CN111814460B - External knowledge-based drug interaction relation extraction method and system - Google Patents

Abstract

The invention provides a method and a system for extracting a drug interaction relation based on external knowledge, wherein the method comprises the following steps: analyzing and processing the contents of the drug database, extracting and generating drug pairs with interaction, and simultaneously storing all drug description information to form a drug interaction data set with the drug description information; constructing a drug description system information training model, and training through the drug interaction data set to obtain and store an optimal model; and combining the optimal model with a BilSTM-Att-CapsNet model to obtain an EK-BilSTM-Att-CapsNet model, identifying a drug entity of a drug interaction data set, searching and storing corresponding drug description information in a drug database, and finally training the combined model to obtain a final relation extraction model. The method can relieve the problem of large difference of extraction results of different relation categories, and improve the extraction effect.

Description

Method and system for extraction of drug interaction relationship based on external knowledge

technical field

The invention belongs to the technical field of natural language processing, and in particular relates to a method for extracting drug interaction relationships based on external knowledge.

Background technique

Drug-Drug Interactions (DDI) refers to the synergistic or antagonistic effects between different drugs when a patient takes multiple drugs at the same time, which may cause side effects, increase the cost of treatment, and increase the cost of treatment. The life safety of patients is threatened, so understanding the interaction between drugs is of great significance and value for the diagnosis and treatment of patients and the development of medicine.

At present, the main application methods in the field of drug interaction relationship extraction are: rule-based methods, traditional machine learning-based methods and deep learning-based methods. Rule-based methods generally require the assistance of professionals in the medical field to formulate rules. Due to the diversity of language expressions, it is often difficult to formulate rules to cover all drug interaction relationships, so the recall rate of this method is low; Traditional machine learning methods usually need to use a large number of artificially defined features, such as parts of speech, syntax, grammar and other features, and need to use external natural language processing tools to generate these features, such as part-of-speech tagger, syntax analyzer and other tools, so its extraction performance It is greatly influenced by external natural language processing tools; the method based on deep learning has the ability to automatically learn features, which can reduce the cost of artificial design features and the extraction effect is generally better than the traditional method, but similar to the first two methods, There will be large differences in the extraction results of the model on different relationship categories.

SUMMARY OF THE INVENTION

In view of this, one of the objectives of the present invention is to provide a method for extracting drug interaction relationships based on external knowledge, which can alleviate the problem of large differences in the extraction results of different relationship categories and improve the extraction effect.

In order to achieve the above object, the technical scheme of the present invention is: a method for extracting drug interaction relationship based on external knowledge, comprising the following steps:

Analyze and process the content of the drug database, extract and generate interacting drug pairs, save all drug description information at the same time, and form a drug interaction data set with drug description information;

constructing a drug description information training model, and training through the drug interaction data set to obtain and save the optimal model;

The optimal model is combined with the BiLSTM-Att-CapsNet model to obtain the EK-BiLSTM-Att-CapsNet model, and the drug entities of the drug interaction data set are identified at the same time, and the corresponding drug description information is found in the drug database and saved. The combined model is trained to obtain the final relation extraction model.

Further, the steps of constructing a drug description system information training model, and performing training on the drug interaction data set to obtain and save the optimal model, specifically include:

Receive the description information of the first drug and the description information of the second drug at the same time;

Convert the description information of the first drug and the description information of the second drug into a vector representation;

Obtain the forward information and backward information of the first drug and the second drug description sentences respectively, and then combine the two to represent them as sentences;

Linearly transform the sentence representation, and then perform normalization, and select the category with the highest probability as the predicted category label;

The optimal model is obtained by substituting the loss function into the drug interaction dataset for training.

Further, the method to obtain the sentence representation is:

BiLSTM obtains the forward and backward information of the sentence, and calculates the output of the hidden layer:

表示正向输入的语句，

表示逆序输入的语句，

表示正向输入的语句的输出，

表示逆序输入的语句的输出，H为BiLSTM隐藏层的输出；

A statement representing forward input,

Represents a statement entered in reverse order,

represents the output of the forward input statement,

Represents the output of the sentence input in reverse order, and H is the output of the BiLSTM hidden layer;

After calculating the BiLSTM, the sentence is expressed as:

表示前向输入的最后一个时间步的信息，

表示后向输入的最后一个时间步的信息。

information representing the last time step of the forward input,

Information representing the last time step of the backward input.

