Biomedical event assembly methods, apparatuses, devices, and media

Abstract

This application discloses a biomedical event assembly method, apparatus, device, and medium, relating to the field of natural language processing technology. The method includes: combining trigger words and arguments of the text to be assembled based on their trigger word types to obtain candidate events; tagging the trigger words and arguments of the candidate events using a nested tagging method to obtain candidate instances; encoding the candidate instances using a preset deep learning model to obtain an output representation of the semantic information of the candidate instances; acquiring a first probability result for a preset legal category and a second probability result for a preset illegal category of the output representation; and then determining a target biomedical event from the candidate events based on the first and second probability results. This method can improve the performance of biomedical event assembly.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to biomedical event assembly methods, apparatus, equipment and media. Background Technology

[0002] Biomedical event extraction refers to the automatic extraction of events from biomedical literature and their representation as structured information. Biomedical event information is of great significance for the development of new drugs, precision medicine, and basic biomedical research.

[0003] Events have fixed event types, trigger words, and arguments. The event trigger word identifies the occurrence of the event, while the arguments represent the participants in the event. For example, the sentence "Interestingly, we observed a specific four-fold upregulation of Id1 mRNA in BMP-6-treated B cell" contains a positive regulation event. The trigger word is "upregulation," and it contains two arguments: one is "Id1," which acts as a theme, indicating the subject of the event, and the other is "BMP-6," which acts as a cause, indicating the protein entity that caused the event. In the biomedical field, events such as gene expression and transcription have relatively simple argument composition, containing only one thematic role entity argument, and are called simple events. Binding events can have one or more thematic role entity arguments, and are called multi-faceted events. Regulation, positive regulation, and negative regulation events have the most complex argument composition, requiring one thematic role argument and 0-1 causal role arguments. These arguments may be biomedical entities within the sentence or other events within the sentence; these events are called nested events. Biomedical event assembly is a crucial step in biomedical event extraction. Its purpose is to assemble the trigger words and their arguments into structured event information using a specific method. Traditional event assembly methods assemble according to rules based on the argument type and number for different event types, and are called rule-based assembly. For simple events, rule-based assembly directly constructs an event from the trigger word and each of its thematic role arguments, which is relatively easy. However, for multi-element events and nested events with an indefinite number of arguments, rule assembly cannot adequately meet the assembly requirements of these two types of events, resulting in low performance of biomedical event assembly.

[0004] In summary, improving the performance of biomedical event assembly is a problem that needs to be solved in this field. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for assembling biomedical events, which can improve the performance of biomedical event assembly. The specific solution is as follows:

[0006] In a first aspect, this application discloses a method for assembling biomedical events, comprising:

[0007] Based on the trigger word type of the text to be assembled, the trigger words and arguments of the text to be assembled are combined to obtain the candidate events of the text to be assembled.

[0008] The trigger words and arguments of the candidate events are labeled using a nested labeling method to obtain candidate instances;

[0009] The candidate instances are encoded using a pre-defined deep learning model to obtain an output representation of the semantic information of the candidate instances;

[0010] The first probability result of the preset legal category and the second probability result of the preset illegal category are obtained from the output representation. Then, the target biomedical event is determined from the candidate events based on the first probability result and the second probability result.

[0011] Optionally, the step of combining the trigger words and arguments of the text to be assembled based on the trigger word type to obtain candidate events of the text to be assembled includes:

[0012] If the trigger word type of the text to be assembled is a preset multi-class event, then the entity arguments of the topic role are selected from all the arguments of the text to be assembled.

[0013] The trigger words of the text to be assembled and the entity arguments of the topic role are combined to obtain the candidate events of the text to be assembled.

[0014] Optionally, the step of combining the trigger words and arguments of the text to be assembled based on the trigger word type to obtain candidate events of the text to be assembled includes:

[0015] If the trigger word type of the text to be assembled is a preset nested event, then the arguments of the topic role and the arguments of the cause role are selected from all the arguments of the text to be assembled.

[0016] The trigger words, arguments of the thematic roles, and arguments of the causal roles of the text to be assembled are combined to obtain candidate events of the text to be assembled that include a first preset number of arguments of the thematic roles and a second preset number of arguments of the causal roles.

