CN108182976A - A kind of clinical medicine information extracting method based on neural network - Google Patents
A kind of clinical medicine information extracting method based on neural network Download PDFInfo
- Publication number
- CN108182976A CN108182976A CN201711462492.6A CN201711462492A CN108182976A CN 108182976 A CN108182976 A CN 108182976A CN 201711462492 A CN201711462492 A CN 201711462492A CN 108182976 A CN108182976 A CN 108182976A
- Authority
- CN
- China
- Prior art keywords
- vector
- word
- max
- character
- lstm
- 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
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Machine Translation (AREA)
Abstract
The invention discloses a kind of clinical medicine information extracting methods based on neural network, there is many professional vocabularies, uncommon vocabulary and temporal expression with digital and character composition etc. in usual medicine text, but the character vector obtained using convolutional neural networks can include the morphologic information of word, therefore can be very good to solve the problems, such as this.Two-way LSTM used herein can capture contextual information well simultaneously.In addition, avoiding this process of artificial design feature in machine learning using the method for neural network, the problem of field adapts to can solve.The method of the present invention all yields good result on different data fields, and energy efficiently and accurately intelligently extracts with practical value and research significance information from magnanimity medical data.
Description
Technical field
The present invention relates to biomedical text natural language processing fields, and in particular to a kind of clinic based on neural network
Medical information extracting method.
Background technology
Under the historical background of big data and " intelligent medical treatment ", text mining and the information extraction technique of medical domain have become
" most important thing " focused in recent years for researchers.Medicine text traditional Chinese medicine entity information is extracted, if time, event are medicine
One of vital task of big data processing.But there is data volume Pang with the unstructured medicine text data that natural language is stated
Greatly, it is complicated, generate features, the related researchers such as speed is fast and quickly and accurately to be obtained from a large amount of texts valuable
Knowledge and information be extremely difficult.So how not only intelligently to extract efficiently but also from magnanimity medical data with practical value
And research significance medical knowledge information and carry out the expression of structuring, deeper into grasp that those are unknown, threaten people
The disease information of class health is very urgent.
Existing information extracting method is broadly divided into Rule Extraction method and the method for machine learning.But due to natural language
Complexity, the rule of structure are difficult to cover all entity types;According to based on the machine learning algorithm for having supervision,
Then due to the particularity and complexity of medicine text, the entity information degree of overlapping that different illnesss is included is very low, entity class
Also very diversification, thus different illnesss is directed to, it is required in advance being used as a part of text progress handmarking training text,
And the feature artificially built is difficult to cover the feature of all entity class.When frontier when something goes wrong, can only be by again
Label, training data complete learning model building.However, medical data, which is labeled, to be needed to expend a large amount of of high professional
Time, cost are extremely high.
Invention content
It is an object of the invention to overcome above-mentioned problems of the prior art, a kind of facing based on neural network is provided
Bed medical information extracting method.
In order to achieve the above object, the present invention adopts the following technical scheme that:
Step 1:Word segmentation processing is carried out to training text and test text first, the training text obtained after participle is used
BIO labels are marked.
Step 2:Its corresponding original character vector table is built for 24 English alphabets and other common characters, and with
Biomedical article in PubMed databases is the initial term vector of building of corpus.Text after being segmented based on step 1, is passed through
It tables look-up and obtains each corresponding initial term vector of word and the corresponding original character vector of each character.
Step 3:Structure is carried based on the character vector that step 2 generates and the neural network medicine entity of term vector joint input
Modulus type.Model is divided into encoder, decoder and grader three parts, respectively using CNN networks and Bi-LSTM networks to word
Symbol vector and the input of term vector are encoded, and using Bi-LSTM network decodings, complete to classify using softmax graders.
Step 4:Training data after being marked using BIO trains above-mentioned model, passes through the reality in comparative training data
BIO labels and this category of model obtain after BIO labels difference, adjustment model parameter is with Optimum Classification performance.
Step 5:Using test data, to step 4, trained model is tested, and is obtained eventually by softmax graders
To BIO sequence labels extract medicine entity.
The step 2, includes the following steps:
Step 2.1:Its corresponding character vector, specific practice are initialized to existing all English characters using random number
Be for initialization vector per one-dimensional, all fromIn the range of generate at random one number carry out assignment,
Wherein dim is the dimension of character vector, and all original character vectors are gathered together and generate an original character vector table,
The size of dim is between 30 to 50;
Step 2.2:For all characters in training text and test text, generated by searching for step 2.1 initial
Character vector table obtains its corresponding original character vector.
