CN110020671A - The building of drug relationship disaggregated model and classification method based on binary channels CNN-LSTM network - Google Patents
The building of drug relationship disaggregated model and classification method based on binary channels CNN-LSTM network Download PDFInfo
- Publication number
- CN110020671A CN110020671A CN201910174269.4A CN201910174269A CN110020671A CN 110020671 A CN110020671 A CN 110020671A CN 201910174269 A CN201910174269 A CN 201910174269A CN 110020671 A CN110020671 A CN 110020671A
- Authority
- CN
- China
- Prior art keywords
- text
- drug
- relationship
- layer
- disaggregated model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Treatment And Welfare Office Work (AREA)
Abstract
The invention discloses a kind of drug relationship disaggregated model construction methods based on binary channels CNN-LSTM network, parent drugs text set is pre-processed, backout is carried out to pretreated drug text each in pretreated drug text set, obtains backward text set；Using pretreated drug text set as positive sequence text set；Training neural network, obtains drug relationship disaggregated model；Neural network includes parallel positive sequence Text character extraction layer and backward Text character extraction layer, Fusion Features layer and classification layer；Positive sequence Text character extraction layer and backward Text character extraction layer include the convolution block set gradually and shot and long term Memory Neural Networks block；The present invention is extracted the local feature of drug text using CNN, is extracted the global characteristics of drug text respectively using LSTM by constructing binary channels CNN-LSTM network, and the drug relationship feature extracted is more abundant, so that classification accuracy rate improves.
Description
Technical field
The present invention relates to the building of drug relationship disaggregated model and classification methods, and in particular to one kind is based on binary channels CNN-
The drug relationship disaggregated model of LSTM network constructs and classification method.
Background technique
Drug relationship refers to while or taking comprehensive effect caused by two or more drugs whithin a period of time.This
Kind effect can be divided into synergistic effect, antagonistic effect and non-interaction.Mutual antagonistic effect between drug can cause patient
Serious health risk.Drug relationship extracts the typical relationship that (DDIE) task is natural language processing field and extracts task,
It is intended to detect and identify the semantic relation of drug pair, to drug safety accident is reduced, the development of biomedical technology is promoted to have
Significance.
In recent years, the expert in terms of biomedical and text mining is made that very big effort in DDIE task, also creates
Many methods are made, these methods can be mainly divided into three classes: the method for rule-based mode, the side based on statistical machine learning
Method and method based on deep learning.Although the method for rule-based mode can targetedly be gone in identification target text very much
Entity relationship, but this method is there are three very serious drawback: (1) needing to expend a large amount of man power and material and remove research mesh
Text is marked, the information extraction quality for otherwise making rule cannot be guaranteed；(2) it is needed when laying down a regulation in the field
Expert a large amount of priori knowledge is provided, and may to make standard because of the reason of subjective consciousness different by different experts
The regular collection of cause；(3) because this method has very strong specific aim to domain knowledge, it is only applicable in the field
Information extraction, generalization ability is generally poor, so the method for rule-based mode does not cause researcher widely to pay close attention to.
Although these method performances based on statistical machine learning are good, it require that coming by fine and cumbersome Feature Engineering
Extract suitable characteristic set.However, the extraction quality of these features depend on existing natural language processing tool, therefore by
To the adverse effect of these tool noises and cost, extracts and be characterized in unordered, characteristic mass hardly results in effective guarantor
Card causes the accuracy rate of classification not high.
Summary of the invention
The drug relationship disaggregated model building based on binary channels CNN-LSTM network that the purpose of the present invention is to provide a kind of
And classification method leads to drug to solve feature randomness that drug relationship classification method in the prior art extracts
The not high problem of the accuracy rate of relationship classification.