[0017] Optionally, the step of using nested tagging to tag the trigger words and arguments of the candidate events to obtain candidate instances includes:

[0018] The trigger words of the candidate events are marked using a nested tagging method, and it is determined whether the current argument of the candidate events meets the preset conditions.

[0019] If the conditions are met, the nested labeling method is used to perform a first labeling on the current argument of the candidate event to obtain a first candidate instance.

[0020] Optionally, after determining whether the current argument of the candidate event satisfies the preset condition, the method further includes:

[0021] If the conditions are not met, the nested labeling method is used to perform a second labeling on the current argument of the candidate event to obtain a second candidate instance.

[0022] Optionally, the step of encoding the candidate instance using a preset deep learning model to obtain an output representation of the semantic information of the candidate instance includes:

[0023] The entity arguments in the candidate instance are replaced with preset words to obtain a first replaced candidate instance, and the text in the first replaced candidate instance is replaced with preset symbols to obtain a second replaced candidate instance. Then, a preset classification symbol is inserted in a preset position in the second replaced candidate instance to obtain a candidate instance to be trained.

[0024] The candidate instances to be trained are iteratively trained using the BERT pre-trained language learning model in the preset deep learning model to obtain the output representation of the semantic information of the candidate instances to be trained.

[0025] Optionally, obtaining the first probability result of the preset legal category and the second probability result of the preset illegal category of the output representation includes:

[0026] The output representation is converted into a corresponding two-dimensional vector and normalized using the softmax activation function to obtain a first probability result of the preset legal category and a second probability result of the preset illegal category of the output representation.

[0027] Secondly, this application discloses a biomedical event assembly apparatus, comprising:

[0028] The candidate event acquisition module is used to combine the trigger words and arguments of the text to be assembled based on the trigger word type of the text to be assembled, so as to obtain the candidate events of the text to be assembled.

[0029] The candidate instance acquisition module is used to mark the trigger words and arguments of the candidate events using a nested tagging method to obtain candidate instances;

[0030] The output representation acquisition module is used to encode the candidate instance using a preset deep learning model to obtain the output representation of the semantic information of the candidate instance.

[0031] The target event determination module is used to obtain a first probability result of a preset legal category and a second probability result of a preset illegal category of the output representation, and then determine the target biomedical event from the candidate events based on the first probability result and the second probability result.

[0032] Thirdly, this application discloses an electronic device, including:

[0033] Memory, used to store computer programs;

[0034] A processor for executing the computer program to implement the steps of the aforementioned disclosed biomedical event assembly method.

[0035] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed biomedical event assembly method.

[0036] As can be seen, this application combines the trigger words and arguments of the text to be assembled based on the trigger word type of the text to be assembled to obtain candidate events of the text to be assembled; it uses a nested tagging method to tag the trigger words and arguments of the candidate events to obtain candidate instances; it uses a preset deep learning model to encode the candidate instances to obtain the semantic information output representation of the candidate instances; it obtains a first probability result of a preset legal category and a second probability result of a preset illegal category of the output representation, and then determines the target biomedical event from the candidate events based on the first probability result and the second probability result. Therefore, this application combines the trigger words and arguments of the text to be assembled based on the trigger word type of the text to be assembled to obtain candidate events of the text to be assembled; it uses a nested tagging method to consider the trigger words and arguments of the candidate events and tags them; it uses a preset deep learning model to obtain the output representation, which can fully capture the semantic information of the candidate instances, thus significantly improving the performance of biomedical event assembly; it uses the first probability result of the preset legal category and the second probability result of the preset illegal category of the output representation to determine the target biomedical event, that is, to determine the correctness of the candidate event, thereby achieving the purpose of assembling biomedical events. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0038] Figure 1 This is a flowchart of a biomedical event assembly method disclosed in this application;

[0039] Figure 2 This is a schematic diagram of a specific biomedical event assembly disclosed in this application;

[0040] Figure 3 This is a flowchart of a specific biomedical event assembly method disclosed in this application;

[0041] Figure 4 This is a schematic diagram of the structure of a biomedical event assembly device disclosed in this application;

[0042] Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0043] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0044] Biomedical event assembly is a crucial step in biomedical event extraction tasks. Its purpose is to assemble trigger words and their arguments into structured event information using a specific method. Traditional event assembly methods, known as rule-based assembly, assemble events according to rules governing the types and numbers of arguments for different event types. For simple events, rule-based assembly can directly construct an event from each trigger word and its arguments for each topic role, making assembly relatively easy. However, for multi-element events with an indefinite number of arguments and nested events, rule-based assembly cannot adequately meet the assembly requirements, resulting in lower performance in biomedical event assembly.

