffine layer of the classification network to predict the category to which the entity type information belongs for each relation.

[0070]

[0071]

[0072] Among them, s j,k (RE) The category representing the predicted relationship between j and k.

[0073] This is the general expression for Biaffine.

[0074] This yields the relationship category between the primary entity and the secondary entity. It's understandable that entity type information includes both the primary and secondary entities.

[0075] S103. Construct a medical knowledge graph by aligning entity type information and relational information with medical terminology standards.

[0076] Entity type information from entity extraction and relation information from relation extraction are extracted, but the extracted results may have problems such as semantic duplication and inconsistent expression, so alignment is required.

[0077] Specifically, frame matching is used, combining element-level matching and structure-level matching. In this application, element-level matching is based on string matching techniques to match entities. The more similar the strings are, the more likely the two entity types represent the same concept. Simultaneously, matching is performed based on constraints in element-level matching, considering the similarity of the attribute sets of two entity types. The higher the attribute similarity, the more likely the two entity types represent the same concept. Structure-level matching utilizes the relationships between conceptual entities for matching. The higher the similarity of the relationship between two entity types, the higher the similarity between the two entity types. The EM algorithm and the Label Propagation iterative algorithm are used for approximate solution.

[0078] like Figure 3 , Figure 4 As shown, there are two entity type information G1 and G2. The similarity between two medical entities within each entity is compared to determine if they are the same thing. Each medical entity is itself a string with several attributes as constraints, and its structure consists of its adjacent nodes and the set of n-hop adjacent nodes. The string similarity s... str Similarity s of attribute sets attr Similarity of relational structures rel The similarity is summed up as the similarity between two entity types. String similarity can be calculated through string matching. For attribute set similarity, let A1 be the set of attribute names for entity E1 and A2 be the set of attribute names for entity E2, then...

[0079]

[0080] Where a1 is any attribute name in E1, a2 is any attribute name in E2, and s im is the text similarity between the two attribute names. For each a1, the above formula finds the most similar attribute name a2 in the attribute name set A2 of E2, calculates the similarity between a1 and a2, and then adds all the similarities together.

[0081] For the similarity of relational structures, a GCN graph convolutional network is used for calculation, such as... Figure 3 As shown, entity alignment is based on the distance between the outputs of the two graph GCNs. Let entity e1 be in graph G1 and entity e2 be in graph G2. Calculate the distance:

[0082]

[0083] Where f(x,y) is the 1 norm (the sum of the absolute values ​​of the vector elements), hs is the structural embedding of the entity, ds is the output dimension, and β is a hyperparameter used to measure the distance between structures and other parameters. The more similar the entities are, the closer they are.

[0084]

[0085] s = s str +s attr +s rel

[0086] The similarity between entity type information, relation information, and the standards in the SNOMED-CT medical terminology standard is calculated. If the similarity is greater than a threshold, the medical entity in this entity type information is considered to be the same medical entity as the one in SNOMED-CT, and they are aligned. Without introducing a large number of manual features, equivalent entity alignment pairs are automatically extracted from a large-scale heterogeneous knowledge graph, saving the cost of manual entity alignment and annotation.

[0087] Similarly, if the similarity between the two is less than or equal to the set threshold, then the medical entity in the entity type information is considered to be a separate medical entity that is not covered by SNOMED-CT, and the medical text containing this medical entity is retained.

[0088] The alignment operation described above requires extracting entity information, relation information, and using SNOMED-CT to obtain the SPO triples, namely the main entity, the object entity, and the relation. However, these are only the constituent elements of knowledge; further processing is required. First, an ontology is established in the medical field, considering common medical terms and related concepts, such as pneumonia. Here, the focus is on clinical manifestations and treatment plans. The ontology can be generated manually (using ontology editing software) or automatically using computer-aided, data-driven methods. Finally, it is modified and verified through algorithmic evaluation and manual review.

[0089] Next, a quality rating is conducted, which quantifies the credibility of the knowledge. Those with lower credibility are identified, effectively ensuring the quality of the knowledge.

[0090] Finally, the knowledge graph is constructed, visualized, and applied. For example, neo4j and py2neo can be used to build the knowledge graph. Neo4j is an open-source graph database, while py2neo is a Python module. Python is used to manipulate the neo4j graph database. The graph database is created using the `Graph()` method, nodes are created using the `Node()` method, and relationships are created using the `Relationship()` method. Finally, a web interface is used for visualization and related application development.

[0091] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0092] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0093] Based on the same technical concept, and corresponding to any of the above embodiments, this application also provides a pipeline-type highly robust medical knowledge graph standardization construction device.

[0094] like Figure 5 As shown, a pipeline-type robust medical knowledge graph standardization construction device includes: an entity extraction module 501, used to suspend and group all candidate entity word spans in medical texts, and extract entity type information from the candidate entity word spans;

[0095] The relation extraction module 502 extracts relations from entity type information, predicts the category of each relation between entity type information, and outputs relation information.