In order to realize above-mentioned task, the invention adopts the following technical scheme:
A kind of drug relationship disaggregated model construction method based on binary channels CNN-LSTM network, the method according to
Lower step executes:
Step 1 obtains parent drugs text set；
Drug relationship in parent drugs text each in parent drugs text set is labeled, drug relationship mark is obtained
Label collection；
Step 2 pre-processes the parent drugs text set, obtains pretreated drug text set；
The pretreatment includes Text normalization, text size is fixed and text vector mapping；
Step 3 carries out backward behaviour to pretreated drug text each in the pretreated drug text set
Make, obtains backward text set；
Using the pretreated drug text set as positive sequence text set；
Step 4, using the positive sequence text set and backward text set as input, by the drug relationship tally set
As output, training neural network obtains drug relationship disaggregated model；
The neural network includes the parallel positive sequence Text character extraction layer and backward text feature set gradually
Extract layer, Fusion Features layer and classification layer；
The positive sequence Text character extraction layer and the backward Text character extraction layer include the volume set gradually
Block and shot and long term Memory Neural Networks block.
Further, the convolution block is provided with 4.
Further, each convolution block includes the batch regularization sublayer set gradually, convolution sublayer, activation letter
Number sublayer, convolution sublayer, activation primitive sublayer and pond beggar layer.
Further, the activation primitive in the activation primitive sublayer is ReLU function.
Further, the Fusion Features layer includes full articulamentum.
Further, the classification layer includes Softmax function layer.
A kind of drug relationship classification method based on binary channels CNN-LSTM network, to drug text to be sorted according to
Lower step executes:
Step A, the drug text to be sorted is pre-processed using the method for step 2 in claim 1, is obtained
Pretreated drug text；
It step B, will be described in the pretreated drug text input to any one of claim 1-6 claim
Drug relationship disaggregated model in, obtain classification results.
The present invention has following technical characterstic compared with prior art:
1, a kind of building of drug relationship disaggregated model and classification based on binary channels CNN-LSTM network provided by the invention
Method is extracted the local feature of drug text using CNN, is distinguished using LSTM by constructing binary channels CNN-LSTM network
The global characteristics of drug text are extracted, the drug relationship feature extracted is more abundant, so that classification accuracy rate improves；
2, a kind of building of drug relationship disaggregated model and classification based on binary channels CNN-LSTM network provided by the invention
Method completes feature extraction by the way that the positive sequence text of drug relationship text and backward text are respectively fed to CNN-LSTM network
Journey, compared to single pass LSTM network, the drug relationship feature extracted is more comprehensive, so that classification accuracy rate improves；
3, a kind of building of drug relationship disaggregated model and classification based on binary channels CNN-LSTM network provided by the invention
Method simplifies the process of drug characteristic vector extraction, improves drug relationship classification by extracting drug Text eigenvector
Accuracy；
4, a kind of building of drug relationship disaggregated model and classification based on binary channels CNN-LSTM network provided by the invention
Method is not necessarily to manual intervention and pertinent arts using the parent drugs relational text comprising multiple pharmaceutical entities as input,
It does not need manually to extract complicated text feature, generalization ability is strong.
Detailed description of the invention
Fig. 1 is the classification of drug model structure provided in one embodiment of the present of invention；
Fig. 2 is the convolution block internal structure chart provided in one embodiment of the present of invention.
Specific embodiment
It makes explanations first to the term occurred in specific embodiment:
Shot and long term Memory Neural Networks (LSTM): LSTM network is by input gate, forgetting door, out gate and memory unit structure
At LSTM passes through the door control mechanism of this complexity, can effectively learn the long-term Dependency Specification of input data, in text data
It has a wide range of applications in processing with the serialization informations such as track data.
Convolutional neural networks (CNN): a kind of comprising convolutional calculation and with the feedforward neural network of depth structure.
Embodiment one
As shown in Figure 1, disclosing a kind of drug relationship classification based on binary channels CNN-LSTM network in the present embodiment
Model building method, the method execute according to the following steps:
Step 1 obtains parent drugs text set；
Drug relationship in parent drugs text each in parent drugs text set is labeled, drug relationship mark is obtained
Label collection；
The biomedical text acquired in the present embodiment can be adopted by modes such as Biomedical literature and papers
Collection, the text of acquisition can be for document and paper partly or wholly, but needs to guarantee that text semantic expression is complete.
