Genome variation named entity recognition method and system based on feature fusion and multi-task learning

Abstract

The invention relates to a genome variation named entity recognition method and system based on feature fusion and multi-task learning. The method comprises the steps of obtaining biomedical text data with genome variation entity labels; performing data enhancement on the biomedical text data to generate diversified enhanced biomedical text data; training a named entity recognition model based on the enhanced biomedical text data, and performing named entity recognition on the to-be-recognized biomedical text through the trained named entity recognition model to obtain a recognition result; the named entity recognition model comprises an attention-enhanced multi-scale feature fusion module, a first BILSTM layer, an NER module, a token classification module and an entity boundary division module; according to the method, the problem of data scarcity is solved, meanwhile, the attention-enhanced multi-scale feature fusion module is provided to solve the problem that feature representation is inaccurate due to the fact that the entity length is too long and character features are complex, and the recognition accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of entity recognition, and in particular relates to a genome variation named entity recognition method and system based on feature fusion and multi-task learning. Background Art

[0002] Named Entity Recognition (NER) refers to the process of locating entity boundaries from unstructured text and correctly classifying them into predefined categories. It is an important task of information extraction. With the development of biomedical and Internet technologies, biomedical literature has shown an exponential growth trend. Named entity recognition has become a new research opportunity to mine and extract valuable information from massive literature. Genomic variant entity recognition aims to extract genomic variant entities from biomedical literature and classify them.

[0003] Traditional methods for identifying genomic variant entities in biomedical literature include rule-based and dictionary-based methods and machine learning-based methods. Rule-based and dictionary-based methods detect entity boundaries by constructing feature rules and regular expressions, and then use dictionaries to classify entities, such as MutationFinder, EMU and other tools. However, this method is time-consuming and labor-intensive, and is heavily dependent on the completeness of rules and dictionaries, and has poor generalization ability; machine learning-based methods classify entities through feature extraction and classifiers. For example, tmVar inputs features into CRF for prediction, and then uses a regular expression-based post-processing module to reduce false negatives; Avada combines regular expressions and machine learning to extract genomic variant information; however, this type of method requires sufficient feature engineering skills and domain knowledge, and it is difficult to capture complex and deep data features. Models based on deep learning can often overcome the above problems, but the following two problems still exist:

[0004] (1) The complex characteristics of genomic variant entities lead to inaccurate recognition results: Different from named entities in general fields, genomic variant entities have diverse naming rules, complex character characteristics, and some entities are long, such as IVS8+4A>G, p.Pro246HisfsX13, and Trp-241was replaced with Ala or Gln. Their complexity greatly increases the difficulty of entity recognition. In the current biomedical named entity recognition methods based on deep learning, the model is not designed according to the characteristics of genomic variant entities, and the implicit features of the text cannot be effectively extracted, resulting in poor recognition of some entities. For example, the DeepVar model uses a general field label alignment method during preprocessing, which exacerbates label imbalance and label discontinuity.

[0005] (2) Data scarcity: The size and quality of the training data set often have a decisive impact on the performance of the model. In the biomedical field, there is often a problem of scarce training data sets. In particular, high-quality data sets related to gene mutations are very sparse. Solving the problem of scarce training data is crucial to improving the overall performance of the model. Summary of the invention

[0006] In order to solve the problems existing in the background technology, one aspect of the present invention provides a genome variation named entity recognition method based on feature fusion and multi-task learning, comprising:

[0007] S1: Obtain biomedical text data with genomic variant entity annotations;

[0008] S2: Perform data augmentation on biomedical text data to generate diverse augmented biomedical text data;

[0009] S3: training a named entity recognition model based on enhanced biomedical text data, and performing named entity recognition on the biomedical text to be recognized by using the trained named entity recognition model to obtain a recognition result; the named entity recognition model includes: an attention-enhanced multi-scale feature fusion module, a first BILSTM layer, a NER module, a token classification module, and an entity boundary division module;

[0010] The attention-enhanced multi-scale feature fusion module is used to extract a first comprehensive feature vector sequence of the input biomedical text; the first BILSTM layer is used to integrate position information into the first comprehensive feature vector sequence to generate a second comprehensive feature vector sequence; the NER module is mainly used to perform named entity recognition on the input biomedical text according to the second comprehensive feature vector sequence; the token classification module and the entity boundary division module are respectively used to construct a token classification subtask and an entity boundary division subtask according to the second comprehensive feature vector sequence to assist the training of the named entity recognition model.

