CN108875809A - The biomedical entity relationship classification method of joint attention mechanism and neural network - Google Patents
The biomedical entity relationship classification method of joint attention mechanism and neural network Download PDFInfo
- Publication number
- CN108875809A CN108875809A CN201810554915.5A CN201810554915A CN108875809A CN 108875809 A CN108875809 A CN 108875809A CN 201810554915 A CN201810554915 A CN 201810554915A CN 108875809 A CN108875809 A CN 108875809A
- Authority
- CN
- China
- Prior art keywords
- word
- sentence
- vector
- entity
- biomedical
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Abstract
A kind of biomedical entity relationship classification method for combining attention mechanism and neural network, belongs to biomedical and data mining technology field, to solve biomedical entity relationship classification problem, main points be include S1, based on referring to the text-processing parsed;S2, mode input vector of the building based on attention mechanism;S3, biomedical entity relationship disaggregated model of the building based on two-way LSTM;S4, biomedical entity relationship classification is carried out using relationship disaggregated model.The sentence that the present invention is directed in biological document devises the parsing of the reference based on following, then from the basic unit word for constituting sentence, the insertion vector of word is weighting in using attention mechanism, the weight of the prominent keyword that the classification of biomedical entity relationship is had a major impact, make the relationship between candidate entity is apparent to be illustrated, carries out biomedical entity relationship classification.
Description
Technical field
The present invention relates to biomedical and data mining technology field, especially a kind of joint attention mechanism and mind
Biomedical entity relationship classification method through network.
Background technique
With the development of the bioinformatics of data-driven, found between the biomedical entity of prediction by calculation method
Relationship becomes a kind of trend.Text mining method based on calculating can be from a large amount of available biometric databases and non-structured text
Middle discovery mode and knowledge.Currently, the newest unstructured data of magnanimity is hidden in specialized database or scientific literature.In
It is to detect and predict that biomedical entity relationship is one effective and feasible from document and database using Text Mining Technology
Approach.In addition, this can also automate the database annotation process realized by having been manually done, biomedical map is additionally aided
Building.
Biomedical entity relationship research is excavated in traditional slave text mainly utilizes the machine in statistical machine learning theory
Device learning method such as support vector machines.These methods are dependent on well-designed core or the feature carefully designed.Feature
Design needs domain expert, but also is a labor-intensive skill sex work based on testing repeatedly.In addition, this
A little methods are limited to the generalization ability for not meeting word.Currently, method neural network based can be simple non-thread by constructing
Property model automatically learns the multi-level expression of non-structured text, has shown in each task of natural language processing field
Its potentiality are shown.There are two main neural network framework, that is, convolutional neural networks and Recognition with Recurrent Neural Network at present.The former is more
Suitable for learning continuous local mode.Although the latter can learn discontinuous global schema, its with biasing characteristic, i.e., after
The information of input is more dominant to target.
However, due to the general long sentence of Biomedical literature text and complicated clump sentence is in the majority, and the spy with scientific language
Property.Moreover, the word important to final relationship might not but appear in the rear portion of sentence.Although above-mentioned different text mining sides
Method explores the interactive relation between various classification biomedicine entities, but the biomedical entity in the majority in long complex sentence
It is not also very satisfactory on interaction classification performance.
Summary of the invention
The object of the present invention is to provide a kind of based on the neural network framework of attention mechanism in Biomedical literature
It has marked relationship between entity and has carried out more accurate and effectively relationship classification method.
The present invention solves technical solution used by prior art problem:A kind of joint attention mechanism and nerve net
The biomedical entity relationship classification method of network, includes the following steps:
S1, the text-processing based on reference parsing:The open data set marked of acquisition, using text-processing technology, refers to
Generation parsing and technology of prunning branches handle sentence, and processing method is as follows:
A1, initial treatment:One special symbol replaces the digital substring for being not belonging to biomedical entity substring.It deletes not
Bracket comprising candidate entity;Extensive for method, all biomedical entities are replaced with entity*, and wherein * indicates 0,
1,2,…;Delete the sentence of sentence or two entities with same symbol only comprising an entity;
A2, the reference based on following handle parsing:For in biological document text have colon ":" include
The sentence pattern of " following [refer to word] ", if two entities of candidate centering be located at colon ":" both sides, benefit
It is replaced with following rule, wherein [w] * indicates one or more word symbols:
Sentence pattern 1:entity1[w]*following[cataphora word]:[w]*entity2[w]*.
Rule 1:entity1[w]*following entity2.
Sentence pattern 2:[w]*following[cataphora word][w]*entity2:[w]*entity1[w]*.
Rule 2:[w]*following entity1[w]*entity2.