Step 2.3:Using GLoVe term vectors model method disclosed in Stamford, the biology in PubMed databases is chosen
Medical article generates initial word vector table for corpus.
Step 2.4:For all words in training text and test text, generated by searching for step 2.3 initial
Term vector table obtains its corresponding initial term vector.
The step 3, includes the following steps:
Step 3.1:Using step 2.2 generate original character vector, by form each word character its it is corresponding just
The beginning character vector generation original character matrix that is stitched together is sent into convolutional neural networks and is encoded (as unit of each word).
The original character matrix of convolutional neural networks is input to for each, first passes around a convolutional layer, using convolution kernel by group
Original character vector into each word adjacent character carries out convolution.Then by the Input matrix of convolutional layer output a to maximum
Pond layer is directed to each row vector of convolutional layer output matrix, utilizes that one-dimensional generation of maximum pond layer choosing access value maximum
The information that the entire row vector of table includes, then after maximum pond layer, output one it is identical with original character vector dimension to
Amount.
Step 3.2:The initial term vector generated using step 2.3, by the corresponding initial word of words all in each sentence
Vector is stitched together to be fed through in a Bi-LSTM and be encoded.It is included in wherein two-way LSTM there are two LSTM layers, one is
Forward direction LSTM, one is backward LSTM.T-th of word being then directed in a sentence is obtained to LSTM comprising the using preceding
One word to t-th word context information corresponding vector hft, obtained using backward LSTM comprising t-th of word to the end
The corresponding vector h of one word context informationbt, vector is stitched together, the term vector h as t-th of wordt=(hft,
hbt)。
Step 3.3:If the corresponding character vectors of each word i of CNN layers of output are { c1, c2..., cdim, Bi-
The corresponding term vectors of the LSTMencoding layers of each word i of output are { wh1, wh2..., whn, then it is normalized, i.e.,
If cmaxThat maximum for character vector numerical value is one-dimensional, if whmaxFor word to numerical quantity it is maximum that is one-dimensional, then final character
Vector is { c1/cmax, c2/cmax..., cdim/cmax, final term vector is { wh1/whMax,wh2/whmax..., whn/whmax}.It will
Two above vector is spliced to obtain the corresponding final vector m of each wordi={ c1/cmax, c2/cmax..., cdim/cmax,
wh1/whmax, wh2/whmax..., whn/whmax}.The corresponding final vector of words all in each sentence is cascaded up to be formed
Then final vector matrix is input to Bi-LSTM networks as unit of sentence and is decoded.
Step 3.4:By the decoded output vectors of Bi-LSTM by final softmax layers, obtain to each word most
Whole BIO label results.
Compared with prior art, the present invention has technique effect beneficial below:
The present invention is based on the methods of neural network, realize that extraction is believed with clinical medicine from medical unstructured text data
Relevant entity information is ceased, such as time, event expression.There is many professional vocabularies, uncommon vocabulary in usual medicine text
And temporal expression with digital and character composition etc., it is difficult accurately to carry it using the morphology syntactic feature artificially designed
It takes.But the character vector obtained using convolutional neural networks can include the morphologic information of word, therefore can be very good
Solve the problems, such as this.Two-way LSTM used herein can capture contextual information well simultaneously.In addition, use neural network
Method avoids this process of artificial design feature in machine learning, can solve the problem of field adapts to.The present invention's
Method all yields good result on different data fields, and energy efficiently and accurately is intelligently extracted from magnanimity medical data
With practical value and research significance information.
Description of the drawings
Fig. 1 is the method flow diagram of the clinical medicine information extracting method of the present invention;
Specific embodiment
With reference to test example and specific embodiment, the present invention is described in further detail.But this should not be understood
Range for the above-mentioned theme of the present invention is only limitted to following embodiment, all to belong to this based on the technology that the content of present invention is realized
The range of invention.
The training data and test data that embodiment uses are THYME corpus disclosed in Semval-2017.
Embodiment 1:It is mainly used to extract the time serial message in clinical medicine text.Training data and test data are equal
Intestinal cancer text data in THYME corpus.
Step 1:Word segmentation processing is carried out to training text and test text first, the training text obtained after participle is used
BIO labels are marked.Wherein B represents the beginning that this word is an information sequence, and it is one information of composition that I, which represents this word,
The intermediate word of sequence, O represent this word and are not belonging to any information sequence.