[0011] Preferably, the data enhancement of the biomedical text data includes: separating the text in the biomedical text from the genomic variation entities; performing one or more operations of case conversion, synonym replacement, random deletion or random insertion on the separated text to obtain enhanced text; introducing the separated genomic variation entities into the dbSNP and / or DisGenNET databases, and performing synonym replacement on the genomic variation entities; recombining the processed enhanced text and the genomic variation entities after synonym replacement to obtain enhanced biomedical text; and combining the enhanced biomedical text and the original biomedical text data to form enhanced biomedical text data.

[0012] Preferably, the attention-enhanced multi-scale feature fusion module comprises: a pre-trained BIOBERT model, a character random embedding layer, a CNN module, a second BILSTM layer, an attention module, a maximum pooling layer and a feature fusion layer;

[0013] The character random embedding layer is used to generate random embeddings for characters in the input biomedical text to obtain a character embedding representation sequence for each word in the biomedical text;

[0014] The BIOBERT model is used to obtain word-level features of input biomedical text;

[0015] The CNN module and the maximum pooling layer are cascaded to obtain local character features of each word in the biomedical text according to the character embedding representation sequence of each word in the biomedical text;

[0016] The second BILSTM layer and the attention module are cascaded to obtain global character-level features of each word of the biomedical text according to the character embedding representation sequence of the biomedical text;

[0017] The feature fusion layer is used to fuse the word-level features, local character features and global character-level features of each word in the biomedical text to obtain a first comprehensive feature vector sequence of the biomedical text.

[0018] Preferably, the feature fusion of the word-level features, local character features and global character-level features of each word in the biomedical text comprises:

[0019]

[0020] Among them, h i Represents the first comprehensive feature vector of the i-th word; and They represent the global character-level features, local character features, and word-level features of the i-th word respectively; λ1, λ2, and λ3 represent the weight coefficients of each type of feature.

[0021] Preferably, the training process of the named entity recognition model includes:

[0022] The first comprehensive feature vector sequence is respectively input into the NER module, token classification module and entity boundary division module to perform named entity recognition task, token classification task and entity boundary division task on the input biomedical text; the corresponding cross entropy loss function is constructed based on the prediction result of each task and the corresponding true label, the total loss function of the model is constructed based on the loss function of each task, and the parameters of the named entity recognition model are tuned based on the constructed total loss function of the model.

[0023] Preferably, the total loss function Loss of the model includes:

[0024] Loss = α·L NER +β·L Boundary +γ·L Token

[0025] Among them, L NER represents the loss function of the named entity recognition task, L Boundary represents the loss function of the entity boundary segmentation task; L Token Represents the loss function for the token classification task.

[0026] Preferably, the cross entropy loss function includes:

[0027]

[0028] Among them, L∈{L NER ,L Boundary ,L Token},y i represents the true label distribution, represents the label distribution predicted by the model, k represents the number of different label categories in the dataset, ω i is the weight corresponding to label i.

[0029] Another aspect of the present invention provides a genome variation named entity recognition system based on feature fusion and multi-task learning, wherein the system is applied to the genome variation named entity recognition method based on feature fusion and multi-task learning, and comprises: a data enhancement module, a multi-task model training module, a data acquisition module and a classification module;

[0030] A data enhancement module is used to enhance the biomedical text data to generate diversified enhanced biomedical text data;

[0031] A multi-task model training module for training named entity recognition models based on augmented biomedical text data;

[0032] A data acquisition module, used for acquiring biomedical text to be recognized in real time;

[0033] The classification module is used to perform named entity recognition on the biomedical text to be identified using the trained named entity recognition model.