A3, sentence beta pruning:Each sentence in corpus is trimmed to fixed input length:It calculates between all candidate entities
Maximum spacing after, select the text size n bigger than this spacing as the input length of sentence.In order to reach this fixation
Input length n, the sentence of n be greater than for length, in addition to retain two entities and between all texts, and before and after retaining entity
The word of same number deletes word extra in sentence based on this strategy;It is less than the sentence of n for length, it is special with one
Symbol behind sentence patch not enough this length each input sentence.
S2, mode input vector of the building based on attention mechanism:The mode input vector include input feature vector to
Amount, input attention vector sum sentence vector, processing method are as follows:
The building of B1, input feature value:Sentence S={ the w of a given beta pruning1,w2,…,wi,…,wn, each word wi
It is expressed as three features:Word itself word, the PoS label of word and position;Position feature reflects current word wiBe mentioned
The relative distance of two candidate entities;The position sequence with current word in sentence subtracts the position sequence of entity 1 and entity 2 respectively
Obtain word wiDistance d relative to entity 1 and entity 21And d2;The PoS label for combining a word and it distinguishes a word in difference
Semanteme in sentence;By the parsing of Stamford resolver, processed sentence can obtain PoS of each word in sentence above
Label;Each feature group has an insertion dictionary;Assuming thatIt is the insertion dictionary of k-th of feature group, m herekIt is one super
Parameter indicates the dimension of feature insertion vector, lkIt is dictionary VkThe number of features for including;Word in each insertion dictionary can be by
Random initializtion, or it is vector initialising with word insertion trained in advance;To a word wiFor, by looking into corresponding insertion
It is a real value row vector that dictionary, which can map index of each feature in dictionary, to obtain word wiWord,
The corresponding insertion vector of PoS and position, is expressed asAndWith
B2, the building for inputting attention vector:Vector is embedded in using initial word of the attention mechanism into B1
It generates the word towards candidate entity and is embedded in vector;It is equal to the row vector α of sentence maximum length n using two lengthjTo quantify
Each word w in one sentenceiWith j-th candidates entity ejThe degree of correlation factor, wherein { 1,2 } j ∈;αjIt is defined as follows
Shown in formula:
Wherein,WithIt is word w respectivelyiWith j-th candidates entity ejWord be embedded in vector, score function regards face as
To the function of candidate entity, it is defined as follows:
Wherein, symbol dot indicates two vectorsWithOn dot product operations.m1It is the dimension of word insertion vector.It is related
The degree factorWithAct on word wiInitial word be embedded in vectorOn;The synergy α of the twoiIt is expressed asαiIt is applied to initial word insertion vectorOn regard as towards candidate entity word insertion vectorIt is defined
It is represented byWherein symbol * indicates the multiplication that step-by-step calculates;
The building of B3, sentence vector:
Construct vectorTo indicate word wiSemanteme, wherein xi∈Rm, m=m1+m2+
2m3, m2And m3It is the dimension of PoS and position insertion vector respectively;" | | " indicate attended operation.Then, sentence S can be expressed as
One real-valued vectors array Semb=[x1,x2,…,xi,...,xn]。
S3, biomedical entity relationship disaggregated model of the building based on two-way LSTM:It is deposited using two-way with length term
The Recognition with Recurrent Neural Network of storage unit LSTM carries out supervised learning modeling;Embeding layer therein can be realized and be obtained in B1 in step S2
The vector of each feature indicates;Attention layers of input are realized the building for inputting attention vector in step S2 in B2;Merge
Layer realizes the building of sentence vector in B3 in step S2;Two-way LSTM layers uses two-way LSTM network, it is by a forward direction LSTM
With a backward LSTM composition.For word wiFor, two LSTM collect available from the front to the back and from the front to the back respectively
Contextual information;The two-way LSTM layers of output in moment n is the connection of LSTM output vector in both directionhn
Indicate the high-level semantic of entire sentence;Utilize the logistical regression classifier (logistic for having softmax function
Regression) the classifier as candidate entity;Softmax function is with two-way LSTM layers of output hnIt is defeated as input
Y indicates probability distribution of the candidate translation example on each different relationship class label out;The probability distribution of jth class label is expressed as
P (y=j | S)=softmax (hnWs+bs), S indicates sentence, WsIt is the weight matrix for needing to learn, bsIt is the biasing for needing to learn
Vector;The corresponding label of the classification of maximum probability is the relationship type of candidate translation example, is expressed as
Wherein C is the set of presumable label of classifying in biological corpus;It sets for predicting that the loss function of error is damaged as cross entropy
Lose functionL is the sample number marked in training set, what subscript k indicated to be classified for k-th
Sentence, θ is all parameters in model, using RMSprop optimization algorithm (Resilient mean square
Propagation the parameter of loss function) is updated, training method is as follows:
e1:It is accumulative that parameter learning rate η in RMSprop optimization algorithm, momentum term parameter beta, initial rate v, gradient are set
Measure rate of decay ρ, gradient cumulative amount r and maximum number of iterations maxIter and most small quantities of m.