Further, the predicted category labels are obtained in the following ways:

First calculate the linear transformation of the sentence representation:

h ^* =[h ₁ ; h ₂ ];

output=W ^(fc) ·h ^* +b ^(fc) ;

Among them, W ^(fc) and b ^(fc) are the weight parameters and bias parameters of the fully connected layer, respectively, h ₁ ∈ R ^N represents the sentence representation of the first drug description information through the BiLSTM layer, h ₂ ∈ R ^N represents the second The drug description information is represented by sentences in the BiLSTM layer, N represents the number of BiLSTM hidden layer units, output is the output of linear transformation, h ^* ∈ R ^2N is the splicing of the sentence representations of the first drug and the second drug;

Normalization is performed according to linear transformation, and the largest category probability is selected as the predicted category label:

代表预测类别标签，output代表所述句子线性变换的输出，softmax(output)为归一化处理。in,

Represents the predicted category label, output represents the output of the linear transformation of the sentence, and softmax(output) is normalized.

Further, the loss function is:

以one-hot向量表示，λ是L2正则化的超参数，θ为在模型中进行训练得到。where y∈Rm represents the true class label, ^m represents the number of class labels, y and

Represented as a one-hot vector, λ is the hyperparameter of L2 regularization, and θ is obtained by training in the model.

Further, the loss function used by the EK-BiLSTM-Att-CapsNet model is:

L=T _k max(0,m ⁺ -||v _k ||) ² +λ(1-T _k )max(0,||v _k ||-m ^- ) ²

Among them, Tk is the indicator function of classification, _k is the indicator coefficient, m ⁺ is the upper boundary, ||v _k || is the length of the kth capsule, and ^m- is the lower boundary.

In view of this, the second purpose of the present invention is to provide a drug interaction relationship extraction system based on external knowledge, which can alleviate the problem of large differences in results in different relationship category extraction.

To achieve the above purpose, the technical solution of the present invention is: a system for extracting drug interaction relationships based on external knowledge, comprising:

The drug information dataset building module is used to analyze and process the content of the drug database, extract and generate interacting drug pairs, and retain the description information of all drugs to form a drug interaction dataset with drug description information;

a drug description information model, connected with the drug information data set building module, used to construct a drug description information training model, train on the drug interaction data set, and then save the optimal model;

The EK-BiLSTM-Att-CapsNet model, connected with the drug description information model, is used to identify the drug entity of the drug interaction data set, and then find the corresponding drug description information in the drug database and save it, while the drug The optimal model saved by the description information model is combined with the BiLSTM-Att-CapsNet model, and the combined model is trained to obtain the final relation extraction model.

Further, the drug description information model includes: an input layer, an embedding layer, a BiLSTM layer, a fully connected layer, and an output layer; wherein,

The input layer receives the description information of the first medicine and the description information of the second medicine at the same time, the first medicine description sentence is represented by p, p={p ₁ ,...p _i ...,p _n }; the second medicine The description sentence is represented by q, q={q ₁ ,...q _i ...,q _n }, p _i and q _i respectively represent the i-th word of the two drug description sentences;

表示，

第二药物描述语句向量用

表示，

表示单词嵌入的维度；The embedding layer converts the drug description sentence of the input layer into a vector representation, and the first drug description sentence vector is represented by

express,

The second drug description sentence vector is used

express,

represents the dimension of word embedding;

The BiLSTM layer is connected with the embedding layer, and is used for using the BiLSTM network to obtain the forward information and the backward information of the description sentences of the first drug and the second drug respectively, and then combine the two to represent the sentences;

The fully connected layer is connected to the BiLSTM layer and is used to linearly transform the sentence representation of the BiLSTM layer;

The output layer is connected to the fully-connected layer, and is used for normalizing the output of the fully-connected layer, and selecting the relationship category with the largest category probability as the predicted relationship category.

Further, the method to obtain the sentence representation is:

表示正向输入的语句，

表示逆序输入的语句，

表示正向输入的语句的输出，

表示逆序输入的语句的输出，H为BiLSTM层中BiLSTM隐藏层的输出，

表示前向输入的最后一个时间步的信息，

表示后向输入的最后一个时间步的信息。

A statement representing forward input,

Represents a statement entered in reverse order,

represents the output of the forward input statement,

Represents the output of the sentence input in reverse order, H is the output of the BiLSTM hidden layer in the BiLSTM layer,

information representing the last time step of the forward input,

Information representing the last time step of the backward input.