Step 2:Its corresponding original character vector table is built for 24 English alphabets and other common characters, and with
Biomedical article in PubMed databases is the initial term vector of building of corpus.Text after being segmented based on step 1, is passed through
It tables look-up and obtains each corresponding initial term vector of word and the corresponding original character vector of each character.
Its detailed step is as follows:
Step 2.1:Its corresponding character vector, specific practice are initialized to existing all English characters using random number
Be for initialization vector per one-dimensional, all fromIn the range of generate at random one number carry out assignment,
Wherein dim is the dimension of character vector, and all original character vectors are gathered together and generate an original character vector table,
The size of dim is set herein as 30.If x is i-th of character of original character vector table, then the corresponding original character vectors of x are
{cpi1, cpi2..., cpidim}。
Step 2.2:For all characters in training text and test text, generated by searching for step 2.1 initial
Character vector table obtains its corresponding original character vector.
Step 2.3:Using GLoVe term vectors model method disclosed in Stamford, the biology in PubMed databases is chosen
Medical article generates initial word vector table for corpus, and the corresponding vector dimension of each word is 300 dimensions.
Step 2.4:For all words in training text and test text, generated by searching for step 2.3 initial
Term vector table obtains its corresponding initial term vector.
Step 3:Structure is carried based on the character vector that step 2 generates and the neural network medicine entity of term vector joint input
Modulus type.Model is divided into encoder, decoder and grader three parts, respectively using CNN networks and Bi-LSTM networks to word
Symbol vector and the input of term vector are encoded, and using Bi-LSTM network decodings, complete to classify using softmax graders.
Its detailed step is as follows:
Step 3.1:Using step 2.2 generate original character vector, by form each word character its it is corresponding just
The beginning character vector generation original character matrix that is stitched together is sent into convolutional neural networks and is encoded (as unit of each word).
The original character matrix of convolutional neural networks is input to for each, first passes around a convolutional layer, using convolution kernel by group
Original character vector into each word adjacent character carries out convolution.Then by the Input matrix of convolutional layer output a to maximum
Pond layer is directed to each row vector of convolutional layer output matrix, utilizes that one-dimensional generation of maximum pond layer choosing access value maximum
The information that the entire row vector of table includes, then after maximum pond layer, output one it is identical with original character vector dimension to
Amount.Then for each word i, corresponding character vector is ci={ c1, c2..., cdim}.By this step, can extract each
The morphologic feature of word.For such as " 2013-12-29 " etc. using character and number composition be difficult to extract its meaning of a word and
The temporal expression of syntactic feature can be very good to capture its character composition form by this step.
Step 3.2:The initial term vector generated using step 2.3, by the corresponding initial word of words all in each sentence
Vector is stitched together to be fed through in a Bi-LSTM and be encoded.It is included in wherein two-way LSTM there are two LSTM layers, one is
Forward direction LSTM, one is backward LSTM.T-th of word being then directed in a sentence is obtained to LSTM comprising the using preceding
One word to t-th word context information corresponding vector hft, i.e. the hidden layer of the corresponding forward direction LSTM of t-th of word is defeated
Go out.The corresponding vector h comprising t-th of word to a last word context information is obtained using backward LSTMbt, i.e., t-th
The hidden layer output of the corresponding backward LSTM of word.Two vectors are stitched together, the term vector h as t-th of wordt=(hft,
hbt).The syntax meaning of a word feature of a word context can be obtained using this step, then for shaped like " a few days " etc. by
The temporal expression of the phrase form of multiple word compositions, can be very good to extract.
Step 3.3:If the corresponding character vectors of each word i of CNN layers of output are ci={ c1, c2..., cdim, Bi-
The corresponding term vectors of encoding layers of each word i of output of LSTM are hi={ wh1, wh2..., whn, then normalizing is carried out to it
Change, that is, set cmaxThat maximum for character vector numerical value is one-dimensional, if whmaxFor word to numerical quantity it is maximum that is one-dimensional, then finally
Character vector be { c1/cmax, c2/cmax..., cdim/cmax, final term vector is { wh1/whmax, wh2/whmax..., whn/
whmax}.Two above vector is spliced to obtain the corresponding final vector m of each wordi={ c1/cmax, c2/cmax...,
cdim/cmax, wh1/whmax, wh2/whmax..., whn/whmax}.By the corresponding final vector cascade of words all in each sentence
Get up to form final vector matrix, Bi-LSTM networks are then input to as unit of sentence and are decoded.