e2:Gradient cumulative amount r=0, the number of iterations iterCount=0 are initialized, initialization error current and last time are accidentally
Difference is infinity, i.e. currError=lastError=0, random initializtion parameter θ;
e3:Parameter θ more in new model as follows:
θ←θ+v
Wherein ⊙ indicates the multiplication that step-by-step calculates.
e4:The number of iterations iterCount adds 1, calculates error current according to the loss function in step S3, if current miss
Difference is greater than last time error, that is, currError>LastError or the number of iterations are equal to maximum number of iterations, that is, iterCount
=maxIter then meets the condition of convergence and goes to step e5;Otherwise, lastError=currError continues e3;
e5:All parameter θs in preservation model are into file.
S4, biomedical entity relationship is predicted:The parameter value θ that read step e5 training obtains passes to the network mould in S3
Type;The sentence including at least two biomedical entities in Biomedical literature is extracted, matching principle constructs two-by-two according to entity
Relationship example;Initial treatment is carried out to text using the method in step S1, refers to parsing and beta pruning;Then using in step S2
The method of B1, obtain respectively word in sentence, the PoS label of each word and current word and two entities with respect to away from
From looking into corresponding vector dictionary and obtain the index of these features;Based on input, model can export each pair of relationship example each
Probability value in classification, wherein that maximum one kind of probability value is the corresponding class label of candidate translation example, to obtain candidate
Relationship type between entity.
The beneficial effects of the present invention are:The present invention innovatively proposes one by introducing input attention mechanism
The biomedical entity relationship classification method of kind joint attention mechanism and neural network.This method is to a certain extent gram
The biasing defect of Recognition with Recurrent Neural Network LSTM is taken, this defect makes LSTM neglect one when handling biomedical text long sentence
Information before the sentence important to final classification result a bit.Moreover, this method can efficiently identify word in long complicated sentence
Between short distance mode and remote mode, thus realize automatically and efficiently in biological document entity carry out relation
Class.It is evaluated and tested in corpus on tri- data sets of DrugBank, Medline and Overall in embodiment DDIExtraction 2013
Primary evaluation index F-score on increase rate be respectively 3.3%, 21.7% and 6.2%, demonstrate the method for the present invention pair
The validity that entity relationship is classified in Biomedical literature.
Detailed description of the invention
Fig. 1 is the flow diagram of relationship classification method of the present invention;
Fig. 2 is the biomedical entity relationship disaggregated model signal of present invention joint attention mechanism and neural network
Figure;
Fig. 3 is the visualization that the embodiment of the present invention inputs attention.
Specific embodiment
Below in conjunction with the drawings and the specific embodiments, the present invention will be described:
Embodiment:
According to the above-mentioned description for being directed to method and system specific embodiment involved in the present invention, in conjunction with specific embodiments
It is illustrated.
The present embodiment uses DrugBank the and Medline data set in 2013 evaluation and test task of DDIExtraction,
In be divided into training set and test set again.DrugBank training set and test set separately include 31270 and 1221 sentences, also distinguish
Represent the sentence in the sentence and biomedical article in Service functions.In experimentation, in two datasets
Training dataset merge as training set, using two test sets are constant and the union of two test sets when test
Overall。
It is a kind of joint attention mechanism and neural network biomedical entity relationship classification method specific steps such as
Under:
1, based on the text-processing for referring to parsing:The above-mentioned data set marked is acquired, using text-processing technology, is referred to
After parsing is handled, each sentence in corpus is trimmed to fixed input length.It calculates between all candidate entities most
After big spacing, each 5 words before and after two candidate entities are kept respectively.Then, 85 are set as by the sentence length n of beta pruning.
2, the mode input vector based on attention mechanism is constructed:The mode input vector include input feature vector to
Amount, input attention vector sum sentence vector, steps are as follows:
For input feature vector term vector, the pre-training corpus that word indicates includes two parts, amounts to about 2.5G byte.One
It point is the abstract of related article in the Medline before obtained and key word of the inquiry " drug " in PubMed 2016.
Another part is the corpus in 2013 evaluation and test task of DDIExtraction.PoS vector has used marked PoS label
DDIExtraction 2013 evaluates and tests the sentence in task corpus as training corpus.Two kinds of insertion vector all passes through out
Source tool word2vec tool training, using Skip-Gram model and the negative method of sampling.Position insertion vector uses random initial
Change assignment, institute's directed quantity is initialized to meet the random value of normal state standard profile.The dimension of word insertion vector is set as m1=
200, PoS and position insertion vector dimension m2、m3It is respectively set to 10.