Further, the predicted relationship category is obtained in the following manner:

h ^* =[h ₁ ; h ₂ ];

output=W ^(fc) ·h ^* +b ^(fc) ;

代表预测的关系类别，output代表所述全连接层的输出，softmax(output)为归一化处理。Among them, W ^(fc) and b ^(fc) are the weight parameters and bias parameters of the fully connected layer, respectively, h ₁ ∈ R ^N represents the sentence representation of the first drug description information through the BiLSTM layer, h ₂ ∈ R ^N represents the second The drug description information is represented by sentences in the BiLSTM layer, N represents the number of BiLSTM hidden layer units, output is the output of linear transformation, h ^* ∈ R ^2N is the splicing of the sentence representations of the first drug and the second drug;

Represents the predicted relationship category, output represents the output of the fully connected layer, and softmax(output) is normalized.

The invention provides a drug interaction relationship extraction method based on external knowledge, which processes information in an external drug database, constructs a data set with drug description information, then conducts model training on the data set, and saves the optimal Finally, the optimal model is combined with the drug relationship extraction model to extract the drug relationship, so as to better utilize the existing knowledge in the drug database, alleviate the problem of large differences in the extraction results of different relationship categories, and improve the performance of the drug relationship. Extraction effect.

Description of drawings

In order to illustrate the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that are required to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention, and for those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative effort.

1 is a schematic structural diagram of an embodiment of an external knowledge-based drug interaction relationship extraction system of the present invention;

Fig. 2 is the structural representation of the EK-BiLSTM-Att-CapsNet model of the present invention;

FIG. 3 is a flowchart of an embodiment of a method for extracting drug interaction relationships based on external knowledge of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The examples are given to better illustrate the present invention, but the content of the present invention is not limited to the examples. Therefore, those skilled in the art make non-essential improvements and adjustments to the embodiments according to the above-mentioned contents of the invention, which still belong to the protection scope of the present invention.

Example 1

Referring to FIG. 1, it is a schematic structural diagram of a drug interaction relationship extraction system based on external knowledge of the present invention, specifically, a drug interaction relationship extraction system based on external knowledge, including:

The drug information data set building module 1 is used to analyze and process the content of the drug database, extract and generate interacting drug pairs, and at the same time retain the description information of all drugs to form a drug interaction data set with drug description information;

In this embodiment, a commonly used drug database, such as the DrugBank database, can be used, which contains relevant information such as drug type, drug structure, drug description information and other drugs that interact with the drug;

In this embodiment, the drug information data set building module 1 finds that the description information of the drug is a detailed introduction of the drug by analyzing the drug information in the database, and the description information generally includes the disease type, composition, Information such as action and effect, that is, the description information of the drug may imply the potential relationship between two drugs, which is helpful to judge whether there is an interaction relationship, so the system selects the description information of the drug.

The drug description information model 2, which is connected with the drug information data set building module, is used to construct a drug description information training model, and train it on the drug interaction data set, and then save the optimal model;

The drug description information model 2 in this embodiment further includes: an input layer 201, an embedding layer 202, a BiLSTM layer 203, a fully connected layer 204, and an output layer 205; wherein,

The input layer 201 receives the description information of two medicines at the same time. In this embodiment, it is set as the description information of the first medicine and the description information of the second medicine. The first medicine description sentence is represented by p, p={p ₁ ,. ..p _i ..., p _n }; the second drug description sentence is represented by q, q={q ₁ , ... q _i ..., q _n }, p _i and qi _represent two drugs respectively describe the ith word of the sentence;

表示，

第二药物描述语句向量用

表示，

表示单词嵌入的维度；The embedding layer 202 converts the drug description sentence of the input layer 201 into a vector representation, and the first drug description sentence vector is represented by

express,

The second drug description sentence vector is used

express,

represents the dimension of word embedding;

The BiLSTM layer 203 is connected to the embedding layer 202, and is used to obtain the long sequence dependency information of the input sentence. By using the BiLSTM network, the forward information and backward information of the description sentences of the first drug and the second drug are obtained respectively, and then the two are obtained. combined, expressed as a sentence;

In this embodiment, the method for obtaining the sentence representation in the BiLSTM layer 203 is:

表示正向输入的语句，

表示逆序输入的语句，

表示正向输入的语句的输出，

表示逆序输入的语句的输出，H为BiLSTM层中BiLSTM隐藏层的输出，

表示前向输入的最后一个时间步的信息，

表示后向输入的最后一个时间步的信息。

A statement representing forward input,

Represents a statement entered in reverse order,

represents the output of the forward input statement,

Represents the output of the sentence input in reverse order, H is the output of the BiLSTM hidden layer in the BiLSTM layer,

information representing the last time step of the forward input,

Information representing the last time step of the backward input.