[0045] Therefore, this application provides a biomedical event assembly scheme that can improve the performance of biomedical event assembly.

[0046] See Figure 1 As shown in the embodiments, this application discloses a biomedical event assembly method, including:

[0047] Step S11: Based on the trigger word type of the text to be assembled, combine the trigger words and arguments of the text to be assembled to obtain candidate events of the text to be assembled.

[0048] In this embodiment, it can be understood that a candidate multivariate event includes a trigger word, one or more participating topic role arguments, and the other topic arguments of the trigger word are referred to as non-participating topic role arguments; while a candidate nested event includes a trigger word, a participating topic role argument, and an optional participating cause role argument, and the other topic arguments of the trigger word are also referred to as non-participating topic role arguments, and the other cause arguments are referred to as non-participating cause role arguments. It should be noted that the premise of biomedical event assembly is that the trigger word and arguments in the text to be assembled have been obtained, where the trigger word includes its corresponding event type, and the argument includes its corresponding trigger word role. The trigger word refers to a word or phrase appearing at a certain position in the text, and the argument refers to an entity mention or other pre-assembled event appearing at a certain position in the text to be assembled.

[0049] Step S12: Use nested tagging to tag the trigger words and arguments of the candidate events to obtain candidate instances.

[0050] It is understood that in this embodiment, when using nested tagging to tag the trigger words and arguments of candidate events, the trigger words are tagged with the symbol "#", that is, "#" is inserted before and after the trigger word. Arguments that participate in the topic role are tagged with "@", arguments that do not participate in the topic role are tagged with "$", arguments that participate in the cause role are tagged with "&", and arguments that do not participate in the cause role are tagged with "%".

[0051] Step S13: Encode the candidate instance using a preset deep learning model to obtain the output representation of the semantic information of the candidate instance.

[0052] In this embodiment, the step of encoding the candidate instances using a preset deep learning model to obtain the output representation of the semantic information of the candidate instances specifically includes: replacing entity arguments in the candidate instances with preset words to obtain a first replaced candidate instance, replacing the text in the first replaced candidate instance with preset symbols to obtain a second replaced candidate instance, and then inserting preset classification symbols into preset positions in the second replaced candidate instance to obtain a candidate instance to be trained; iteratively training the candidate instance to be trained using the BERT pre-trained language learning model in the preset deep learning model to obtain the output representation of the semantic information of the candidate instance to be trained. For example... Figure 2 As shown, the first step is to blind the entity arguments involved in the candidate instance text, for example, by replacing the entity arguments with "gene" to obtain the first replaced candidate instance. The second step is to symbolize the text and convert it into a word segment sequence to obtain the second replaced candidate instance. The third step is to insert the classification symbol [CLS] at the beginning to obtain the candidate instance to be trained. The fourth step is to input the candidate instance to be trained into the BERT (Bidirectional Encoder Representations from Transformer) pre-trained language model. This model contains multiple iterative modules from bottom to top. Each module uses a multi-head self-attention mechanism to transform the lower layer input and then output it to the upper layer. The final output representation of the model is the context representation corresponding to each word segment, that is, the output representation of the semantic information of the candidate instance.

[0053] Step S14: Obtain the first probability result of the preset legal category and the second probability result of the preset illegal category of the output representation, and then determine the target biomedical event from the candidate events based on the first probability result and the second probability result.