[0096] The medical knowledge graph construction module 503 is used to align entity type information, relationship information, and medical terminology standards to construct a medical knowledge graph.

[0097] Storage module 504 is used to store entity type information and relationship information.

[0098] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0099] The apparatus described above is used to implement a pipeline-based, highly robust medical knowledge graph standardization construction method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0100] Based on the same technical concept, corresponding to any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a pipeline-based, highly robust medical knowledge graph standardization construction method as described in any of the above embodiments.

[0101] Figure 6 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0102] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0103] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0104] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0105] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0106] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0107] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0108] The electronic devices described above are used to implement the corresponding medical knowledge graph construction methods in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0109] Based on the same technical concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium that stores computer instructions for causing the computer to execute a pipelined, highly robust medical knowledge graph standardization construction method as described in any of the above embodiments.

[0110] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0111] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute a pipeline-based, highly robust medical knowledge graph standardization construction method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0112] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0113] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0114] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0115] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A pipeline-based, highly robust medical knowledge graph standardization construction method, characterized in that, The method includes: Each character in the medical text is used as a start word, and the start word and the characters following it are used as end words. An attention mask matrix is ​​applied to suspend all candidate entity words sequentially, with the span from the start word to the end word as the distance. Each suspending mark is visible to another suspending mark within the same span in the attention mask matrix, but invisible to suspending marks in other spans. The suspending marks of all candidate entity word spans are appended to the medical text as a training instance, and the sequentially suspended marks are sorted. The suspending marks are divided into groups of size K, with adjacent spans preferentially grouped into the same group, where K is the maximum number of spans in a group. The suspending marks of each group are processed in parallel, and a pre-trained language model is used to obtain a contextualized representation of the entity word span in each group. Based on the contextualized representation, a neural network structure is used to output entity type information. Based on the contextualized representation and the entity type information encoding, the relationship categories between the entity type information are obtained, and the relationship information is output. By aligning the entity type information and relationship information with medical terminology standards, a medical information knowledge graph is constructed.

2. The method according to claim 1, characterized in that, Based on the contextualized representation and the entity type information encoding, the relationship categories between the entity type information are obtained, and the relationship information is output, including: Obtain the contextualized representation vector of the entity word span in each group obtained using a pre-trained language model; The entity type information is obtained and encoded into a vector of fixed dimensions; By concatenating the contextualized representation vector and the entity type information into a fixed-dimensional vector, the category of each relationship between the subject entity and the object entity in the entity type information is predicted.

3. The method according to claim 2, characterized in that, The method of concatenating the contextualized representation vector and the entity type information into a fixed-dimensional vector predicts the category of each relationship between the subject entity and the object entity in the entity type information, including: The concatenated contextualized representation vector and the entity type information are encoded into fixed-dimensional vectors and passed in parallel through FFN-head and FFN-tail based on feedforward neural networks to obtain the FFN-head output vector and FFN-tail output vector, respectively. The FFN-head output vector and the FFN-tail output vector are passed through a classification network to predict the category of each relationship between the main entity and the guest entity.

4. The method according to claim 1, characterized in that, Aligning the entity type information and the relationship information with medical terminology standards includes: By using frame matching, combined with element-level and structure-level matching, string similarity is determined. Similarity of attribute sets Similarity of relational structures The similarity between the entity type information and the medical terminology standard is calculated comprehensively; If the similarity between the entity type information and the medical terminology standard is greater than a threshold, then the medical entity in the entity type information and the medical terminology standard are considered to be the same medical entity. If the similarity between the entity type information and the medical terminology standard is less than and / or equal to a threshold, then the medical entity in the entity type information is considered to be a different medical entity from the medical terminology standard, and the original medical text is retained.

5. The method according to claim 1, characterized in that, Before aligning the entity type information and the relationship information, the method further includes: The entity type information and the relationship information are stored using Neo4j.

6. A pipeline-type highly robust medical knowledge graph construction device, characterized in that, include: The entity extraction module is used to take each character in the medical text as a start word and the start word and the characters following the start word as end words. An attention mask matrix is ​​applied to suspend all candidate entity words sequentially, with the span from the start word to the end word as the distance. Each suspended marker is visible to another suspended marker within the same span in the attention mask matrix, but invisible to suspended markers in other spans. The suspended markers for all candidate entity word spans are appended to the medical text as a training instance, and the sequentially suspended markers are sorted. The suspended markers are divided into groups of size K, with adjacent spans preferentially grouped into the same group, where K is the maximum span size in a group. The suspended markers in each group are processed in parallel, and a pre-trained language model is used to obtain a contextualized representation of the entity word span in each group. Based on the contextualized representation, a neural network structure is used to output entity type information. The relation extraction module is used to obtain the relationship categories between the entity type information based on the contextualized representation and the entity type information encoding, and output the relation information. The medical knowledge graph construction module is used to align the entity type information, the relationship information, and the medical terminology standard to construct a medical information knowledge graph; The storage module is used to store the entity type information and the relationship information.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method as described in any one of claims 1 to 5.

Images