The fully connected layer 204 is connected to the BiLSTM layer 203. Since the output dimension of the BiLSTM layer 203 is different from the number of node relationship categories of the output layer 205, the fully connected layer 204 is used to linearly transform the sentence representation of the BiLSTM layer;

The output layer 205 is connected to the fully connected layer 204, and is used to normalize the output of the fully connected layer 204, and select the category with the highest probability as the predicted relationship category, that is, convert the value corresponding to each category into the corresponding probability, all The sum of the class probability values is 1, and the one with the highest probability is selected as the predicted relation class

Further, the predicted class labels are obtained in the following ways:

h ^* =[h ₁ ; h ₂ ];

output=W ^(fc) ·h ^* +b ^(fc) ;

代表预测的关系类别，output代表全连接层的输出，softmax(output)为归一化处理；Among them, W ^(fc) and b ^(fc) are the weight parameters and bias parameters of the fully connected layer, respectively, h ₁ ∈ R ^N represents the sentence representation of the first drug description information through the BiLSTM layer, h ₂ ∈ R ^N represents the second The drug description information is represented by sentences in the BiLSTM layer, N represents the number of BiLSTM hidden layer units, output is the output of linear transformation, h ^* ∈ R ^2N is the splicing of the sentence representations of the first drug and the second drug;

Represents the predicted relationship category, output represents the output of the fully connected layer, and softmax(output) is normalized;

Still further, the drug interaction data set with drug description information constructed in the drug information data set building module 1 is repeatedly trained in the drug description information model 2, and finally the model that obtains the optimal result is saved.

EK-BiLSTM-Att-CapsNet model 3, connected with the drug description information model 2, is used to identify drug entities in the drug interaction dataset, and then find the corresponding drug description information in the drug database and save it, and save the drug description information model at the same time The optimal model is combined with the BiLSTM-Att-CapsNet model, and the combined model is trained to obtain the final relation extraction model.

In this embodiment, in order to make full use of the external drug description information and drug interaction knowledge to improve the performance of the system, the optimal model obtained by the drug description information model 2 and the attention-based drug relationship extraction model (BiLSTM-Att -CapsNet model) in combination, with reference to FIG. 2, it is a schematic structural diagram of an embodiment of the EK-BiLSTM-Att-CapsNet model in the present invention, the attention-based drug relationship extraction model structure refers to the current related technical documents, as in this implementation In the example, the BiLSTM-Att-CapsNet model includes a BILSTM layer, an attention layer, a capsule network layer, and an output layer. The BILSTM layer and output layer in this model are not the same structure as the BILSTM layer and output layer in the drug description information model 2. In order to distinguish, in this embodiment, the BiLSTM-Att-CapsNet model including the BILSTM layer and the output layer are named as the BAC-BILSTM layer and the BAC-output layer in FIG. 2 ;

In this embodiment, because it is necessary to know whether the prediction is accurate or not, the weighted fully connected layer and the output layer of the drug description information model repeatedly train the model until the optimal model is obtained. The relationship prediction output by the output layer is the data to be predicted. prediction results.

Further, in this embodiment, the description information of the drug is used as a part of the low-level capsule network in the capsule network layer, and then dynamically transmitted to the high-level capsule network in the capsule network layer, so as to better use the drug description information. , In addition, the drug description information model is obtained by training on the constructed drug interaction data set, which is obtained by processing the existing information in the drug database, so the existing external drug interaction knowledge can be fully utilized.

Preferably, the loss function formula adopted by the EK-BiLSTM-Att-CapsNet model in this embodiment is:

L=T _k max(0,m ⁺ -||v _k ||) ² +λ(1-T _k )max(0,||v _k ||-m ^- ) ²

Among them, T _k is the indicator function of classification, k is the indicator coefficient, m ⁺ is the upper boundary, ||v _k || is the length of the k-th capsule, m ^- is the lower boundary; when k exists, T _k is 1 , otherwise _Tk is 0.