[0054] In this embodiment, obtaining the first probability result of the preset legal category and the second probability result of the preset illegal category of the output representation specifically includes: converting the output representation into a corresponding two-dimensional vector, and performing normalization calculation using the softmax activation function to obtain the first probability result of the preset legal category and the second probability result of the preset illegal category of the output representation. Using fully connected layers (FC) and a normalization module, binary classification is performed on the output representation corresponding to the classification symbol generated by the above module to determine whether the candidate event in the candidate instance constitutes a legal event. The fully connected layer converts the high-dimensional output representation corresponding to the classification symbol [CLS] into a two-dimensional vector, and then performs normalization using the softmax activation function to obtain the probabilities of the two types, namely the first probability result of the preset legal category and the second probability result of the preset illegal category. The class with the higher probability is selected as the final result, for example... Figure 2 The first probability result of the current candidate event is 0.98, and the second probability result is 0.02. Therefore, the first probability result of 0.98 is the final result, and the current candidate event is determined to be a legitimate biomedical event. That is, the combination of the trigger word and the argument is a correct biomedical event, and the current candidate event is the target biomedical event.

[0055] As can be seen, this application combines the trigger words and arguments of the text to be assembled based on the trigger word type of the text to be assembled to obtain candidate events of the text to be assembled; it uses a nested tagging method to tag the trigger words and arguments of the candidate events to obtain candidate instances; it uses a preset deep learning model to encode the candidate instances to obtain the semantic information output representation of the candidate instances; it obtains a first probability result of a preset legal category and a second probability result of a preset illegal category of the output representation, and then determines the target biomedical event from the candidate events based on the first probability result and the second probability result. Therefore, this application combines the trigger words and arguments of the text to be assembled based on the trigger word type of the text to be assembled to obtain candidate events of the text to be assembled; it uses a nested tagging method to consider the trigger words and arguments of the candidate events and tags them; it uses a preset deep learning model to obtain the output representation, which can fully capture the semantic information of the candidate instances, thus significantly improving the performance of biomedical event assembly; it uses the first probability result of the preset legal category and the second probability result of the preset illegal category of the output representation to determine the target biomedical event, that is, to determine the correctness of the candidate event, thereby achieving the purpose of assembling biomedical events.

[0056] This application discloses a specific method for assembling biomedical events. Compared to the previous embodiment, this embodiment further explains and optimizes the technical solution. See also... Figure 3 As shown, it specifically includes:

[0057] Step S21: Based on the trigger word type of the text to be assembled, combine the trigger words and arguments of the text to be assembled to obtain candidate events of the text to be assembled.

[0058] In this embodiment, the step of combining the trigger words and arguments of the text to be assembled based on the trigger word type to obtain candidate events of the text to be assembled specifically includes: if the trigger word type of the text to be assembled is a preset multi-type event, then filtering out entity arguments of the topic role from all arguments of the text to be assembled; combining the trigger words and entity arguments of the topic role to obtain candidate events of the text to be assembled. For preset multi-type events, such as binding events, this type of event only has entity arguments of the topic role, but the number can be one or more, so it is necessary to arbitrarily combine the trigger words and the entity arguments of all its topic roles.

[0059] In this embodiment, the step of combining the trigger words and arguments of the text to be assembled based on the trigger word type to obtain candidate events of the text to be assembled specifically includes: if the trigger word type of the text to be assembled is a preset nested event, then selecting arguments of the topic role and arguments of the cause role from all arguments of the text to be assembled; combining the trigger words, arguments of the topic role, and arguments of the cause role of the text to obtain candidate events of the text to be assembled containing a first preset number of arguments of the topic role and a second preset number of arguments of the cause role. For preset nested events, such as control events, including control, positive control, and negative control, it must contain arguments of the first preset number of topic roles and arguments of the second preset number of cause roles, that is, it must contain an argument of one topic role and 0 to 1 arguments of cause roles, and the arguments of both roles can be entities or another event. At this point, the trigger word is first combined with the arguments of each topic role to generate candidate events. If the trigger word also contains arguments of the cause role, then it is combined with the arguments of each topic role and the arguments of each cause role to form candidate events.