When building input attention vector, vector is embedded in using input word of the attention mechanism into B1;Benefit
It is equal to the row vector of sentence maximum length 85 with two lengthTo quantify in a sentence
Each word wiWith j-th candidates entity ejThe degree of correlation factor, wherein { 1,2 } j ∈.Score function isWherein i≤85.Then, the degree of correlation factorWithAct on word wiInitial word
It is embedded in vectorOn;The synergy α of the twoiIt is expressed asSuch as the sentence in sentence corpus
“Synergism was also noted when entity0 was combined with entity1and
Entity2. ", above formula α is utilizediThe result of the word grade attention calculated is as shown in Figure 3.It can be sent out from figure
Existing, word " synergism ", " combined " and " when " has higher attention weight relative to other words.Due to two
Real relationship between a candidate's entity entity0 and entity1 is " effect ", therefore, what these were calculated
Attention value is reasonable.Next, αiIt is applied to initial word insertion vectorOn obtain towards candidate entity word insertion
VectorWherein symbol * indicates the multiplication that step-by-step calculates.
Finally, including that word is embedded in vectorPoS is embedded in vectorAnd position is embedded in vectorWithInside all
Vector connect into a new vectorIndicate word wiSemanteme, wherein xi∈Rm, m=230.
Then, sentence S can be expressed as a real-valued vectors array Semb=[x1,x2,...,xi,...,x85]。
3, classified using the framework based on two-way LSTM to biomedical entity relationship and carry out supervised learning modeling:Before
It states the training set data that step process is crossed and is transmitted to two-way LSTM Recognition with Recurrent Neural Network progress supervised learning as shown in Fig. 2
Modeling;The two-way LSTM network used, it is made of a forward direction LSTM and a backward LSTM.For word wiFor, two
A LSTM collects available contextual information from the front to the back and from the front to the back respectively.The adaptive calculating answered the door of three of LSTM according to
By pervious state ht-1With current input state xt, formula is as follows:
it=σ (Wi·xt+Ui·ht-1+bi)
ft=σ (Wf·xt+Uf·ht-1+bf)
ot=σ (Wo·xt+Uo·ht-1+bo)
Wherein σ indicates sigmoid function, and the value range of three doors is [0,1].After having three doors, by former
Cell state Ct-1And candidate stateSynergy determine current cell state Ct.The output h of LSTM unittIt is by exporting
The fixed cell state of thresholding updates shown in following formula:
BLSTM is in the connection that the output of moment n=85 is LSTM output vector in both directionWherein,
LSTM implies unit number and is set as inputting identical dimension 230, h with LSTMnDimension be 460.
Using the logistical regression classifier with softmax function as the classifier of candidate entity;Softmax letter
Number is with two-way LSTM layers of output hnAs input, exporting y indicates candidate translation example on each different relationship class label
Probability distribution;The probability distribution of jth class label is expressed as p (y=j | S)=softmax (hnWs+bs), S indicates sentence, WsIt is
The weight matrix for needing to learn, bsIt is the bias vector for needing to learn;The corresponding label of the classification of maximum probability is candidate translation example
Relationship type, be expressed asSymbol C=5.Next it is used to predict the loss function of error
For cross entropy loss functionL is the sample number marked in training set, and subscript k indicates kth
A sentence being classified.
4, the parameter of training entity relationship disaggregated model:The parameter of loss function is updated with RMSprop optimization algorithm,
Middle parameter learning rate η=0.001, momentum term parameter beta=0.9.Ginseng is adjusted using 5 times of cross validations of Sentence-level on training set
Number saves the parameter of acquisition into file with the performance of optimization system.
5, biomedical entity relationship is predicted:It reads the parameter value θ that file training obtains in 4 and passes to net shown in Fig. 2
Network model;Extracting in test set includes at least two biomedical sentences, and according to entity, matching principle building relationship is real two-by-two
Example;Text is handled using the method for B1 in step S1 and S2, respectively the PoS of the word in acquisition sentence, each word
Label and relative distance with two entities, look into corresponding vector dictionary and obtain the index of these features and pass to Fig. 2 mould
Embeding layer in type;Based on input, model exports probability value of each pair of relationship example in each classification, and wherein probability value is maximum
That one kind be the corresponding class label of candidate translation example, thus the relationship type between obtaining entity.
For the validity of verification method, three kinds of control methods of experimental selection:
(1) the method RAIHANI based on SVM:System RAIHANI devises many rule and feature, such as chunk, touching
Send out word, negative filtering and SAME_BLOK etc..Classifier also individually designed many of this method for each subtype
Different features.
(2) the method jointAB-LSTM based on LSTM:JointAB-LSTM has combined two LSTM networks, one of them
Attention technology has been used at pooling layers, has been on 2013 corpus of DDIExtraction based on neural network framework method
In existing best one of LSTM method.