In this embodiment, the drug description information model 2 is trained separately, the data set used is a large amount of drug interaction data existing in the drug database, and the model is trained according to the relationship prediction of the output layer of the model, and the optimal model is saved. The main purpose is to explore the connection between drug description information and drug interaction relationship by using a large amount of related knowledge in the drug database; then use the training set of public datasets (such as DDIE2011 and DDIE2013) to train the overall model, at this time Input the training set data of DDIE2011 and DDIE2013 into the optimal drug description information model saved just now, and get the intermediate results. There is no need to train the parameters of the optimal model, and then combine the intermediate results with the capsule network model. perform relation extraction.

Example 2

Based on the system of Embodiment 1, this embodiment provides a method for extracting drug interaction relationships based on external knowledge. Referring to FIG. 3, it is a schematic flowchart of the method, specifically: a drug interaction relationship extraction method based on external knowledge method, including the following steps:

S400: Analyse and process the content of the drug database, extract and generate interacting drug pairs, save all drug description information at the same time, and form a drug interaction data set with the drug description information; then perform step S402;

In this embodiment, a commonly used drug database, such as the DrugBank database, contains relevant information such as drug type, drug structure, drug description information and other drugs that interact with the drug; The knowledge is analyzed and processed to extract the interacting drug pairs and generate non-interacting drug pairs, while retaining the description information of each drug, so as to construct a drug interaction data set with drug description information.

S402: Build a drug description system information training model, and train it through the drug interaction data set to obtain and save the optimal model; then perform step S404;

In this embodiment, the description information of two medicines, such as the description information of the first medicine and the description information of the second medicine, is first received at the same time, and then the description information of the first medicine and the description information of the second medicine are converted into vector representations ; Obtain the forward information and backward information of the first drug and the second drug description sentence respectively, and then combine the two as a sentence representation; then perform a linear transformation on the sentence representation, and then perform a normalization process to select the category The highest probability is used as the predicted category label; the specific calculation method can refer to the establishment method of the drug description information model 2 in the embodiment;

Further, the description information p of the first medicine can be expressed as h ₁ ∈ R ^N after passing through the BiLSTM layer 203 , and the description information q of the second medicine can be expressed as h ₂ ∈ R ^N after passing through the BiLSTM layer 203 , and the two or splicing, that is, h ^* =[h ₁ ; h ₂ ]∈R ^2N is sent to the fully connected layer 204, where N represents the number of hidden layer units in the BiLSTM layer 203; since the output dimension of BiLSTM is 2N, the output layer The node is the number of relationship categories m, so the fully connected layer needs to be used for linear transformation, as shown in the following formula, where W ^(fc) and b ^(fc) are the weight parameters and bias parameters of the fully connected layer, respectively, and output is the output of the linear transformation:

output=W ^(fc) ·h ^* +b ^(fc) .

Preferably, in order to improve the generalization ability of the model and reduce the risk of overfitting of the model, the dropout mechanism is applied in the fully connected layer, and the idea is to use a certain ratio to make the nodes not work.

Finally, in this step, the loss function is substituted into the drug interaction data set for training to obtain the optimal model; specifically, the softmax function is used in this layer to normalize the output of the fully connected layer, that is, the values corresponding to each category are converted For the corresponding probability, the sum of the probability values of all categories is 1, and the one with the highest probability is selected as the predicted relationship category (predicted category label):

代表预测类别标签，output代表句子线性变换的输出，softmax(output)为归一化处理。in,

Represents the predicted category label, output represents the output of the linear transformation of the sentence, and softmax(output) is normalized.

Further, the loss function is:

以one-hot向量表示，λ是L2正则化的超参数，θ为在模型中进行训练得到，该θ主要是目的是为了防止模型出现过拟合现象，当θ值为最佳值时，此时的药物描述系信息训练模型即为最优模型。本步骤中还使用步骤S400中构建的带有药物描述信息的药物相互用数据集进行训练，并将取得最优结果的模型进行保存。where y∈Rm represents the true class label, ^m represents the number of class labels, y and

Represented by a one-hot vector, λ is the hyperparameter of L2 regularization, and θ is obtained by training in the model. The main purpose of this θ is to prevent the model from overfitting. When the θ value is the best value, this The information training model of the drug description system at the time is the optimal model. In this step, the drug interaction data set with drug description information constructed in step S400 is also used for training, and the model that obtains the optimal result is saved.