Step 3.4:If the corresponding output vectors of the decoded each word i of Bi-LSTM are yi={ f1, f2..., ft, it will
yiBy final softmax layers, a three-dimensional vector { x is obtained1, x2, x3, it represents respectively and this word is classified as B, I, 0
Probability.Then that maximum one-dimensional corresponding label of numerical value is the final classification label of time word.It is just obtained by this step
The BIO label result final to each word.
Step 4:Training data after being marked using BIO trains above-mentioned model, using passing through the reality in comparative training data
The BIO labels on border obtained with this category of model after BIO labels difference, adjustment model parameter is with Optimum Classification performance.Experiment
The loss function used is log-likelihood loss function
Step 5:Using test data, to step 4, trained model is tested, and is obtained eventually by softmax graders
To BIO sequence labels extract medicine entity.
Embodiment 2:It is mainly used to extract the event information in clinical medicine text, such as craniotomy, bleed,
Cancer, in order to show the ability that there is the present invention field to adapt to, training text intestinal cancer textual data in THYME corpus
According to test text cancer of the brain text data in THYME databases.Training text be about intestinal cancer, test text be about
The cancer of the brain, so just needing to solve the problems, such as that field adapts to.Such as Chemotherapy (chemotherapy) is either in source domain also mesh
Mark field is all an event, and craniotomy (opening cranium art) is only an event in target domain.
Step 1:Training text and test text are carried out using Stamford natural language processing tool at participle first
Reason, the training text obtained after participle is marked with BIO labels.It is opening for sequence of events that wherein B, which represents this word,
Head, it is the intermediate word for forming a sequence of events that I, which represents this word, and O represents this word and is not belonging to any information sequence.
Step 2:Its corresponding original character vector table is built for 24 English alphabets and other common characters, and with
Biomedical article in PubMed databases is the initial term vector of building of corpus.Text after being segmented based on step 1, is passed through
It tables look-up and obtains each corresponding initial term vector of word and the corresponding original character vector of each character.
Its detailed step is as follows:
Step 2.1:Its corresponding character vector, specific practice are initialized to existing all English characters using random number
Be for initialization vector per one-dimensional, all fromIn the range of generate at random one number carry out assignment,
Wherein dim is the dimension of character vector, and all original character vectors are gathered together and generate an original character vector table,
The size of dim is set herein as 30.If x is i-th of character of original character vector table, then the corresponding original character vectors of x are
{cpi1, cpi2..., cpidim}。
Step 2.2:For all characters in training text and test text, generated by searching for step 2.1 initial
Character vector table obtains its corresponding original character vector.
Step 2.3:Using GLoVe term vectors model method disclosed in Stamford, the biology in PubMed databases is chosen
Medical article generates initial word vector table for corpus, and the corresponding vector dimension of each word is 300 dimensions.
Step 2.4:For all words in training text and test text, generated by searching for step 2.3 initial
Term vector table obtains its corresponding initial term vector.
Step 3:Structure is carried based on the character vector that step 2 generates and the neural network medicine entity of term vector joint input
Modulus type.Model is divided into encoder, decoder and grader three parts, respectively using CNN networks and Bi-LSTM networks to word
Symbol vector and the input of term vector are encoded, and using Bi-LSTM network decodings, complete to classify using softmax graders.
Its detailed step is as follows:
Step 3.1:Using step 2.2 generate original character vector, by form each word character its it is corresponding just
The beginning character vector generation original character matrix that is stitched together is sent into convolutional neural networks and is encoded (as unit of each word).
The original character matrix of convolutional neural networks is input to for each, a convolutional layer is first passed around, utilizes the convolution kernel of 3*3
The original character vector for forming each word adjacent character is subjected to convolution.Then by the Input matrix of convolutional layer output to one
Maximum pond layer is directed to each row vector of convolutional layer output matrix, utilizes that of maximum pond layer choosing access value maximum
Dimension represents the information that entire row vector includes, then after maximum pond layer, output one is identical with original character vector dimension
Vector.Then for each word i, corresponding character vector is ci={ c1, c2..., cdim}.By this step, can extract
The morphologic feature of each word.Due to often having many uncommon specialized vocabularies in medicine text, but they have
Certain word-building rule, such as with ant-, anti-is that the word of prefix has confrontation, the meaning cancelled, inhibited, antacid (systems
Sour agent), antibiotic (antibiotic), antieoagulant (anti-freezing) etc. can have been extracted more fully by this step
Information.