[0060] It is understandable that the text to be assembled can have multiple trigger words, and each trigger word has a different type, such as one being a preset multi-event and another being a preset nested event. For example, the text to be assembled is: "Specifically, Foxp3 physically interacts with RORgammat". 1 , and this interaction inhibitsRORgammat 2 The function contains a binding event E1 and a reverse control event E2. Superscripts are added to distinguish the two "RORgammat" entity arguments. The trigger word for binding event E1 is "interacts," and it contains arguments for two thematic roles: "Foxp3" and "RORgammat." 1 The reverse regulation event E2 is a pre-defined nested event whose trigger word is "inhibits" and contains an argument "RORgammat" for a thematic role. 2 Given an argument E1 representing the cause role of an event type, where E1 denotes the bound class event in the example sentence, the candidate event generation process is as follows:

[0061] (1) The trigger word "interacts" belongs to the bound event category. Therefore, the following candidate events are obtained by arbitrarily combining the trigger word with the entity arguments of all its thematic roles:

[0062] Candidate event 1: Trigger word: interactions, entity arguments of participating thematic roles: Foxp3, entity arguments of non-participating thematic roles: RORgammat1;

[0063] Candidate event 2: Trigger word: interactions, entity arguments of participating thematic roles: RORgammat1, entity arguments of non-participating thematic roles: Foxp3;

[0064] Candidate event 3: Trigger word: interactions, entity arguments of participating thematic roles: Foxp3, RORgammat1, entity arguments of non-participating thematic roles: none;

[0065] (2) The trigger word "inhibits" belongs to the reverse control type of event. Therefore, by combining the trigger word with the arguments of each of its thematic roles, the following candidate events are obtained:

[0066] Candidate event 4: Trigger word: inhibits, Argument involving the topic role: RORgammat2, Argument not involving the cause role: E1;

[0067] In addition, the trigger word also has a causal argument, so the trigger word needs to be combined with a thesis argument and a causal argument to obtain the following candidate events:

[0068] Candidate event 5: Trigger word: inhibits, Argument for participating topic role: RORgammat2, Argument for participating cause role: E1, Non-participating argument: None.

[0069] Step S22: Use nested tagging method to tag the trigger words of the candidate events, and determine whether the current argument of the candidate events meets the preset conditions.

[0070] In this embodiment, it can be understood that the preset condition is that the current argument participates in the candidate event.

[0071] Step S23: If the conditions are met, the current argument of the candidate event is marked with the nested marking method to obtain the first candidate instance.

[0072] In this embodiment, the arguments of the topic role in the candidate event are marked with "@", and the arguments of the cause role are marked with "&". If the argument is another sub-event, the trigger word and its participating arguments of the sub-event are marked not only in the sub-event but also nested according to the role of the sub-event in the candidate event.

[0073] Step S24: If the condition is not met, the nested labeling method is used to perform a second labeling on the current argument of the candidate event to obtain a second candidate instance.

[0074] In this embodiment, arguments of the topic role that do not participate in the candidate event are marked with "$", and arguments of the cause role are marked with "%". If the argument is another sub-event, only the trigger word of that sub-event is marked. It should be noted that the trigger word of the sub-event itself is also marked with "#".

[0075] Step S25: Encode the first candidate instance or the second candidate instance using a preset deep learning model to obtain the output representation of the semantic information of the first candidate instance or the second candidate instance.

[0076] In this embodiment, after candidate events 1 to 5 are labeled, the entity arguments are replaced with genes, the text is symbolically converted into word fragment sequences, and a classification symbol [CLS] is inserted at the beginning to obtain candidate instances 1 to 5 to be trained:

[0077] Candidate Example 1: "[CLS]Specifically, @gene@physically#interacts#with$gene$, and this interaction inhibits RORgammat 2 function."

[0078] Candidate instance 2: "[CLS]Specifically, $gene$physically#interacts#with@gene@, and this interaction inhibits RORgammat 2 function."

[0079] Candidate instance 3: “[CLS]Specifically, @gene@physically#interacts#with@gene@, and this interaction inhibits RORgammat 2 function."

[0080] Candidate instance 4: "[CLS]Specifically, gene physically%#interacts#%with gene, and this interaction#inhibits#@gene@function."

[0081] Candidate instance 5: “[CLS]Specifically, &@gene@&physically&#interacts#&with&@gene@&, and this interaction#inhibits#@gene@function.”

[0082] Step S26: Obtain the first probability result of the preset legal category and the second probability result of the preset illegal category of the output representation, and then determine the target biomedical event from the candidate events based on the first probability result and the second probability result.