[0034] Another aspect of the present invention provides a genomic variation named entity recognition device based on feature fusion and multi-task learning, comprising a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory, and is used to execute the computer program stored in the memory, so that the genomic variation named entity recognition device based on feature fusion and multi-task learning performs the genomic variation named entity recognition method based on feature fusion and multi-task learning.

[0035] Another aspect of the present invention provides a computer-readable storage medium storing a program, which, when executed by a processor, implements the method for genomic variation named entity recognition based on feature fusion and multi-task learning.

[0036] The present invention has at least the following beneficial effects

[0037] According to the characteristics of scarce training data and genomic variant entities, the present invention proposes a data enhancement method and introduces an attention-enhanced multi-scale feature fusion model to capture implicit information in the context from multiple levels and improve the ability to extract text features. At the same time, the method based on the multi-task learning strategy divides the original task into entity boundary recognition and token classification subtasks, and improves the overall performance and generalization ability of the model through the correlation between tasks. This method effectively addresses the problems of data scarcity, complex characters of genomic variant entities, long entity length, and diverse naming rules, and significantly improves the accuracy of recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0039] Figure 2 It is a schematic diagram of the method framework of the present invention;

[0040] Figure 3 It is a schematic diagram of the framework of the named entity recognition model of the present invention;

[0041] Figure 4 A schematic diagram of loss function construction of the present invention. DETAILED DESCRIPTION

[0042] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0043] See also Figure 1 to Figure 4 The present invention provides a genome variation named entity recognition method based on feature fusion and multi-task learning, comprising:

[0044] S1: Obtain biomedical text data with genomic variant entity annotations;

[0045] S2: Perform data augmentation on biomedical text data to generate diverse augmented biomedical text data;

[0046] S3: training a named entity recognition model based on enhanced biomedical text data, and performing named entity recognition on the biomedical text to be recognized by using the trained named entity recognition model to obtain a recognition result; the named entity recognition model includes: an attention-enhanced multi-scale feature fusion module, a first BILSTM layer, a NER module, a token classification module, and an entity boundary division module;

[0047] The attention-enhanced multi-scale feature fusion module is used to extract a first comprehensive feature vector sequence of the input biomedical text; the first BILSTM layer is used to integrate position information into the first comprehensive feature vector sequence to generate a second comprehensive feature vector sequence; the NER module is mainly used to perform named entity recognition on the input biomedical text according to the second comprehensive feature vector sequence; the token classification module and the entity boundary division module are respectively used to construct a token classification subtask and an entity boundary division subtask according to the second comprehensive feature vector sequence to assist the training of the named entity recognition model.

[0048] In one embodiment, considering that the size and quality of the training data set often have a significant impact on the performance of the model, data enhancement technology is used to alleviate the problem of data scarcity, such as Figure 2 As shown, the processing flow includes:

[0049] The text in the biomedical text is separated from the genomic variation entities; one or more operations of case conversion, synonym replacement, random deletion or random insertion are performed on the separated text to obtain enhanced text; the separated genomic variation entities are introduced into the dbSNP and / or DisGenNET databases to perform synonym replacement on the genomic variation entities; the processed enhanced text and the genomic variation entities after synonym replacement are recombined to obtain enhanced biomedical text; the enhanced biomedical text and the original biomedical text data are combined to form enhanced biomedical text data.

[0050] In one embodiment, in order to effectively mine the complex feature information implicit in biomedical text, an attention-enhanced multi-scale feature fusion method is proposed to learn features of different levels from different angles of the input biomedical text, and use the attention mechanism to selectively combine CNN and BiLSTM to extract specific features in the sequence, including: global character-level feature sequence, local character feature sequence and word-level feature sequence, such as Figure 4 shown.