S404: combine the optimal model with the BiLSTM-Att-CapsNet model to obtain the EK-BiLSTM-Att-CapsNet model; then perform step S406;

The EK-BiLSTM-Att-CapsNet model obtained in this example can refer to the EK-BiLSTM-Att-CapsNet model 3 in Example 1, and the structure diagram refers to Figure 2. The drug information enters the BiLSTM after passing through the BILSTM in the drug description information model. -Att-CapsNet model arrives at the capsule network layer.

Further, the loss function used by the EK-BiLSTM-Att-CapsNet model is:

L=T _k max(0,m ⁺ -||v _k ||) ² +λ(1-T _k )max(0,||v _k ||-m ^- ) ²

Among them, Tk is the indicator function of classification, _k is the indicator coefficient, m ⁺ is the upper boundary, ||v _k || is the length of the kth capsule, and ^m- is the lower boundary.

S406: Identify the drug entities of the drug interaction data set, find and save the corresponding drug description information in the drug database, and finally train the combined model to obtain a final relationship extraction model.

In this embodiment, the drug entity is the description information of the drug. Assuming that the BiLSTM-Att-CapsNet model inputs the original sentence S _original , the shortest dependency path between the two drug entities in the sentence is S _sdp , and the two The description information of drug entities in the DrugBank database is represented by des ₁ and des ₂ , respectively;

和

再将二者分别送入BAC-BiLSTM层，即可得到原语句的长序列依赖信息H_original和最短依存路径的长序列依赖信息表示H_sdp，考虑到语句中不同单词的重要性不同，在原语句H_original上应用单词级别的注意力机制，即可得到加权后的

为了缓解使用最短依存路径可能造成的噪声干扰问题，在H_original和H_sdp应用句子级别的注意力机制，可以有效的将原语句信息和最短依存路径信息相融合，用H_all表示；As shown in the upper part of Fig. 2, the S _original and S _sdp represented by text are passed through the embedding layer to obtain a vector represented by

and

Then send the two to the BAC-BiLSTM layer, respectively, to obtain the long sequence dependency information H _original of the original sentence and the long sequence dependency information representation H _sdp of the shortest dependency path. Considering the different importance of different words in the sentence, in the original sentence Apply the word-level attention mechanism on H _original to get the weighted

In order to alleviate the noise interference problem that may be caused by using the shortest dependency path, the sentence-level attention mechanism is applied to H _original and H _sdp , which can effectively fuse the original sentence information and the shortest dependency path information, which is represented by _Hall ;

As shown in the lower part of Figure 2, after the first drug description information des ₁ and the second drug description information des ₂ are passed through the embedding layer, they are respectively sent to the BiLSTM layer 203, and the corresponding BiLSTM hidden layer output H _des1 can be obtained. and H _des2 , by combining the two, H _des =[H _des1 ; H _des2 ] can be obtained.

_{Convolve Hall} and H _des respectively to obtain the low-level capsule representation of the capsule network layer, u _all =(u ₁ ,u ₂ ,..., _um ) and u _des =(u ₁ ,u ₂ ,. ..,u _n ), where m and n represent the number of low-level capsules generated by _{Hall and the number of capsules generated by the drug description information H des ,} respectively. The two together form the low-level capsule u=[u _all ; u _des ] of the capsule network layer. Through the dynamic routing mechanism of the capsule network layer, the amount of information transmitted from the low-level capsule to the high-level capsule can be dynamically determined, and the drug description can be dynamically used. information as external knowledge.

Example 3

In this example, the effectiveness of the system in Example 1 and the method in Example 2 is verified, and the DDIExtraction2013 data set is used for experiments. The data set is composed of 792 medical texts in DrugBank and 233 abstracts in MedLine. composition, and annotated drug entities in advance. The details of the dataset are shown in Table 1:

Table 1 DDIExtraction2013 data description

In this chapter, the length of the drug description information is set to 50, and the word is converted into a 300-dimensional Glove word vector. For words that do not appear in the Glove vocabulary, the word vector of the word is represented by random initialization. The number of iterations of the capsule network Set to 3, Batchsize to 64, dropout to 0.5, learning rate to 0.001, the number of iterations of the drug description information model is set to 50, the number of iterations of the EK-BiLSTM-Att-CapsNet model is set to 35, and the model loss function where m ⁺ is 0.9 and ^m- is 0.1; the hyperparameter λ for L2 regularization is 0.25.