[0083] For more detailed work procedures regarding step S26, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0084] Therefore, this application does not assemble candidate events based on simple rules, but requires different combinations based on different trigger word types to obtain different candidate events, thereby dealing with preset multi-class events and preset nested events, and improving the performance of biomedical event assembly.

[0085] See Figure 4 As shown in the figure, this application discloses a biomedical event assembly apparatus, comprising:

[0086] The candidate event acquisition module 11 is used to combine the trigger words and arguments of the text to be assembled based on the trigger word type of the text to be assembled, so as to obtain the candidate events of the text to be assembled.

[0087] The candidate instance acquisition module 12 is used to mark the trigger words and arguments of the candidate events using a nested tagging method to obtain candidate instances;

[0088] The output representation acquisition module 13 is used to encode the candidate instance using a preset deep learning model to obtain the output representation of the semantic information of the candidate instance.

[0089] The target event determination module 14 is used to obtain a first probability result of a preset legal category and a second probability result of a preset illegal category of the output representation, and then determine the target biomedical event from the candidate events based on the first probability result and the second probability result.

[0090] As can be seen, this application combines the trigger words and arguments of the text to be assembled based on the trigger word type of the text to be assembled to obtain candidate events of the text to be assembled; it uses a nested tagging method to tag the trigger words and arguments of the candidate events to obtain candidate instances; it uses a preset deep learning model to encode the candidate instances to obtain the semantic information output representation of the candidate instances; it obtains a first probability result of a preset legal category and a second probability result of a preset illegal category of the output representation, and then determines the target biomedical event from the candidate events based on the first probability result and the second probability result. Therefore, this application combines the trigger words and arguments of the text to be assembled based on the trigger word type of the text to be assembled to obtain candidate events of the text to be assembled; it uses a nested tagging method to consider the trigger words and arguments of the candidate events and tags them; it uses a preset deep learning model to obtain the output representation, which can fully capture the semantic information of the candidate instances, thus significantly improving the performance of biomedical event assembly; it uses the first probability result of the preset legal category and the second probability result of the preset illegal category of the output representation to determine the target biomedical event, that is, to determine the correctness of the candidate event, thereby achieving the purpose of assembling biomedical events.

[0091] In some specific embodiments, the candidate event acquisition module 11 includes:

[0092] The first combination unit is used to select entity arguments of topic roles from all arguments of the text to be assembled if the trigger word type of the text to be assembled is a preset multi-type event; and to combine the trigger words of the text to be assembled and the entity arguments of the topic roles to obtain candidate events of the text to be assembled.

[0093] In some specific embodiments, the candidate event acquisition module 11 includes:

[0094] The second combination unit is used to, if the trigger word type of the text to be assembled is a preset nested event, filter out the arguments of the topic role and the arguments of the cause role from all the arguments of the text to be assembled; combine the trigger word, the arguments of the topic role, and the arguments of the cause role of the text to be assembled to obtain a candidate event of the text to be assembled containing a first preset number of arguments of the topic role and a second preset number of arguments of the cause role.

[0095] In some specific embodiments, the candidate instance acquisition module 12 includes:

[0096] The first candidate instance acquisition unit is used to mark the trigger words of the candidate event using a nested marking method, and to determine whether the current argument of the candidate event meets a preset condition; if it does, the nested marking method is used to mark the current argument of the candidate event to obtain the first candidate instance.

[0097] In some specific embodiments, the candidate instance acquisition module 12 includes:

[0098] The second candidate instance acquisition unit is used to perform a second labeling on the current argument of the candidate event using the nested labeling method if the conditions are not met, so as to obtain a second candidate instance.

[0099] In some specific embodiments, the output representation acquisition module 13 includes:

[0100] The candidate instance acquisition unit is used to replace entity arguments in the candidate instance with preset words to obtain a first replaced candidate instance, and replace the text in the first replaced candidate instance with preset symbols to obtain a second replaced candidate instance, and then insert a preset classification symbol in a preset position in the second replaced candidate instance to obtain a candidate instance to be trained.