It is at least needed in the parent drugs text comprising two drug target title words, the two drug target title words are
It is related to the drug word of drug relationship classification, remaining is other words, such as parent drugs text in the present embodiment are as follows:
“Some quinolones,including ciprofloxacin,have been associated with transient
elevations in serum creatinine in patients receiving cyclosporine
Concomitantly ", wherein " quinolones ", " ciprofloxacin " and " cyclosporine " is medicine name word,
Remaining word is other words.
In the data set used herein, between 0 to 150 words, most text size is distributed in the length of text
Between 20 to 60 words, and backward rely on the phenomenon that (such as the grammatical phenomenons such as attributive clause) 46% is accounted in data set.
Drug relationship label includes 5 kinds, is that advice suggests respectively, effect effect, mechanism drug action mechanism, int
Positive and irrelevant false.
Step 2 pre-processes the parent drugs text set, obtains pretreated drug text set；
The pretreatment includes Text normalization, text size is fixed and text vector mapping；
In the present embodiment, patent is utilized " based on multilayer convolutional Neural to the pretreatment mode of parent drugs text set
The drug relationship classification method of network " in drug text set and processing mode.In each of parent drugs text set
The different different length of parent drugs text formatting, and medicine name word is complicated and uncommon, when being classified using neural network,
It is readily incorporated error, therefore is pre-processed firstly the need of to the parent drugs text of acquisition, including parent drugs are literary
All words in this carry out morphology normalization, i.e., the morphology of all words is unified；By drug target title word using unified
Naming method be named and be replaced original drug target title word in the form after naming, concrete operations with
Following steps:
All words in the parent drugs text set are normalized step 2.1, are named with Unified Form, then benefit
The drug target title word is replaced by the drug target title word after being named with Unified Form, after being normalized
Drug text set；
Wherein, normalization includes that morphology normalizes and name normalization；
To make drug text in classification, it can more accurately classify, reduce error and introduce, therefore by parent drugs
Each of text word carries out morphology normalization, converts them to unified format.To every in parent drugs text
One word carries out morphology normalization, the parent drugs text after being normalized, until each in parent drugs text set
Each of a parent drugs text word all have passed through morphology normalization, the parent drugs text set after being normalized.
In order to improve the generalization of neural network, by all drug target title prefixes in drug text first with unified shape
Formula name, Unified Form are the form of " X serial number ", and wherein X can be any English word, such as " day ",
" interaction " etc., serial number is with the sequence serial number of English form, such as " one, two, three " etc., and by the system
Title after one name replaces the title of former drug target word, the drug target title word after replacing be " drugone ",
" drugtwo " and " drugthree " etc. obtain pretreated drug text set there is no influencing between drug text.
Step 2.2, by each drug text uniform length in the drug text set after normalization, obtain length it is fixed after
Drug text set；
The length of each of pretreated drug text set drug text is fixed as n, for length
Text less than n is filled, and the mode of filling can be for by the way of full odd jobs random number, then drug text can indicate
Are as follows:
S=w1w2w3...wn
Step 2.3, it is fixed to each length in drug text set after fixed of the length after drug text into
Row vector mapping, obtains pretreated drug text set；
Since neural network can not be handled the text of natural language form directly as input, by drug text
Originally it is mapped as the text vector of digital form, comprising the following steps:
A, term vector table is constructed, the term vector table is made of word and corresponding digital term vector；
Term vector table is made of word and corresponding digital term vector, each word is corresponding unique in term vector table
The term vector of digital form as far as possible inserts more words in the table, and term vector table is enable to cover more word.
For more significant term vectors can be converted out, in the present embodiment, studied using Stanford university NLP small
Group provides GloVe (Global Vectors for Word Representation) model term vector table, including
2196016 term vectors, the dimension of each term vector are 300.If inputting the word in urtext not in this term vector table,
Then the often one-dimensional of the term vector of the word is initialized to 0.
B, the fixed drug text of each length in drug text set is mapped by way of tabling look-up, is obtained pre-
Treated drug text set.
For each of n dimension drug text word all by way of looking into the term vector table, it is mapped to one
Each of n dimension drug text word is all mapped to the term vector of d dimension, therefore by the vector of a d dimension in this manner
One length is that the parent drugs text S of n is just mapped as the text vector of one (n × d) dimension:
It and include that the parent drugs text S that m length is n is just mapped as a m × (n × d) text for one
Vector set includes m (n × d) text vectors tieed up in drug text set.