Further, the ablation experiment was designed using the model of Example 1, and the obtained results were shown in Table 2:

Table 2 Results of ablation experiments

Model P(%) R(%) F1(%) BiLSTM-Att-CapsNet 79.69 70.35 74.73 +Drug description information 79.47 71.01 75.00 +Drug description information+DDIs information in DrugBank 80.11 71.32 75.46 Model P(%) R(%) F1(%)

It can be seen from the results of the ablation experiment that after directly adding the external drug description information, the F1 value of the model is increased by 0.27%, indicating that there is information on the drug interaction relationship judgment in the drug description information. By using this information, the model can be improved. extraction effect. After training with the drug interaction dataset with drug description information in DrugBank, the effect of the model is improved by 0.73%, which effectively proves that the existing knowledge in the external drug database can effectively improve the extraction effect of the model. The results of ablation experiments show that the model proposed in this chapter can effectively use the existing external drug description information, and at the same time can use a large number of drug interaction knowledge in DrugBank, which is helpful to improve the extraction results of the model.

Further, the results of different relationship categories are compared to verify the effectiveness of the present invention. Through the analysis of the data set, it can be seen that the number of different categories in the data set is quite different, so it may cause the prediction results of different categories to be quite different. The experimental results of the invention model and related models in different categories are compared, and the specific information is shown in Table 3, (the models in the following table are all prior art models):

Table 3 Experimental results on different relation categories

From the above table, it can be seen that the maximum difference in F1 value between different categories of the comparison model is between 31.35% and 36.15%, and the maximum difference in F1 value between different categories of the system model of the present invention is only 25.34%, which is greatly reduced It can be seen that compared with other models, this model can effectively alleviate the problem of large differences in extraction results caused by large differences in the number of samples of different types.

Finally, in this embodiment, a comparative experiment is made between the method of the present invention and the existing relationship extraction method, and the experimental results of the system model of the present invention on the DDIExtraction2013 data set are compared with the results of the existing model on the data set. Compared.

The experimental results of different models on the DDIExtraction2013 dataset are shown in Table 4:

Table-4 Experimental results of different models on the DDIExtraction2013 dataset

Model/method proposer P(%) R(%) F1(%) CNN/Liu et al. (2016) 75.72 64.66 69.75 BR-LSTM/Xuetal.(2018) 71.52 70.79 71.15 PM-BLSTM/Zhouetal. (2018) 75.8 70.38 72.99 System of the present invention 80.11 71.32 75.46

According to Table 4, the F1 value of the model in the present invention is higher than the F1 value of other models in the table, that is, compared with the existing model, the system model of the present invention can better automatically extract the interaction between drugs effect relationship.

The embodiments of the present invention have been described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments, which are merely illustrative rather than restrictive. Under the inspiration of the present invention, without departing from the scope of protection of the present invention and the claims, many forms can be made, which all belong to the protection of the present invention.

Claims (10)

1. A drug interaction relation extraction method based on external knowledge is characterized by comprising the following steps:

analyzing and processing the contents of the drug database, extracting and generating drug pairs with interaction, and simultaneously storing all drug description information to form a drug interaction data set with the drug description information;

constructing a drug description system information training model, and training through the drug interaction data set to obtain and store an optimal model;

and combining the optimal model with a BilSTM-Att-CapsNet model to obtain an EK-BilSTM-Att-CapsNet model, identifying a drug entity of a drug interaction data set, searching and storing corresponding drug description information in a drug database, and finally training the combined model to obtain a final relation extraction model.

2. The method according to claim 1, wherein the step of constructing a drug description system information training model and training through the drug interaction data set to obtain and store an optimal model specifically comprises:

receiving description information of a first medicine and description information of a second medicine at the same time;

converting the description information of the first medicine and the description information of the second medicine into vector representation;

respectively acquiring forward information and backward information of the first medicine and the second medicine description sentences, and then combining the forward information and the backward information to represent as sentences;

performing linear transformation on the sentence expression, then performing linear normalization processing, and selecting the maximum class probability as a prediction class label;

and substituting the loss function into the drug interaction data set for training to obtain an optimal model.

3. The method of claim 2, wherein the sentence representation is derived by:

the BilSTM obtains the forward information and the backward information of the statement, and calculates to obtain the output of a hidden layer:

a statement that represents a forward input is presented,

a statement that is input in a reverse order is represented,

the output of the sentence representing the forward input,

representing the output of the statement input in the reverse order, wherein H is the output of a BilSTM hidden layer;

after the BilSTM is calculated, the sentence is expressed as:

information indicating the last time step of the forward input,

information indicating the last time step of the backward input.

4. The method of claim 3, wherein the prediction class label is derived by:

firstly, linear transformation expressed by sentences is calculated:

h^*＝[h₁；h₂]；

output＝W^(fc)·h^*+b^(fc)；

wherein, W^(fc)And b^(fc)Weight parameter and bias parameter, h, of the fully-connected layer, respectively₁∈R^NSentence representation, h, representing the first drug description information through the BilSTM layer₂∈R^NRepresenting the sentence representation of the second medicine description information through a BilSTM layer, N representing the number of units of a BilSTM hidden layer, output being the output of linear transformation, h^*∈R^2NSplicing sentence representations of the first medicine and the second medicine;

carrying out normalization processing according to linear transformation, and selecting the maximum class probability as a prediction class label:

wherein,

represents a prediction class label, output represents the output of the sentence linear transformation, and softmax (output) is the normalization process.

5. The method of claim 4, wherein the loss function is:

wherein y ∈ R^mRepresenting true class labels, m representing the number of class labels, y and

and in one-hot vector representation, lambda is a hyper-parameter of L2 regularization, and theta is obtained by training in a model.

6. The method of claim 5, wherein the EK-BilSTM-Att-CapsNet model uses a loss function of:

L＝T_kmax(0,m⁺-||v_k||)²+λ(1-T_k)max(0,||v_k||-m^-)²

wherein, T_kFor the indicative function of the classification, k is an indicative coefficient, m⁺To the upper boundary, | | v_kI is the length of the kth capsule, m^-Is the lower boundary.

7. An external knowledge-based drug interaction relationship extraction system, comprising:

the drug information data set construction module is used for analyzing and processing the content of the drug database, extracting and generating drug pairs with interaction, and simultaneously reserving the description information of all drugs to form a drug interaction data set with drug description information;

the medicine description information model is connected with the medicine information data set construction module and used for constructing a medicine description information training model, training on the medicine interaction data set and storing an optimal model;

and the EK-BilSTM-Att-CapsNet model is connected with the drug description information model and is used for identifying the drug entity of the drug interaction data set, searching and storing corresponding drug description information in a drug database, simultaneously combining the optimal model stored by the drug description information model with the BilSTM-Att-CapsNet model, and training the junction model to obtain a final relation extraction model.

8. The system of claim 7, wherein the medication description information model comprises: an input layer, an embedding layer, a BilSTM layer, a full connection layer and an output layer; wherein,

the input layer receives description information of a first drug and description information of a second drug at the same time, the first drug description sentence is represented by p, and p ═ p₁,...p_i...,p_n}; the second drug descriptive statement is denoted by q, q ═ q₁,...q_i...,q_n}，p_iAnd q is_iThe ith word representing two drug description sentences, respectively;

the embedded layer converts the medicine description statement of the input layer into vector representation, and the first medicine description statement vector is used

It is shown that,

second drug description statement vector

It is shown that,

a dimension representing word embedding;

the BilSTM layer is connected with the embedded layer and used for respectively acquiring the forward information and the backward information of the description sentences of the first medicine and the second medicine by using a BilSTM network, and then combining the forward information and the backward information to be used as sentence representation;

the full connection layer is connected with the BilSTM layer and used for carrying out linear transformation on sentence representation of the BilSTM layer;

and the output layer is connected with the full connection layer and used for carrying out normalization processing on the output of the full connection layer and selecting the maximum class probability as a predicted relation class.

9. The system of claim 8, wherein the sentence representation is derived by:

a statement that represents a forward input is presented,

a statement that is input in a reverse order is represented,

the output of the sentence representing the forward input,

the output of the statement representing the input in the reverse order, H is the output of the BilSTM hidden layer in the BilSTM layer，

Information indicating the last time step of the forward input,

information indicating the last time step of the backward input.

10. The system of claim 9, wherein the predicted relationship category is derived by:

h^*＝[h₁；h₂]；

output＝W^(fc)·h^*+b^(fc)；

wherein, W^(fc)And b^(fc)Weight parameter and bias parameter, h, of the fully-connected layer, respectively₁∈R^NSentence representation, h, representing the first drug description information through the BilSTM layer₂∈R^NRepresenting the sentence representation of the second medicine description information through a BilSTM layer, N representing the number of units of a BilSTM hidden layer, output being the output of linear transformation, h^*∈R^2NSplicing sentence representations of the first medicine and the second medicine;

represents the predicted relationship class, output represents the output of the fully-connected layer, and softmax (output) is the normalization process.

Images