(3) the method MCCNN based on convolutional neural networks:MCCNN has used multichannel word to be embedded in vector, is
Based on best one of CNN method existing in network architecture method on 2013 corpus of DDIExtraction.
In addition, (2) and (3), which are removed, uses common text-processing technology, also excluded using filtering technique incoherent
Negative example.
Table 1 gives 4 kinds of methods including the method for the present invention and 3 control methods on experiment corpus described previously
The F-score evaluation index test result of Overall test set co-relation classification, F-score here refer in more classification often
The micro- average Micro_F used.F-score is the evaluation index that text field relationship extracts frequent accepted standard, it determines
Justice is as follows:
Wherein Micro_P indicates precision, and Micro_R indicates recall rate, TPi(true positives, real positive example) generation
Table sort device is predicted as practical in the positive example example of the i-th class relationship being also the number of positive example, FPi(false positives, it is empty
False positive example) it represents classifier and is predicted as actually being negative in the positive example example of the i-th class relationship the number of example, FNi(false
Negatives, false negative example) represent classifier prediction be negative example example in be actually the i-th class relationship positive example number.Precision
Micro_P and recall rate Micro_R considers looking into for algorithm respectively and parasexuality and looks into full property.But the two indexs cannot be more fully
The performance of a categorizing system is embodied, therefore, is generally used for playing a balance between precision Micro_P and recall rate Micro_R
The Micro_F value of effect evaluates the overall performance of an algorithm.
Value in table 1 with runic is the peak of each relationship classification on corresponding data collection, and symbol "-" indicates not corresponding
Value (similarly hereinafter)." DEC " indicates the relationship detection for only carrying out binary, i.e., either with or without relationship." CLA " indicates relationship classification.
" MEC ", " EFF ", " ADV " and " INT " indicates " mechanism ", " effect ", " advice " and " int " type respectively.
Table 1 not homologous ray in OVERALLThe performance of test set co-relation classification compares
Table 2 give including the method for the present invention and wherein 3 kinds of methods including 2 control methods in experiment language described previously
Expect the F-score evaluation index test result of upper three test set co-relations classification (CLA), wherein joint AB-LSTM method
Do not provide experiment corresponding experimental result.
Performance of the homologous ray on three data sets does not compare table 2
Listed experimental result can be seen that method proposed by the invention obtains on experimental data set from above-mentioned two table
The detection (DEC) and classification (CLA) performance obtained.This also illustrate in this method by based on refer to parsing text-processing,
The input vector of attention mechanism construction and suitable model for the classification of biological document entity relationship, truly have it is very high
Performance boost.Tri- data sets of DrugBank, Medline and Overall in corpus are evaluated and tested in DDIExtraction 2013
On primary evaluation index F-score on increase rate be respectively 3.3%, 21.7% and 6.2%, demonstrate the method for the present invention
The validity classified to entity relationship in Biomedical literature.
The above content is combine specific optimal technical scheme further detailed description of the invention, and it cannot be said that
Specific implementation of the invention is only limited to these instructions.For those of ordinary skill in the art to which the present invention belongs, exist
Under the premise of not departing from present inventive concept, a number of simple deductions or replacements can also be made, all shall be regarded as belonging to of the invention
Protection scope.
Claims (6)
1. a kind of biomedical entity relationship classification method for combining attention mechanism and neural network, which is characterized in that packet
Include following steps:S1, the text-processing based on reference parsing;S2, mode input vector of the building based on attention mechanism;
S3, biomedical entity relationship disaggregated model of the building based on two-way LSTM;S4, biomedicine is carried out using relationship disaggregated model
Entity relationship classification.