[0051] Preferably, the attention-enhanced multi-scale feature fusion module comprises: a pre-trained BIOBERT model, a character random embedding layer, a CNN module, a second BILSTM layer, an attention module, a maximum pooling layer and a feature fusion layer;

[0052] The character random embedding layer is used to generate random embeddings for characters in the input biomedical text to obtain a character embedding representation sequence for each word in the biomedical text;

[0053] The BIOBERT model is used to obtain word-level features of input biomedical text;

[0054] The CNN module and the maximum pooling layer are cascaded to obtain local character features of each word in the biomedical text according to the character embedding representation sequence of each word in the biomedical text;

[0055] The second BILSTM layer and the attention module are cascaded to obtain global character-level features of each word of the biomedical text according to the character embedding representation sequence of the biomedical text;

[0056] The feature fusion layer is used to fuse the word-level features, local character features and global character-level features of each word in the biomedical text to obtain a first comprehensive feature vector sequence of the biomedical text.

[0057] After randomly initializing the character embedding, BILSTM and attention mechanism are combined to extract entity-related global character-level features. When the t-th word passes through the BILSTM network to obtain its hidden state, its context is represented as:

[0058]

[0059] in, and They represent the forward and backward outputs at time t, respectively. Then, the self-attention mechanism is used to effectively capture the relationship between the two representations. Get the global character-level features as input The expression is as follows:

[0060]

[0061] Among them, Softmax is the normalized index, c t is the context vector, and let Get the attention weights, W1, W2, W C is the weight matrix, are parameter vectors, each of which is randomly initialized.

[0062] Then, CNN and maximum pooling operations are combined to extract local character feature sequences. The local processing capability of the CNN network is used to process the input character embedding representation sequence, and the information of the redundant filling position is blocked to extract the character feature sequence, as follows:

[0063]

[0064] Among them, U is the convolution kernel size, * indicates the convolution operation, and Mask indicates the zero padding of the input sequence. Then the local salient features are extracted through the maximum pooling operation. The local character features at time t are expressed as follows:

[0065]

[0066] in, Represents local word-level representation.

[0067] In addition, character feature sequences are crucial for named entity recognition. The present invention uses the pre-trained model BIOBERT in the biomedical field to obtain word-level features. Then, a series strategy is used to weight the fusion of multi-level features to obtain a richer unified feature representation to better capture semantic and syntactic information. The final fusion representation is as follows:

[0068]

[0069] Among them, h i Represents the first comprehensive feature vector of the i-th word; and They represent the global character-level features, local character features, and word-level features of the i-th word respectively; λ1, λ2, and λ3 represent the weight coefficients of each type of feature.

[0070] Preferably, the training process of the named entity recognition model includes:

[0071] The first comprehensive feature vector sequence is respectively input into the NER module, token classification module and entity boundary division module to perform named entity recognition task, token classification task and entity boundary division task on the input biomedical text; the corresponding cross entropy loss function is constructed based on the prediction result of each task and the corresponding true label, the total loss function of the model is constructed based on the loss function of each task, and the parameters of the named entity recognition model are tuned based on the constructed total loss function of the model.

[0072] Preferably, the total loss function Loss of the model includes:

[0073] Loss = α·L NER +β·L Boundary +γ·L Token

[0074] Among them, L NER represents the loss function of the named entity recognition task, L Boundary represents the loss function of the entity boundary segmentation task; L Token Represents the loss function for the token classification task.