Step 3 carries out backward behaviour to pretreated drug text each in the pretreated drug text set
Make, obtains backward text set；
Using the pretreated drug text set as positive sequence text set；
There are also the language phenomenons such as the text, such as attribute postposition of backward in natural language, in the present solution, to make extraction
Feature is more comprehensive, using positive sequence drug text and backward drug text respectively to neural metwork training, obtains disaggregated model.
When carrying out backout to drug text, by the reversed order in text vector, such as the vector of one 1 dimension
[0.21 0.35 0.62 0.85 0.96], after backward are as follows: [0.96 0.85 0.62 0.35 0.21].
Step 4, using the positive sequence text set and backward text set as input, by the drug relationship tally set
As output, training neural network obtains drug relationship disaggregated model；
The neural network includes parallel positive sequence Text character extraction layer and backward Text character extraction layer, feature
Fused layer and classification layer；
The positive sequence Text character extraction layer is identical as the backward Text character extraction layer structure, includes successively
The convolution block and shot and long term Memory Neural Networks block of setting.
In the present embodiment, in order to improve the accuracy rate that drug relationship is classified, the structure of neural network has been carried out again
Design then these local features are sent to as shown in Figure 1, extracting the local feature of text using convolution block first
LSTM model extracts the global characteristics and temporal aspect of text to supplement, but it is the text that can handle positive sequence that this, which is also, if
The text modified backward of that attributive clause etc is encountered, processing capacity is still very weak, so using two identical features
Extract layer handles the positive sequence and backward of input text respectively, and then the positive sequence and backward feature extracted is combined, and obtains most
Whole text feature；Text feature is exported into classification layer again later and is classified.
In the present embodiment, the number of convolution block is not accurate enough lower than 4 local features extracted, the number of convolution block
Higher than 4, it may appear that the phenomenon that over-fitting, cause feature extraction to fail, therefore as a preferred embodiment, convolution block
It is provided with 4.
Optionally, each convolution block includes batch regularization sublayer, the convolution sublayer, activation primitive set gradually
Sublayer, convolution sublayer, activation primitive sublayer and pond beggar layer.
In the present embodiment, as shown in Fig. 2, positive sequence drug text and backward drug text can be sent to convolution block it
After be introduced into batch regularization layer, the effect of batch regularization layer is that input data is made more to meet normal distribution, meets normal state
The speed of the sample training of distribution can greatly improve, and accuracy rate can also improve.
In the present embodiment, the data after regularization are sent into convolutional layer and carry out convolution operation, the parameter setting of convolutional layer
Are as follows: the number of convolution unit filter 128.
Enter activation primitive later, data meaningless after convolution are deleted, as a preferred embodiment, swashing
Function living is Relu function.
Obtained data are sent into pond layer, pond layer uses maximum by the operation for repeating the above convolution sum activation
Pondization operation, such as, a pond window size is 2*2, by the pond window of this 2*2 after convolution sum activation
It is slided in data, number maximum in window is elected as representative during sliding, it is a with regard to how many to slide how many times
It represents, these is then used to represent the representative as initial data.Such do is advantageous in that: not losing guaranteeing that text spy demonstrate,proves
Under the premise of mistake, data are reduced, accelerate the training of network.
Positive sequence drug text and backward drug text are passing through 4 convolution blocks identical in this way and then are entering length
Global characteristics present in drug relationship text and temporal aspect are obtained in phase Memory Neural Networks block.
In the present embodiment, shot and long term Memory Neural Networks block interior joint number is set as 64.
Optionally, the Fusion Features layer includes full articulamentum.
In the present embodiment, positive sequence drug text and backward drug text are respectively fed to above-mentioned CNN-LSTM net
After network, positive sequence text feature and backward text feature are respectively obtained, the two features are sent to full articulamentum simultaneously.Such as
Say that positive text feature and backward text feature there are 100, then constructing a first layer is 200 nodes, the second layer 100
The full articulamentum of a node, positive text feature is sent into 100 nodes before first layer, after backward text feature is sent into first layer
Then this 200 features are fused together by 100 nodes in this way.