[0101] The iterative training unit is used to iteratively train the candidate instance to be trained using the BERT pre-trained language learning model in the preset deep learning model, so as to obtain the output representation of the semantic information of the candidate instance to be trained.

[0102] In some specific embodiments, the target event determination module 14 includes:

[0103] The probability result acquisition unit is used to convert the output representation into a corresponding two-dimensional vector and perform normalization calculation using the softmax activation function to obtain a first probability result of the preset legal category and a second probability result of the preset illegal category of the output representation.

[0104] Furthermore, embodiments of this application also provide an electronic device. Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0105] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Specifically, it may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the following steps:

[0106] Based on the trigger word type of the text to be assembled, the trigger words and arguments of the text to be assembled are combined to obtain the candidate events of the text to be assembled.

[0107] The trigger words and arguments of the candidate events are labeled using a nested labeling method to obtain candidate instances;

[0108] The candidate instances are encoded using a pre-defined deep learning model to obtain an output representation of the semantic information of the candidate instances;

[0109] The first probability result of the preset legal category and the second probability result of the preset illegal category are obtained from the output representation. Then, the target biomedical event is determined from the candidate events based on the first probability result and the second probability result.

[0110] In some specific embodiments, the processor executes a computer program stored in the memory, specifically implementing the following steps:

[0111] If the trigger word type of the text to be assembled is a preset multi-class event, then the entity arguments of the topic role are selected from all the arguments of the text to be assembled.

[0112] The trigger words of the text to be assembled and the entity arguments of the topic role are combined to obtain the candidate events of the text to be assembled.

[0113] In some specific embodiments, the processor executes a computer program stored in the memory, specifically implementing the following steps:

[0114] If the trigger word type of the text to be assembled is a preset nested event, then the arguments of the topic role and the arguments of the cause role are selected from all the arguments of the text to be assembled.

[0115] The trigger words, arguments of the thematic roles, and arguments of the causal roles of the text to be assembled are combined to obtain candidate events of the text to be assembled that include a first preset number of arguments of the thematic roles and a second preset number of arguments of the causal roles.

[0116] In some specific embodiments, the processor executes a computer program stored in the memory, specifically implementing the following steps:

[0117] The trigger words of the candidate events are marked using a nested tagging method, and it is determined whether the current argument of the candidate events meets the preset conditions.

[0118] If the conditions are met, the nested labeling method is used to perform a first labeling on the current argument of the candidate event to obtain a first candidate instance.

[0119] In some specific embodiments, the processor executes a computer program stored in the memory, specifically implementing the following steps:

[0120] If the conditions are not met, the nested labeling method is used to perform a second labeling on the current argument of the candidate event to obtain a second candidate instance.

[0121] In some specific embodiments, the processor executes a computer program stored in the memory, specifically implementing the following steps:

[0122] The entity arguments in the candidate instance are replaced with preset words to obtain a first replaced candidate instance, and the text in the first replaced candidate instance is replaced with preset symbols to obtain a second replaced candidate instance. Then, a preset classification symbol is inserted in a preset position in the second replaced candidate instance to obtain a candidate instance to be trained.

[0123] The candidate instances to be trained are iteratively trained using the BERT pre-trained language learning model in the preset deep learning model to obtain the output representation of the semantic information of the candidate instances to be trained.

[0124] In some specific embodiments, the processor, by executing a computer program stored in the memory, may further include the following steps:

[0125] The output representation is converted into a corresponding two-dimensional vector and normalized using the softmax activation function to obtain a first probability result of the preset legal category and a second probability result of the preset illegal category of the output representation.

[0126] In this embodiment, the power supply 23 is used to provide operating voltage for various hardware devices on the electronic device; the communication interface 24 can create a data transmission channel between the electronic device and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0127] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0128] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0129] The operating system 221 manages and controls the various hardware devices and computer programs 222 on the electronic device to enable the processor 21 to perform calculations and processing on the massive amounts of data 223 in the memory 22. The operating system can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the biomedical event assembly method executed by the electronic device as disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the electronic device from external devices, as well as data collected by its own input / output interface 25.

[0130] Furthermore, embodiments of this application also disclose a computer-readable storage medium storing a computer program, which, when loaded and executed by a processor, implements the method steps performed during the biomedical event assembly process disclosed in any of the foregoing embodiments.