2. the biomedical entity relationship point of a kind of joint attention mechanism and neural network according to claim 1
Class method, which is characterized in that the step:
S1, the text-processing based on reference parsing:The open data set marked of acquisition, using text-processing technology to text into
Row initialization process, using based on following reference dissection process and technology of prunning branches sentence is handled;
S2, mode input vector of the building based on attention mechanism:The mode input vector include input feature value,
Attention vector sum sentence vector is inputted, processing method is as follows:
The building of B1, input feature value:Sentence S={ the w of a given beta pruning1,w2,…,wi,…,wn, each word wiBy table
It is shown as three feature vectors:Word itself word, the PoS label of word and position, are expressed asAndWith
B2, the building for inputting attention vector:Using initial word insertion vector of the attention mechanism into B1 with life
Vector is embedded at the word towards candidate entity;It is equal to the row vector α of sentence maximum length n using two lengthjTo quantify one
Each word w in a sentenceiWith the degree of correlation factor of j-th candidates entity, αjIt is defined as follows shown in formula:
Wherein, { 1,2 } j ∈,WithIt is word w respectivelyiWith j-th candidates entity ejWord be embedded in vector, score function quilt
Regard the function towards candidate entity as, is defined as follows:
Wherein, symbol dot indicates two vectorsWithOn dot product operations;m1It is the dimension of word insertion vector;Degree of correlation
The factorWithAct on word wiInitial word be embedded in vectorOn, the synergy α of the twoiIt is expressed as
αiIt is applied to initial word insertion vectorOn regard as towards candidate entity word insertion vectorIts definition is expressed asWherein symbol * indicates the multiplication that step-by-step calculates;
The building of B3, sentence vector:
Construct vectorTo indicate word wiSemanteme, wherein xi∈Rm, m=m1+m2+2m3, m2With
m3It is the dimension of PoS and position insertion vector respectively;" | | " indicate attended operation;Sentence S is expressed as a real-valued vectors array
Semb=[x1,x2,…,xi,...,xn];
S3, biomedical entity relationship disaggregated model of the building based on two-way LSTM:List is stored with length term using two-way
The Recognition with Recurrent Neural Network of member carries out supervised learning modeling;Including:Embeding layer inputs attention layers, merges layer, two-way
LSTM layers;Using the logistical regression classifier with softmax function as the classifier of candidate translation example;Softmax letter
Number is with two-way LSTM layers of output hnAs input, exporting y indicates candidate translation example on each different relationship class label
Probability distribution;The probability distribution of jth class label is expressed as p (y=j | S)=softmax (hnWs+bs), S indicates sentence, WsIt is
The weight matrix for needing to learn, bsIt is the bias vector for needing to learn;The corresponding label of the classification of maximum probability is candidate translation example
Relationship type, be expressed asWherein C is the collection of presumable label of classifying in biological corpus
It closes;The loss function for predicting error is set as cross entropy loss functionL is trained
The sample number of mark is concentrated, subscript k indicates that k-th of sentence being classified, θ are all parameters in model;
S4, biomedical entity relationship is predicted:Initial treatment is carried out to text using the method in step S1, parsing is referred to and cuts
Branch;Then using the method for B1 in step S2, word in sentence, the PoS label of each word and current single are obtained respectively
The relative distance of word and two entities looks into corresponding vector dictionary and obtains the index of these features;Based on input, model can be defeated
Probability value of each pair of relationship example in each classification out, wherein that maximum one kind of probability value is the corresponding class of candidate translation example
Distinguishing label, thus the relationship type between obtaining entity.
3. the biomedical entity relationship point of a kind of joint attention mechanism and neural network according to claim 2
Class method, it is characterised in that in step S1, reference processing and sentence beta pruning of the initial treatment based on following, table
It is shown as:
A1, initial treatment:One special symbol replaces the digital substring for being not belonging to biomedical entity substring;Deletion does not include
The bracket of candidate entity;Extensive for method, all biomedical entities are replaced with entity*, and * indicates 0,1,2 ...;
Delete the sentence of sentence or two entities with same symbol only comprising an entity;
A2, the reference based on following handle parsing:For in biological document text with ":" include " following
The sentence pattern of [refer to word] ", if two entities of candidate centering be located at ":" both sides, using following rule into
Row replacement, wherein [w] * indicates one or more word symbols:
Sentence pattern 1:entity1[w]*following[cataphoraword]:[w]*entity2[w]*.
Rule 1:entity1[w]*following entity2.
Sentence pattern 2:[w]*following[cataphora word][w]*entity2:[w]*entity1[w]*.
Rule 2:[w]*following entity1[w]*entity2.
A3, sentence beta pruning:Each sentence in corpus is trimmed to fixed input length:It calculates between all candidate entities most
After big spacing, select the text size n bigger than this spacing as the input length of sentence;In order to reach the input of this fixation
Length n is greater than length the sentence of n, in addition to retain two entities and between all texts, and retain identical before and after entity
The word of number deletes word extra in sentence based on this strategy;It is less than the sentence of n for length, with a special symbol
Each input sentence of patch not enough this length number behind sentence.
4. the biomedical entity relationship point of a kind of joint attention mechanism and neural network according to claim 2
Class method, which is characterized in that in step S2, described in step B2AndWithAcquisition methods are as follows:
Current word w is used respectivelyiPosition sequence in sentence subtracts the position sequence of entity 1 and entity 2, and word w can be obtainediRelative to entity 1
With the distance d of entity 21And d2;The PoS label for combining a word and it distinguishes semanteme of the word in different sentences;Pass through this
Processed sentence obtains PoS label of each word in sentence above for smooth good fortune resolver parsing;Each feature group have one it is embedding
Enter dictionary;Assuming thatIt is the insertion dictionary of k-th of feature group, wherein mkIt is the dimension of a feature insertion vector, lkIt is word
Allusion quotation VkThe number of features for including;Word in each insertion dictionary can be embedded in by random initializtion, or with word trained in advance
It is vector initialising;To a word wiFor, rope of each feature in dictionary can be mapped by looking into corresponding insertion dictionary
Drawing symbol is a real value row vector, to obtain word wiThe corresponding insertion vector of word, PoS and position,With
AndWith
5. the biomedical entity relationship point of a kind of joint attention mechanism and neural network according to claim 2
Class method, which is characterized in that in step s3, two-way LSTM layers uses two-way LSTM network, it is by a forward direction LSTM and one
A backward LSTM composition;The two-way LSTM layers of output in moment n is the connection of LSTM output vector in both directionhnIndicate the high-level semantic of entire sentence.