Step 3.2:The initial term vector generated using step 2.3, by the corresponding initial word of words all in each sentence
Vector is stitched together to be fed through in a Bi-LSTM and be encoded.It is included in wherein two-way LSTM there are two LSTM layers, one is
Forward direction LSTM, one is backward LSTM.T-th of word being then directed in a sentence is obtained to LSTM comprising the using preceding
One word to t-th word context information corresponding vector hft, i.e. the hidden layer of the corresponding forward direction LSTM of t-th of word is defeated
Go out.The corresponding vector h comprising t-th of word to a last word context information is obtained using backward LSTMbt, i.e., t-th
The hidden layer output of the corresponding backward LSTM of word.Two vectors are stitched together, the term vector h as t-th of wordt=(hft,
hbt), the syntax meaning of a word feature of a word context can be obtained using this step.
Step 3.3:If the corresponding character vectors of each word i of CNN layers of output are ci={ c1, c2..., cdim, Bi-
The corresponding term vectors of encoding layers of each word i of output of LSTM are hi={ wh1, wh2..., whn, then normalizing is carried out to it
Change, that is, set cmaxThat maximum for character vector numerical value is one-dimensional, if whmaxFor word to numerical quantity it is maximum that is one-dimensional, then finally
Character vector be { c1/cmax, c2/cmax..., cdim/cmax, final term vector is { wh1/whmax, wh2/whmax..., whn/
whmax}.Two above vector is spliced to obtain the corresponding final vector m of each wordi={ c1/cmax, c2/cmnax...,
cdim/cmax, wh1/whmax, wh2/whmax..., whn/whmax}.By the corresponding final vector cascade of words all in each sentence
Get up to form final vector matrix, Bi-LSTM networks are then input to as unit of sentence and are decoded.By using CNN and
Bi-LSTM automatically selects feature, avoids artificial design feature this process, can solve the problem of field adapts to.
Step 3.4:If the corresponding output vectors of the decoded each word i of Bi-LSTM are yi={ f1, f2..., ft, it will
yiBy final softmax layers, a three-dimensional vector { x is obtained1, x2, x3, it represents respectively and this word is classified as B, I, 0
Probability.Then that maximum one-dimensional corresponding label of numerical value is the final classification label of time word.It is just obtained by this step
The BIO label result final to each word.
Step 4:Training data after being marked using BIO trains above-mentioned model, using passing through the reality in comparative training data
The BIO labels on border obtained with this category of model after BIO labels difference, adjustment model parameter is with Optimum Classification performance.Experiment
The loss function used is log-likelihood loss function
Step 5:Using test data, to step 4, trained model is tested, and is obtained eventually by softmax graders
To BIO sequence labels extract medicine entity.
Claims (3)
1. a kind of clinical medicine information extracting method based on neural network, which is characterized in that include the following steps:
Step 1:Word segmentation processing is carried out to training text and test text first, the training text obtained after participle is marked with BIO
Label are marked;
Step 2:Its corresponding original character vector table is built, and with other common characters with PubMed for 24 English alphabets
Biomedical article in database is the initial term vector of building of corpus, and the text after being segmented based on step 1 is obtained by tabling look-up
Obtain each corresponding initial term vector of word and the corresponding original character vector of each character;
Step 3:Build the neural network medicine entity extraction mould based on the character vector that step 2 generates with term vector joint input
Type, model are divided into encoder, decoder and grader three parts, respectively using CNN networks and Bi-LSTM networks to character to
It measures the input with term vector to be encoded, using Bi-LSTM network decodings, completes to classify using softmax graders;
Step 4:Training data after being marked using BIO trains above-mentioned model, is marked by the practical BIO in comparative training data
The difference of the BIO labels after being obtained with this category of model is signed, adjustment model parameter is with Optimum Classification performance;
Step 5:Using test data, to step 4, trained model is tested, and is obtained eventually by softmax graders
BIO sequence labels extract medicine entity.
A kind of 2. clinical medicine information extracting method based on neural network according to claim 1, which is characterized in that institute
Step 2 is stated, is included the following steps:
Step 2.1:Its corresponding character vector is initialized to existing all English characters using random number, specific practice is needle
To initialization vector per one-dimensional, all fromIn the range of generate at random one number carry out assignment, wherein
Dim is the dimension of character vector, and all original character vectors are gathered together and generate an original character vector table, dim
Size between 30 to 50;
Step 2.2:For all characters in training text and test text, the original character generated by searching for step 2.1
Vector table obtains its corresponding original character vector;
Step 2.3:Using GLoVe term vectors model method disclosed in Stamford, the biomedicine in PubMed databases is chosen
Article generates initial word vector table for corpus;
Step 2.4:For all words in training text and test text, by searching for the initial word that step 2.3 generates to
Scale obtains its corresponding initial term vector.