[0131] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0132] The present invention provides a detailed description of a biomedical event assembly method, apparatus, device, and medium. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for assembling biomedical events, characterized in that, include: Based on the trigger word type of the text to be assembled, the trigger words and arguments of the text to be assembled are combined to obtain the candidate events of the text to be assembled. The trigger words and arguments of the candidate events are labeled using a nested labeling method to obtain candidate instances; The candidate instances are encoded using a pre-defined deep learning model to obtain an output representation of the semantic information of the candidate instances; Obtain a first probability result of a preset legal category and a second probability result of a preset illegal category from the output representation, and then determine the target biomedical event from the candidate events based on the first probability result and the second probability result; The step of using nested tagging to tag the trigger words and arguments of the candidate events to obtain candidate instances includes: The trigger words of the candidate events are marked using a nested tagging method, and it is determined whether the current argument of the candidate event meets a preset condition; if it does, the current argument of the candidate event is marked first using the nested tagging method to obtain a first candidate instance; the preset condition is that the current argument participates in the candidate event. The step of encoding the candidate instances using a preset deep learning model to obtain the output representation of the semantic information of the candidate instances includes: The entity arguments in the candidate instance are replaced with preset words to obtain a first replaced candidate instance. The text in the first replaced candidate instance is replaced with preset symbols to obtain a second replaced candidate instance. Then, a preset classification symbol is inserted into a preset position in the second replaced candidate instance to obtain a candidate instance to be trained. The candidate instance to be trained is iteratively trained using the BERT pre-trained language learning model in the preset deep learning model to obtain the output representation of the semantic information of the candidate instance to be trained.

2. The biomedical event assembly method according to claim 1, characterized in that, The method of combining the trigger words and arguments of the text to be assembled based on the trigger word type of the text to be assembled to obtain candidate events of the text to be assembled includes: If the trigger word type of the text to be assembled is a preset multi-class event, then the entity arguments of the topic role are selected from all the arguments of the text to be assembled. The trigger words of the text to be assembled and the entity arguments of the topic role are combined to obtain the candidate events of the text to be assembled.

3. The biomedical event assembly method according to claim 1, characterized in that, The method of combining the trigger words and arguments of the text to be assembled based on the trigger word type of the text to be assembled to obtain candidate events of the text to be assembled includes: If the trigger word type of the text to be assembled is a preset nested event, then the arguments of the topic role and the arguments of the cause role are selected from all the arguments of the text to be assembled. The trigger words, arguments of the thematic roles, and arguments of the causal roles of the text to be assembled are combined to obtain candidate events of the text to be assembled that include a first preset number of arguments of the thematic roles and a second preset number of arguments of the causal roles.

4. The biomedical event assembly method according to claim 1, characterized in that, After determining whether the current argument of the candidate event satisfies the preset conditions, the method further includes: If the conditions are not met, the nested labeling method is used to perform a second labeling on the current argument of the candidate event to obtain a second candidate instance.

5. The biomedical event assembly method according to any one of claims 1 to 4, characterized in that, The step of obtaining the first probability result of the preset legal category and the second probability result of the preset illegal category of the output representation includes: The output representation is converted into a corresponding two-dimensional vector and normalized using the softmax activation function to obtain a first probability result of the preset legal category and a second probability result of the preset illegal category of the output representation.

6. A biomedical event assembly device, characterized in that, The steps for implementing the biomedical event assembly method as described in any one of claims 1 to 5 include: The candidate event acquisition module is used to combine the trigger words and arguments of the text to be assembled based on the trigger word type of the text to be assembled, so as to obtain the candidate events of the text to be assembled. The candidate instance acquisition module is used to mark the trigger words and arguments of the candidate events using a nested tagging method to obtain candidate instances; The output representation acquisition module is used to encode the candidate instance using a preset deep learning model to obtain the output representation of the semantic information of the candidate instance. The target event determination module is used to obtain a first probability result of a preset legal category and a second probability result of a preset illegal category of the output representation, and then determine the target biomedical event from the candidate events based on the first probability result and the second probability result.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the biomedical event assembly method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the biomedical event assembly method as described in any one of claims 1 to 5.

Images