6. the biomedical entity relationship point of a kind of joint attention mechanism and neural network according to claim 2
Class method, which is characterized in that in step s3, loss function is updated using RMSprop optimization algorithm.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810554915.5A CN108875809A (en) | 2018-06-01 | 2018-06-01 | The biomedical entity relationship classification method of joint attention mechanism and neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810554915.5A CN108875809A (en) | 2018-06-01 | 2018-06-01 | The biomedical entity relationship classification method of joint attention mechanism and neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108875809A true CN108875809A (en) | 2018-11-23 |
Family
ID=64336156
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810554915.5A Pending CN108875809A (en) | 2018-06-01 | 2018-06-01 | The biomedical entity relationship classification method of joint attention mechanism and neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108875809A (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109710930A (en) * | 2018-12-20 | 2019-05-03 | 重庆邮电大学 | A kind of Chinese Resume analytic method based on deep neural network |
CN109754012A (en) * | 2018-12-29 | 2019-05-14 | 新华三大数据技术有限公司 | Entity Semantics relationship classification method, model training method, device and electronic equipment |
CN109871414A (en) * | 2019-01-15 | 2019-06-11 | 大连交通大学 | Biomedical entity relationship classification method based on the context vector kernel of graph |
CN110060773A (en) * | 2019-04-22 | 2019-07-26 | 东华大学 | Alzheimer's disease progression of the disease forecasting system based on two-way LSTM |
CN110377912A (en) * | 2019-07-24 | 2019-10-25 | 贵州大学 | A kind of relation recognition method based on multichannel deep neural network |
CN110688486A (en) * | 2019-09-26 | 2020-01-14 | 北京明略软件系统有限公司 | Relation classification method and model |
CN111222338A (en) * | 2020-01-08 | 2020-06-02 | 大连理工大学 | Biomedical relation extraction method based on pre-training model and self-attention mechanism |
CN111353306A (en) * | 2020-02-22 | 2020-06-30 | 杭州电子科技大学 | Entity relationship and dependency Tree-LSTM-based combined event extraction method |
CN111859967A (en) * | 2020-06-12 | 2020-10-30 | 北京三快在线科技有限公司 | Entity identification method and device and electronic equipment |
CN111950279A (en) * | 2019-05-17 | 2020-11-17 | 百度在线网络技术(北京)有限公司 | Entity relationship processing method, device, equipment and computer readable storage medium |
CN112036181A (en) * | 2019-05-14 | 2020-12-04 | 上海晶赞融宣科技有限公司 | Entity relationship identification method and device and computer readable storage medium |
CN113032618A (en) * | 2021-03-26 | 2021-06-25 | 齐鲁工业大学 | Music recommendation method and system based on knowledge graph |
WO2022028692A1 (en) * | 2020-08-05 | 2022-02-10 | Siemens Aktiengesellschaft | Enhancement of bootstrapping for information extraction |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106354710A (en) * | 2016-08-18 | 2017-01-25 | 清华大学 | Neural network relation extracting method |
US20170229071A1 (en) * | 2016-02-05 | 2017-08-10 | Hand Held Products, Inc. | Dynamic identification badge |
CN107818141A (en) * | 2017-10-10 | 2018-03-20 | 大连理工大学 | Incorporate the biomedical event extraction method of structuring key element identification |
-
2018
- 2018-06-01 CN CN201810554915.5A patent/CN108875809A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170229071A1 (en) * | 2016-02-05 | 2017-08-10 | Hand Held Products, Inc. | Dynamic identification badge |
CN106354710A (en) * | 2016-08-18 | 2017-01-25 | 清华大学 | Neural network relation extracting method |
CN107818141A (en) * | 2017-10-10 | 2018-03-20 | 大连理工大学 | Incorporate the biomedical event extraction method of structuring key element identification |
Non-Patent Citations (4)
Title |
---|
ZENG DJ ET AL: "《Relation Classification Via Convolutional Deep Neural Network》", 《COLING 2014》 * |
ZHENG WEI ET AL: "《An attention-based effective neural model for drug-drug interactions extraction》", 《BMC BIOINFORMATICS 18》 * |
伊恩•古德费洛著: "《深度学习》", 31 July 2017, 人民邮电出版社 * |
李强等: "《生物医用多孔金属材料的制备及表面改性》", 30 August 2016, 冶金工业出版社 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109710930A (en) * | 2018-12-20 | 2019-05-03 | 重庆邮电大学 | A kind of Chinese Resume analytic method based on deep neural network |
CN109754012A (en) * | 2018-12-29 | 2019-05-14 | 新华三大数据技术有限公司 | Entity Semantics relationship classification method, model training method, device and electronic equipment |