Optionally, the classification layer includes Softmax function layer.
In the present embodiment, full articulamentum and Softmax function layer constitute last portion of drug relationship sorting algorithm
Point, for exporting the drug relationship label of digital vectors form according to the quantity of classification, so that it is determined that last drug relationship point
Each output node of the final result of class, full articulamentum and Softmax function layer represents a drug categories, and classifier is most
The drug label exported eventually is given pharmaceutical entities to the probability for belonging to each drug categories, and the probability value is in [0,1].Example
Such as, it is now assumed that drug relationship there are 2 kinds, relationship and not related, the then output node setting of Softmax function layer have been respectively represented
It is 2, i.e., there are two types of drug relationships, positive and negative are respectively represented, if the number of Softmax function layer output
The drug relationship label of vector form is p [positive, negative]=[0.1,0.9], i.e. Softmax function layer exports
As a result it is 0.1 there are the probability value of positive in, is 0.9 there are the probability value of negative, is then judged with this.?
In the present embodiment, drug relationship includes 5 kinds, is that advice suggests respectively, effect effect, mechanism drug action mechanism, int
Positive and irrelevant false.
The stratification convolution loop neural network is trained with output using above-mentioned input, drug is obtained and closes
Be taxonomical hierarchy convolution loop neural network, the drug relationship text and each drug relationship label be number to
Amount form；Stratification convolution loop neural network described in repetition training n times, with the best stratification volume of this n times training performance
Product Recognition with Recurrent Neural Network as the drug relationship taxonomical hierarchy convolution loop neural network, wherein N >=1.
The training set of one taxonomical hierarchy convolution loop neural network comprising two parts, first is that being inputted after pretreatment
The drug text set of taxonomical hierarchy convolution loop neural network, second is that each drug text in pretreated drug text set
In this corresponding parent drugs text, the drug relationship label between drug target title word is obtained to every in drug text set
The corresponding drug relationship tally set of one drug text is exported as the target of multilayer convolutional network.Likewise, taxonomical hierarchy
The test set of convolution loop neural network also includes two parts, the difference is that during the test, only by pretreated medicine
Object text set is input in trained taxonomical hierarchy convolution loop neural network, taxonomical hierarchy convolution loop nerve
Network can obtain the classification of drug result set of model prediction according to the drug text data of input and trained model parameter, so
The true tag of classification of drug result set and drug relationship is compared afterwards, with the two comparison result classification of assessment stratification volume
The performance of product Recognition with Recurrent Neural Network.
In this example, using 2013 drug relationship data set of DDIExtraction as drug relationship text to classification
Stratification convolution loop neural network is trained and tests, by the 80% of entire data set as training set, 20% as survey
Examination collection, i.e., training set is made of 27792 drug relationship text samples, and test set is by 6409 drug relationship text sample groups
At.Then 10 training are carried out to stratification convolution loop neural network using ready-portioned training set, chosen in 10 training
Final mask of the best model of modelling effect as drug relationship stratification convolution loop neural network.
Embodiment two
A kind of drug relationship classification method based on binary channels CNN-LSTM network, to drug text to be sorted according to
Lower step executes:
Step A, the drug text to be sorted is pre-processed using the method for step 2 in embodiment 1, is obtained pre-
Treated drug text；
Step B, by drug relationship disaggregated model described in the pretreated drug text input embodiment 1
In, obtain classification results.
After training final drug relationship stratification convolution loop neural network, model can predict any drug
Drug relationship involved in relational text, by the drug text input drug relationship stratification convolution loop that drug relationship is unknown
Neural network, the drug that maximum probability is chosen from the digital vectors that drug relationship stratification convolution loop neural network exports close
It is the drug relationship classification results of the drug text as unknown drug relationship.