A kind of 3. clinical medicine information extracting method based on neural network according to claim 1, which is characterized in that institute
Step 3 is stated, is included the following steps:
Step 3.1:The original character vector generated using step 2.2, will form the character of each word its corresponding initial word
The symbol vector generation original character matrix that is stitched together is sent into convolutional neural networks and is encoded (as unit of each word), for
Each is input to the original character matrix of convolutional neural networks, first passes around a convolutional layer, will be formed using convolution kernel every
The original character vector of a word adjacent character carries out convolution, then by the Input matrix of convolutional layer output to a maximum pond
Layer is directed to each row vector of convolutional layer output matrix, using maximum pond layer choosing access value maximum that it is one-dimensional represent it is whole
The information that a row vector includes then after maximum pond layer, exports a vector identical with original character vector dimension;
Step 3.2:The initial term vector generated using step 2.3, by the corresponding initial term vector of words all in each sentence
Be stitched together to be fed through in a Bi-LSTM and be encoded, wherein included in two-way LSTM there are two LSTM layers, one be it is preceding to
LSTM, one is backward LSTM, then t-th of word being directed in a sentence, is obtained using preceding to LSTM comprising first
Word to t-th word context information corresponding vector hft, obtained using backward LSTM comprising t-th of word to last one
The corresponding vector h of word context informationbt, vector is stitched together, the term vector h as t-th of wordt=(hft, hbt);
Step 3.3:If the corresponding character vectors of each word i of CNN layers of output are { c1, c2..., cdim, Bi-LSTM
The corresponding term vectors of the encoding layers of each word i of output are { wh1, wh2..., whn, then it is normalized, that is, sets cmax
That maximum for character vector numerical value is one-dimensional, if whmaxFor word to numerical quantity it is maximum that is one-dimensional, then final character vector
For { c1/cmax, c2/cmax..., cdim/cmax, final term vector is { wh1/whmax, wh2/whmax..., whn/whmax, by more than
Two vectors are spliced to obtain the corresponding final vector m of each wordi{c1/cmax, c2/cmax..., cdim/cmax, wh1/
whmax, wh2/whmax..., whn/whmax, the corresponding final vector of words all in each sentence is cascaded up, and it is final to be formed
Then vector matrix is input to Bi-LSTM networks as unit of sentence and is decoded;
Step 3.4:By the decoded output vectors of Bi-LSTM by final softmax layers, obtain final to each word
BIO marks result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711462492.6A CN108182976A (en) | 2017-12-28 | 2017-12-28 | A kind of clinical medicine information extracting method based on neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711462492.6A CN108182976A (en) | 2017-12-28 | 2017-12-28 | A kind of clinical medicine information extracting method based on neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108182976A true CN108182976A (en) | 2018-06-19 |
Family
ID=62548515
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711462492.6A Pending CN108182976A (en) | 2017-12-28 | 2017-12-28 | A kind of clinical medicine information extracting method based on neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108182976A (en) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109472194A (en) * | 2018-09-26 | 2019-03-15 | 重庆邮电大学 | A kind of Mental imagery EEG signals characteristic recognition method based on CBLSTM algorithm model |
CN109543183A (en) * | 2018-11-16 | 2019-03-29 | 西安交通大学 | Multi-tag entity-relation combined extraction method based on deep neural network and mark strategy |
CN109800411A (en) * | 2018-12-03 | 2019-05-24 | 哈尔滨工业大学(深圳) | Clinical treatment entity and its attribute extraction method |
CN109815478A (en) * | 2018-12-11 | 2019-05-28 | 北京大学 | Medicine entity recognition method and system based on convolutional neural networks |
CN110083838A (en) * | 2019-04-29 | 2019-08-02 | 西安交通大学 | Biomedical relation extraction method based on multilayer neural network Yu external knowledge library |
CN110222338A (en) * | 2019-05-28 | 2019-09-10 | 浙江邦盛科技有限公司 | A kind of mechanism name entity recognition method |