CN109871414A (en) * | 2019-01-15 | 2019-06-11 | 大连交通大学 | Biomedical entity relationship classification method based on the context vector kernel of graph |
CN110060773A (en) * | 2019-04-22 | 2019-07-26 | 东华大学 | Alzheimer's disease progression of the disease forecasting system based on two-way LSTM |
CN110060773B (en) * | 2019-04-22 | 2023-10-27 | 东华大学 | Alzheimer's disease progression prediction system based on bidirectional LSTM |
CN112036181A (en) * | 2019-05-14 | 2020-12-04 | 上海晶赞融宣科技有限公司 | Entity relationship identification method and device and computer readable storage medium |
CN111950279A (en) * | 2019-05-17 | 2020-11-17 | 百度在线网络技术(北京)有限公司 | Entity relationship processing method, device, equipment and computer readable storage medium |
CN110377912A (en) * | 2019-07-24 | 2019-10-25 | 贵州大学 | A kind of relation recognition method based on multichannel deep neural network |
CN110688486A (en) * | 2019-09-26 | 2020-01-14 | 北京明略软件系统有限公司 | Relation classification method and model |
CN111222338A (en) * | 2020-01-08 | 2020-06-02 | 大连理工大学 | Biomedical relation extraction method based on pre-training model and self-attention mechanism |
CN111353306A (en) * | 2020-02-22 | 2020-06-30 | 杭州电子科技大学 | Entity relationship and dependency Tree-LSTM-based combined event extraction method |
CN111859967A (en) * | 2020-06-12 | 2020-10-30 | 北京三快在线科技有限公司 | Entity identification method and device and electronic equipment |
CN111859967B (en) * | 2020-06-12 | 2024-04-09 | 北京三快在线科技有限公司 | Entity identification method and device and electronic equipment |
WO2022028692A1 (en) * | 2020-08-05 | 2022-02-10 | Siemens Aktiengesellschaft | Enhancement of bootstrapping for information extraction |
CN113032618A (en) * | 2021-03-26 | 2021-06-25 | 齐鲁工业大学 | Music recommendation method and system based on knowledge graph |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108875809A (en) | The biomedical entity relationship classification method of joint attention mechanism and neural network | |
CN112214995B (en) | Hierarchical multitasking term embedded learning for synonym prediction | |
CN107992597B (en) | Text structuring method for power grid fault case | |
CN109446338B (en) | Neural network-based drug disease relation classification method | |
CN112001185B (en) | Emotion classification method combining Chinese syntax and graph convolution neural network | |
CN110287481B (en) | Named entity corpus labeling training system | |
Chang et al. | Chinese named entity recognition method based on BERT | |
CN108984526A (en) | A kind of document subject matter vector abstracting method based on deep learning | |
CN108460089A (en) | Diverse characteristics based on Attention neural networks merge Chinese Text Categorization | |
CN112001186A (en) | Emotion classification method using graph convolution neural network and Chinese syntax | |
CN111222318B (en) | Trigger word recognition method based on double-channel bidirectional LSTM-CRF network | |
CN111554360A (en) | Drug relocation prediction method based on biomedical literature and domain knowledge data | |
CN110750645A (en) | Cross-domain false comment identification method based on countermeasure training | |
CN110083836A (en) | A kind of key evidence abstracting method of text prediction result | |
CN115952292B (en) | Multi-label classification method, apparatus and computer readable medium | |
CN111476024A (en) | Text word segmentation method and device and model training method | |
CN111582506A (en) | Multi-label learning method based on global and local label relation | |
Lim et al. | Bibliographic analysis on research publications using authors, categorical labels and the citation network | |
CN112836051A (en) | Online self-learning court electronic file text classification method | |
CN110245234A (en) | A kind of multi-source data sample correlating method based on ontology and semantic similarity | |
Kang et al. | A Research Toward Chinese Named Entity Recognition Based on Transfer Learning | |
CN116245107A (en) | Electric power audit text entity identification method, device, equipment and storage medium | |
CN116070700A (en) | Biomedical relation extraction method and system integrating iterative active learning | |
Kim | Research on Text Classification Based on Deep Neural Network | |
Kohsasih et al. | Sentiment Analysis for Financial News Using RNN-LSTM Network |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181123 |
|
RJ01 | Rejection of invention patent application after publication |