[0075] In one embodiment, in order to effectively solve the problem of inaccurate recognition due to the long length and complex features of genomic variant entities, the original task is divided into entity boundary recognition and token classification subtasks based on the multi-task learning strategy, such as Figure 3 As shown in the figure, it includes three tasks: named entity recognition task, entity boundary demarcation task, and token classification task; the entity boundary demarcation task is to identify the start and end positions of the entity; the token classification task is to classify each token;

[0076] Among them, the named entity recognition task uses CRF as the label decoder, and the task loss function is L NER , the entity boundary segmentation task and token classification task use softmax as the label decoder, and the loss function is L Boundary and L Token In addition, due to the serious imbalance problem of the data set, the number of "non-entity" labels is far greater than the number of "entity" labels. Therefore, when designing the loss function, weights are assigned to different labels to reduce the excessive impact of "non-entity" on the loss function. The weighted cross entropy loss formula is as follows:

[0077]

[0078] Among them, L∈{L NER ,L Boundary ,L Token},y i represents the true label distribution, represents the label distribution predicted by the model, k represents the number of different label categories in the dataset, ω i is the weight corresponding to label i.

[0079] Another aspect of the present invention provides a genome variation named entity recognition system based on feature fusion and multi-task learning, wherein the system is applied to the genome variation named entity recognition method based on feature fusion and multi-task learning, and comprises: a data enhancement module, a multi-task model training module, a data acquisition module and a classification module;

[0080] A data enhancement module is used to enhance the biomedical text data to generate diversified enhanced biomedical text data;

[0081] A multi-task model training module for training named entity recognition models based on augmented biomedical text data;

[0082] A data acquisition module, used for acquiring biomedical text to be recognized in real time;

[0083] The classification module is used to perform named entity recognition on the biomedical text to be identified using the trained named entity recognition model.

[0084] Another aspect of the present invention provides a genomic variation named entity recognition device based on feature fusion and multi-task learning, comprising a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory, and is used to execute the computer program stored in the memory, so that the genomic variation named entity recognition device based on feature fusion and multi-task learning performs the genomic variation named entity recognition method based on feature fusion and multi-task learning.

[0085] Another aspect of the present invention provides a computer-readable storage medium storing a program, which, when executed by a processor, implements the method for genomic variation named entity recognition based on feature fusion and multi-task learning.

[0086] In summary, according to the scarcity of training data and the characteristics of genomic variant entities, the present invention proposes a data enhancement method and introduces an attention-enhanced multi-scale feature fusion model to capture implicit information in the context from multiple levels and improve the ability to extract text features. At the same time, the method based on the multi-task learning strategy divides the original task into entity boundary recognition and token classification subtasks, and improves the overall performance and generalization ability of the model through the correlation between tasks. This method effectively addresses the problems of data scarcity, complex characters of genomic variant entities, long entity length, and diverse naming rules, and significantly improves the accuracy of recognition.

[0087] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A genome variation named entity recognition method based on feature fusion and multi-task learning, characterized in that: include: S1: Obtain biomedical text data with genomic variant entity annotations; S2: Perform data augmentation on biomedical text data to generate diverse augmented biomedical text data; S3: Train the named entity recognition model based on the enhanced biomedical text data, and perform named entity recognition on the biomedical text to be recognized through the trained named entity recognition model to obtain the recognition result; The named entity recognition model includes: an attention-enhanced multi-scale feature fusion module, a first BILSTM layer, a NER module, a token classification module, and an entity boundary division module; The attention-enhanced multi-scale feature fusion module is used to extract a first comprehensive feature vector sequence of the input biomedical text; the first BILSTM layer is used to integrate position information into the first comprehensive feature vector sequence to generate a second comprehensive feature vector sequence; the NER module is mainly used to perform named entity recognition on the input biomedical text according to the second comprehensive feature vector sequence; the token classification module and the entity boundary division module are respectively used to construct a token classification subtask and an entity boundary division subtask according to the second comprehensive feature vector sequence to assist the training of the named entity recognition model.

2. According to claim 1, a genome variation named entity recognition method based on feature fusion and multi-task learning is characterized in that: The data enhancement of biomedical text data includes: separating text from genomic variation entities in the biomedical text; performing one or more operations of case conversion, synonym replacement, random deletion or random insertion on the separated text to obtain enhanced text; introducing the separated genomic variation entities into dbSNP and / or DisGenNET databases, and performing synonym replacement on the genomic variation entities; recombining the processed enhanced text and the genomic variation entities after synonym replacement to obtain enhanced biomedical text; and combining the enhanced biomedical text and original biomedical text data to form enhanced biomedical text data.