In the present embodiment, drug text to be sorted is " Some quinolones have been associated
with transient elevations in serum creatinine in patients receiving
Cyclosporine concomitantly ", first aim medicine name word are quinolones, second target medication name
Word is referred to as cyclosporine, carries out drug relationship point by trained drug relationship stratification convolution loop neural network
Class, the drug relationship digital vectors label of output are as follows:
P [mechanism, advice, effect, int, false]=[0.02,0.09,0.1,0.67,0.12]
I.e. between two drug targets quinolones, cyclosporine is 2% there are the probability of mechanism,
I.e. between two drug targets quinolones, cyclosporine is 9% there are the probability of advice, i.e. two target medicines
Between object quinolones, cyclosporine is 10% there are the probability of effect, i.e. two drug targets
Between quinolones, cyclosporine is 67% there are the probability of int, i.e. two drug target quinolones,
Between cyclosporine is 12% there are the probability of false, wherein being up to 67% there are the probability of int relationship, therefore
It will be between two drug targets quinolones, cyclosporine using drug relationship stratification convolution loop neural network
Relationship is classified as int positive relationship.
The drug relationship classification method and drug in the prior art based on binary channels CNN-LSTM network that this programme provides
Sorting algorithm is compared, performance comparison sheet 1, when evaluating a drug relationship classification method performance quality, accuracy rate, recall rate
It is bigger with F value, illustrate that drug relationship disaggregated model performance is better, from table 1 it follows that drug relationship layer proposed by the present invention
Secondaryization convolution loop neural network will significantly be better than other methods in three accuracy rate, recall rate and F value indexs, this
The drug relationship classification method proposed by the present invention based on the two-way convolution loop neural network of stratification is demonstrated in drug relationship
Possess optimal classification performance in classification problem.
The drug relationship classification method provided by the invention of table 1 and the deemed-to-satisfy4 energy of other drugs relationship classification compare
Claims (7)
1. a kind of drug relationship disaggregated model construction method based on binary channels CNN-LSTM network, which is characterized in that described
Method executes according to the following steps:
Step 1 obtains parent drugs text set；
Drug relationship in parent drugs text each in parent drugs text set is labeled, drug relationship label is obtained
Collection；
Step 2 pre-processes the parent drugs text set, obtains pretreated drug text set；
The pretreatment includes Text normalization, text size is fixed and text vector mapping；
Step 3 carries out backout to pretreated drug text each in the pretreated drug text set, obtains
Obtain backward text set；
Using the pretreated drug text set as positive sequence text set；
Step 4, using the positive sequence text set and backward text set as input, using the drug relationship tally set as
Output, training neural network, obtains drug relationship disaggregated model；
The neural network includes the parallel positive sequence Text character extraction layer and backward Text character extraction set gradually
Layer, Fusion Features layer and classification layer；
The positive sequence Text character extraction layer and the backward Text character extraction layer include the convolution block set gradually
And shot and long term Memory Neural Networks block.
2. special as described in claim 1 based on the drug relationship disaggregated model construction method of binary channels CNN-LSTM network
Sign is that the convolution block is provided with 4.
3. special as claimed in claim 2 based on the drug relationship disaggregated model construction method of binary channels CNN-LSTM network
Sign is that each convolution block includes the batch regularization sublayer set gradually, convolution sublayer, activation primitive sublayer, volume
Product sublayer, activation primitive sublayer and pond beggar layer.
4. special as claimed in claim 3 based on the drug relationship disaggregated model construction method of binary channels CNN-LSTM network
Sign is that the activation primitive in the activation primitive sublayer is ReLU function.
5. special as described in claim 1 based on the drug relationship disaggregated model construction method of binary channels CNN-LSTM network
Sign is that the Fusion Features layer includes full articulamentum.
6. special as described in claim 1 based on the drug relationship disaggregated model construction method of binary channels CNN-LSTM network
Sign is that the classification layer includes Softmax function layer.