CN110222201A (en) * | 2019-06-26 | 2019-09-10 | 中国医学科学院医学信息研究所 | A kind of disease that calls for specialized treatment knowledge mapping construction method and device |
CN110867225A (en) * | 2019-11-04 | 2020-03-06 | 山东师范大学 | Character-level clinical concept extraction named entity recognition method and system |
CN111435410A (en) * | 2019-01-14 | 2020-07-21 | 阿里巴巴集团控股有限公司 | Relationship extraction method and device for medical texts |
CN111476022A (en) * | 2020-05-15 | 2020-07-31 | 湖南工商大学 | Method, system and medium for recognizing STM entity by embedding and mixing L characters of entity characteristics |
CN111553012A (en) * | 2020-04-28 | 2020-08-18 | 广东博智林机器人有限公司 | Home decoration design method and device, electronic equipment and storage medium |
CN112687332A (en) * | 2021-03-12 | 2021-04-20 | 北京贝瑞和康生物技术有限公司 | Method, apparatus and storage medium for determining sites of variation at risk of disease |
CN112687328A (en) * | 2021-03-12 | 2021-04-20 | 北京贝瑞和康生物技术有限公司 | Method, apparatus and medium for determining phenotypic information of clinical descriptive information |
CN112765319A (en) * | 2021-01-20 | 2021-05-07 | 中国电子信息产业集团有限公司第六研究所 | Text processing method and device, electronic equipment and storage medium |
CN113033155A (en) * | 2021-05-31 | 2021-06-25 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Automatic coding method for medical concepts by combining sequence generation and hierarchical word lists |
CN113297852A (en) * | 2021-07-26 | 2021-08-24 | 北京惠每云科技有限公司 | Medical entity word recognition method and device |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105955952A (en) * | 2016-05-03 | 2016-09-21 | 成都数联铭品科技有限公司 | Information extraction method based on bi-directional recurrent neural network |
CN106569998A (en) * | 2016-10-27 | 2017-04-19 | 浙江大学 | Text named entity recognition method based on Bi-LSTM, CNN and CRF |
CN106682220A (en) * | 2017-01-04 | 2017-05-17 | 华南理工大学 | Online traditional Chinese medicine text named entity identifying method based on deep learning |
CN107330032A (en) * | 2017-06-26 | 2017-11-07 | 北京理工大学 | A kind of implicit chapter relationship analysis method based on recurrent neural network |
-
2017
- 2017-12-28 CN CN201711462492.6A patent/CN108182976A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105955952A (en) * | 2016-05-03 | 2016-09-21 | 成都数联铭品科技有限公司 | Information extraction method based on bi-directional recurrent neural network |
CN106569998A (en) * | 2016-10-27 | 2017-04-19 | 浙江大学 | Text named entity recognition method based on Bi-LSTM, CNN and CRF |
CN106682220A (en) * | 2017-01-04 | 2017-05-17 | 华南理工大学 | Online traditional Chinese medicine text named entity identifying method based on deep learning |
CN107330032A (en) * | 2017-06-26 | 2017-11-07 | 北京理工大学 | A kind of implicit chapter relationship analysis method based on recurrent neural network |
Non-Patent Citations (1)
Title |
---|
XUEZHE MA等: "End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF", 《HTTP://ARXIV.ORG/PDF/1603.01354.PDF》 * |
Cited By (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109472194B (en) * | 2018-09-26 | 2022-02-11 | 重庆邮电大学 | Motor imagery electroencephalogram signal feature identification method based on CBLSTM algorithm model |
CN109472194A (en) * | 2018-09-26 | 2019-03-15 | 重庆邮电大学 | A kind of Mental imagery EEG signals characteristic recognition method based on CBLSTM algorithm model |
CN109543183A (en) * | 2018-11-16 | 2019-03-29 | 西安交通大学 | Multi-tag entity-relation combined extraction method based on deep neural network and mark strategy |
CN109800411B (en) * | 2018-12-03 | 2023-07-18 | 哈尔滨工业大学(深圳) | Clinical medical entity and attribute extraction method thereof |
CN109800411A (en) * | 2018-12-03 | 2019-05-24 | 哈尔滨工业大学(深圳) | Clinical treatment entity and its attribute extraction method |
CN109815478A (en) * | 2018-12-11 | 2019-05-28 | 北京大学 | Medicine entity recognition method and system based on convolutional neural networks |
CN111435410A (en) * | 2019-01-14 | 2020-07-21 | 阿里巴巴集团控股有限公司 | Relationship extraction method and device for medical texts |
CN111435410B (en) * | 2019-01-14 | 2023-04-14 | 阿里巴巴集团控股有限公司 | Relationship extraction method and device for medical texts |
CN110083838A (en) * | 2019-04-29 | 2019-08-02 | 西安交通大学 | Biomedical relation extraction method based on multilayer neural network Yu external knowledge library |
CN110222338A (en) * | 2019-05-28 | 2019-09-10 | 浙江邦盛科技有限公司 | A kind of mechanism name entity recognition method |
CN110222201A (en) * | 2019-06-26 | 2019-09-10 | 中国医学科学院医学信息研究所 | A kind of disease that calls for specialized treatment knowledge mapping construction method and device |
CN110222201B (en) * | 2019-06-26 | 2021-04-27 | 中国医学科学院医学信息研究所 | Method and device for constructing special disease knowledge graph |
CN110867225A (en) * | 2019-11-04 | 2020-03-06 | 山东师范大学 | Character-level clinical concept extraction named entity recognition method and system |
CN111553012A (en) * | 2020-04-28 | 2020-08-18 | 广东博智林机器人有限公司 | Home decoration design method and device, electronic equipment and storage medium |
CN111476022B (en) * | 2020-05-15 | 2023-07-07 | 湖南工商大学 | Character embedding and mixed LSTM entity identification method, system and medium for entity characteristics |
CN111476022A (en) * | 2020-05-15 | 2020-07-31 | 湖南工商大学 | Method, system and medium for recognizing STM entity by embedding and mixing L characters of entity characteristics |
CN112765319A (en) * | 2021-01-20 | 2021-05-07 | 中国电子信息产业集团有限公司第六研究所 | Text processing method and device, electronic equipment and storage medium |
CN112765319B (en) * | 2021-01-20 | 2021-09-03 | 中国电子信息产业集团有限公司第六研究所 | Text processing method and device, electronic equipment and storage medium |
CN112687328A (en) * | 2021-03-12 | 2021-04-20 | 北京贝瑞和康生物技术有限公司 | Method, apparatus and medium for determining phenotypic information of clinical descriptive information |
CN112687332A (en) * | 2021-03-12 | 2021-04-20 | 北京贝瑞和康生物技术有限公司 | Method, apparatus and storage medium for determining sites of variation at risk of disease |
CN113033155A (en) * | 2021-05-31 | 2021-06-25 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Automatic coding method for medical concepts by combining sequence generation and hierarchical word lists |
CN113297852A (en) * | 2021-07-26 | 2021-08-24 | 北京惠每云科技有限公司 | Medical entity word recognition method and device |
CN113297852B (en) * | 2021-07-26 | 2021-11-12 | 北京惠每云科技有限公司 | Medical entity word recognition method and device |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108182976A (en) | A kind of clinical medicine information extracting method based on neural network | |
CN105808525B (en) | A kind of field concept hyponymy abstracting method based on similar concept pair | |
CN104268160B (en) | A kind of OpinionTargetsExtraction Identification method based on domain lexicon and semantic role | |
CN108334495A (en) | Short text similarity calculating method and system | |
CN108846017A (en) | The end-to-end classification method of extensive newsletter archive based on Bi-GRU and word vector | |
CN106096664B (en) | A kind of sentiment analysis method based on social network data | |
CN108108355A (en) | Text emotion analysis method and system based on deep learning | |
CN108491372B (en) | Chinese word segmentation method based on seq2seq model | |
CN107122349A (en) | A kind of feature word of text extracting method based on word2vec LDA models | |
CN108804423B (en) | Medical text feature extraction and automatic matching method and system | |
CN103020034A (en) | Chinese words segmentation method and device | |
CN111243699A (en) | Chinese electronic medical record entity extraction method based on word information fusion | |
CN108509409A (en) | A method of automatically generating semantic similarity sentence sample | |
CN112966525B (en) | Law field event extraction method based on pre-training model and convolutional neural network algorithm | |
CN105446955A (en) | Adaptive word segmentation method | |
CN108710611A (en) | A kind of short text topic model generation method of word-based network and term vector | |
CN107357785A (en) | Theme feature word abstracting method and system, feeling polarities determination methods and system | |
CN110717330A (en) | Word-sentence level short text classification method based on deep learning | |
CN111914555B (en) | Automatic relation extraction system based on Transformer structure | |
CN106528776A (en) | Text classification method and device | |
CN108427717A (en) | It is a kind of based on the alphabetic class family of languages medical treatment text Relation extraction method gradually extended | |
CN110222338A (en) | A kind of mechanism name entity recognition method | |
CN108763192A (en) | Entity relation extraction method and device for text-processing | |
CN105159917A (en) | Generalization method for converting unstructured information of electronic medical record to structured information | |
CN115510864A (en) | Chinese crop disease and pest named entity recognition method fused with domain dictionary |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180619 |