3. According to the method for genome variation named entity recognition based on feature fusion and multi-task learning in claim 1, it is characterized in that: The attention-enhanced multi-scale feature fusion module includes: a pre-trained BIOBERT model, a character random embedding layer, a CNN module, a second BILSTM layer, an attention module, a max pooling layer, and a feature fusion layer; The character random embedding layer is used to generate random embeddings for characters in the input biomedical text to obtain a character embedding representation sequence for each word in the biomedical text; The BIOBERT model is used to obtain word-level features of input biomedical text; The CNN module and the maximum pooling layer are cascaded to obtain local character features of each word in the biomedical text according to the character embedding representation sequence of each word in the biomedical text; The second BILSTM layer and the attention module are cascaded to obtain global character-level features of each word of the biomedical text according to the character embedding representation sequence of the biomedical text; The feature fusion layer is used to fuse word-level features, local character features and global character-level features of each word in the biomedical text to obtain a first comprehensive feature vector sequence of the biomedical text.

4. According to claim 3, a genome variation named entity recognition method based on feature fusion and multi-task learning is characterized in that: The feature fusion of the word-level features, local character features and global character-level features of each word in the biomedical text includes: Among them, h i Represents the first comprehensive feature vector of the i-th word; and They represent the global character-level features, local character features, and word-level features of the i-th word respectively; λ1, λ2, and λ3 represent the weight coefficients of each type of feature.

5. According to claim 1, a genome variation named entity recognition method based on feature fusion and multi-task learning is characterized in that: The training process of the named entity recognition model includes: The first comprehensive feature vector sequence is respectively input into the NER module, token classification module and entity boundary division module to perform named entity recognition task, token classification task and entity boundary division task on the input biomedical text; the corresponding cross entropy loss function is constructed based on the prediction result of each task and the corresponding true label, the total loss function of the model is constructed based on the loss function of each task, and the parameters of the named entity recognition model are tuned based on the constructed total loss function of the model.

6. According to claim 5, a method for genome variation named entity recognition based on feature fusion and multi-task learning is characterized in that: The total loss function Loss of the model includes: Loss=α·L NER +β·L Boundary +γ·L Token Among them, L NER represents the loss function of the named entity recognition task, L Boundary represents the loss function of the entity boundary segmentation task; L Token Represents the loss function for the token classification task.

7. According to claim 5, a method for genome variation named entity recognition based on feature fusion and multi-task learning is characterized in that: The cross entropy loss function includes: Among them, L∈{L NER ,L Boundary ,L Token },y i represents the true label distribution, represents the label distribution predicted by the model, k represents the number of different label categories in the dataset, ω i is the weight corresponding to label i.

8. A genomic variation named entity recognition system based on feature fusion and multi-task learning, wherein the system is applied to a genomic variation named entity recognition method based on feature fusion and multi-task learning according to any one of claims 1 to 7, comprising: Data enhancement module, multi-task model training module, data acquisition module and classification module; A data enhancement module is used to enhance the biomedical text data to generate diversified enhanced biomedical text data; A multi-task model training module for training named entity recognition models based on augmented biomedical text data; A data acquisition module, used for acquiring biomedical text to be recognized in real time; The classification module is used to perform named entity recognition on the biomedical text to be identified using the trained named entity recognition model.

9. A genome variation named entity recognition device based on feature fusion and multi-task learning, characterized in that: It comprises a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the genomic variation named entity recognition device based on feature fusion and multi-task learning performs the genomic variation named entity recognition method based on feature fusion and multi-task learning described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a program, characterized in that: When the program is executed by a processor, a genomic variation named entity recognition method based on feature fusion and multi-task learning as described in any one of claims 1 to 7 is implemented.

Images