7. a kind of drug relationship classification method based on binary channels CNN-LSTM network, which is characterized in that drug to be sorted
Text executes according to the following steps:
Step A, the drug text to be sorted is pre-processed using the method for step 2 in claim 1, obtains pre- place
Drug text after reason；
Step B, by medicine described in the pretreated drug text input to any one of claim 1-6 claim
In object relationship disaggregated model, classification results are obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910174269.4A CN110020671B (en) | 2019-03-08 | 2019-03-08 | Drug relationship classification model construction and classification method based on dual-channel CNN-LSTM network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910174269.4A CN110020671B (en) | 2019-03-08 | 2019-03-08 | Drug relationship classification model construction and classification method based on dual-channel CNN-LSTM network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110020671A true CN110020671A (en) | 2019-07-16 |
| CN110020671B CN110020671B (en) | 2023-04-18 |
Family
ID=67189365
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910174269.4A Active CN110020671B (en) | 2019-03-08 | 2019-03-08 | Drug relationship classification model construction and classification method based on dual-channel CNN-LSTM network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110020671B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111243682A (en) * | 2020-01-10 | 2020-06-05 | 京东方科技集团股份有限公司 | Method, device, medium and apparatus for predicting toxicity of drug |
| CN111444961A (en) * | 2020-03-26 | 2020-07-24 | 国家计算机网络与信息安全管理中心黑龙江分中心 | Method for judging internet website affiliation through clustering algorithm |
| CN111898364A (en) * | 2020-07-30 | 2020-11-06 | 平安科技（深圳）有限公司 | Neural network relation extraction method, computer device and readable storage medium |
| CN112860816A (en) * | 2021-03-01 | 2021-05-28 | 三维通信股份有限公司 | Construction method and detection method of interaction relation detection model of drug entity pair |
| WO2021159744A1 (en) * | 2020-09-27 | 2021-08-19 | 平安科技（深圳）有限公司 | Medicine classification method and apparatus, terminal device, and storage medium |
| CN113806531A (en) * | 2021-08-26 | 2021-12-17 | 西北大学 | Drug relationship classification model construction method, drug relationship classification method and system |
| CN115376668A (en) * | 2022-08-30 | 2022-11-22 | 温州城市智慧健康有限公司 | Big data business analysis method and system applied to intelligent medical treatment |
| WO2023173823A1 (en) * | 2022-03-17 | 2023-09-21 | 中国科学院深圳理工大学(筹) | Method for predicting interaction relationship of drug pair, and device and medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170308790A1 (en) * | 2016-04-21 | 2017-10-26 | International Business Machines Corporation | Text classification by ranking with convolutional neural networks |
| CN108363774A (en) * | 2018-02-09 | 2018-08-03 | 西北大学 | A kind of drug relationship sorting technique based on multilayer convolutional neural networks |
| CN108763216A (en) * | 2018-06-01 | 2018-11-06 | 河南理工大学 | A kind of text emotion analysis method based on Chinese data collection |
| CN109299264A (en) * | 2018-10-12 | 2019-02-01 | 深圳市牛鼎丰科技有限公司 | File classification method, device, computer equipment and storage medium |
-
2019
- 2019-03-08 CN CN201910174269.4A patent/CN110020671B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170308790A1 (en) * | 2016-04-21 | 2017-10-26 | International Business Machines Corporation | Text classification by ranking with convolutional neural networks |
| CN108363774A (en) * | 2018-02-09 | 2018-08-03 | 西北大学 | A kind of drug relationship sorting technique based on multilayer convolutional neural networks |
| CN108763216A (en) * | 2018-06-01 | 2018-11-06 | 河南理工大学 | A kind of text emotion analysis method based on Chinese data collection |
| CN109299264A (en) * | 2018-10-12 | 2019-02-01 | 深圳市牛鼎丰科技有限公司 | File classification method, device, computer equipment and storage medium |
Non-Patent Citations (2)
| Title |
|---|
| 杜晓东: "基于深度网络的药物关系抽取算法研究", 《中国优秀硕士学位论文全文数据库》 * |
| 马龙: "基于深度神经网络的药物关系挖掘方法研究", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111243682A (en) * | 2020-01-10 | 2020-06-05 | 京东方科技集团股份有限公司 | Method, device, medium and apparatus for predicting toxicity of drug |
| CN111444961A (en) * | 2020-03-26 | 2020-07-24 | 国家计算机网络与信息安全管理中心黑龙江分中心 | Method for judging internet website affiliation through clustering algorithm |
| CN111444961B (en) * | 2020-03-26 | 2023-08-18 | 国家计算机网络与信息安全管理中心黑龙江分中心 | Method for judging attribution of Internet website through clustering algorithm |
| CN111898364A (en) * | 2020-07-30 | 2020-11-06 | 平安科技（深圳）有限公司 | Neural network relation extraction method, computer device and readable storage medium |
| CN111898364B (en) * | 2020-07-30 | 2023-09-26 | 平安科技（深圳）有限公司 | Neural network relation extraction method, computer equipment and readable storage medium |
| JP2022538823A (en) * | 2020-09-27 | 2022-09-06 | 平安科技（深▲せん▼）有限公司 | Drug classification method and device, terminal device and storage medium |
| JP7227405B2 (en) | 2020-09-27 | 2023-02-21 | 平安科技（深▲せん▼）有限公司 | Drug classification method and device, terminal device and storage medium |
| WO2021159744A1 (en) * | 2020-09-27 | 2021-08-19 | 平安科技（深圳）有限公司 | Medicine classification method and apparatus, terminal device, and storage medium |
| CN112860816A (en) * | 2021-03-01 | 2021-05-28 | 三维通信股份有限公司 | Construction method and detection method of interaction relation detection model of drug entity pair |
| CN113806531A (en) * | 2021-08-26 | 2021-12-17 | 西北大学 | Drug relationship classification model construction method, drug relationship classification method and system |
| CN113806531B (en) * | 2021-08-26 | 2024-02-27 | 西北大学 | Drug relationship classification model construction method, drug relationship classification method and system |
| WO2023173823A1 (en) * | 2022-03-17 | 2023-09-21 | 中国科学院深圳理工大学(筹) | Method for predicting interaction relationship of drug pair, and device and medium |
| CN115376668A (en) * | 2022-08-30 | 2022-11-22 | 温州城市智慧健康有限公司 | Big data business analysis method and system applied to intelligent medical treatment |
| CN115376668B (en) * | 2022-08-30 | 2024-03-08 | 温州城市智慧健康有限公司 | Big data business analysis method and system applied to intelligent medical treatment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110020671B (en) | 2023-04-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110020671A (en) | The building of drug relationship disaggregated model and classification method based on binary channels CNN-LSTM network | |
| CN111078889B (en) | Method for extracting relationship between medicines based on various attentions and improved pre-training | |
| CN104318340B (en) | Information visualization methods and intelligent visible analysis system based on text resume information | |
| CN104391942B (en) | Short essay eigen extended method based on semantic collection of illustrative plates | |
| US10387784B2 (en) | Technical and semantic signal processing in large, unstructured data fields | |
| CN108388651A (en) | A kind of file classification method based on the kernel of graph and convolutional neural networks | |
| CN108182295A (en) | A kind of Company Knowledge collection of illustrative plates attribute extraction method and system | |
| CN106815369A (en) | A kind of file classification method based on Xgboost sorting algorithms | |
| CN108073569A (en) | A kind of law cognitive approach, device and medium based on multi-layer various dimensions semantic understanding | |
| CN106815293A (en) | System and method for constructing knowledge graph for information analysis | |
| CN104142998A (en) | Text classification method | |
| CN112347761B (en) | BERT-based drug relation extraction method | |
| CN107292348A (en) | A kind of Bagging_BSJ short text classification methods | |
| CN112420148A (en) | Medical image report quality control system, method and medium based on artificial intelligence | |
| Xu et al. | Chinese event detection based on multi-feature fusion and BiLSTM | |
| Kakade et al. | A neural network approach for text document classification and semantic text analytics | |
| CN109871414A (en) | Biomedical entity relationship classification method based on the context vector kernel of graph | |
| Wang et al. | Intelligent classification of tobacco leaves based on residual network | |
| Dutta et al. | Fake news prediction: a survey | |
| CN112215006A (en) | Organization named entity normalization method and system | |
| Falissard et al. | A deep artificial neural network based model for underlying cause of death prediction from death certificates | |
| CN110413899A (en) | The storage resource optimization method and system of server storage news | |
| CN110196976A (en) | Sentiment orientation classification method, device and the server of text | |
| CN111797945B (en) | Text classification method | |
| Aluna et al. | Electronic News Sentiment Analysis Application to New Normal Policy during the Covid-19 Pandemic Using Fasttext and Machine Learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |