CN109710925A - Name entity recognition method and device - Google Patents
Name entity recognition method and device Download PDFInfo
- Publication number
- CN109710925A CN109710925A CN201811518491.3A CN201811518491A CN109710925A CN 109710925 A CN109710925 A CN 109710925A CN 201811518491 A CN201811518491 A CN 201811518491A CN 109710925 A CN109710925 A CN 109710925A
- Authority
- CN
- China
- Prior art keywords
- model
- character
- training
- crf
- participle
- 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
- 238000000034 method Methods 0.000 title claims abstract description 70
- 238000012549 training Methods 0.000 claims abstract description 139
- 230000011218 segmentation Effects 0.000 claims description 22
- 239000011159 matrix material Substances 0.000 claims description 21
- 238000012545 processing Methods 0.000 claims description 9
- 239000000203 mixture Substances 0.000 claims description 8
- 238000002372 labelling Methods 0.000 claims description 6
- 238000006073 displacement reaction Methods 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 5
- 235000013399 edible fruits Nutrition 0.000 claims description 4
- 201000010099 disease Diseases 0.000 description 27
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 27
- 230000006870 function Effects 0.000 description 12
- 210000000232 gallbladder Anatomy 0.000 description 12
- 210000001015 abdomen Anatomy 0.000 description 10
- 238000010586 diagram Methods 0.000 description 8
- 208000024891 symptom Diseases 0.000 description 6
- 230000004927 fusion Effects 0.000 description 5
- 238000002192 cholecystectomy Methods 0.000 description 4
- 238000003745 diagnosis Methods 0.000 description 4
- 238000002271 resection Methods 0.000 description 4
- 238000002555 auscultation Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 206010000077 Abdominal mass Diseases 0.000 description 2
- 206010020772 Hypertension Diseases 0.000 description 2
- 230000002612 cardiopulmonary effect Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 206010012601 diabetes mellitus Diseases 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 210000004072 lung Anatomy 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 150000001875 compounds Chemical class 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000005389 magnetism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000877 morphologic effect Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
This disclosure relates to a kind of name entity recognition method and device, including text is labeled according to default mark collection to obtain training corpus, the notation methods of the default type of the mark collection comprising name entity and notation methods and non-name entity；According to the training corpus, the preset multiple identification models of training；Text to be identified is identified using the multiple identification model respectively, obtains the recognition result of each identification model；The recognition result of weight and each identification model based on each identification model, obtains the recognition result of the text to be identified.Text to be identified is identified respectively by using multiple identification models, obtains the recognition result of each identification model, and the precision of name Entity recognition can be effectively able to ascend according to the name entity recognition method and device of the embodiment of the present disclosure.
Description
Technical field
This disclosure relates to pattern-recognition and sorting technique field more particularly to a kind of name entity recognition method and device.
Background technique
Name Entity recognition be intended to identify proper noun in natural language text and significant numeral classifier phrase (refer to by
The phrase that number and quantifier combination are constituted), and classified.Name Entity recognition belongs to unknown word identification in morphological analysis
Scope is essential group of a variety of natural language processing techniques such as information extraction, information retrieval, machine translation and question answering system
At part.
In the related technology, the method for naming Entity recognition is divided into: rule-based and dictionary method, Statistics-Based Method
With the method etc. based on deep learning.Rule-based and dictionary method depends on concrete syntax, field and text style, is easy
Mistake is generated, it is portable poor.Statistics-Based Method also compares the dependence of corpus the more demanding of Feature Selection
Greatly.There are problems that gradient disappearance in training based on the method for deep learning.These methods are difficult to meet to a large amount of at present random
Then, the identification of multi-field name entity.
Summary of the invention
In view of this, being able to ascend name Entity recognition the present disclosure proposes a kind of name entity recognition method and device
Precision.
According to the one side of the disclosure, a kind of name entity recognition method is provided, which comprises according to pre- bidding
Note collection is labeled to obtain training corpus for text, the default type of the mark collection comprising name entity and notation methods,
And the notation methods of non-name entity；According to the training corpus, the preset multiple identification models of training；Using the multiple
Identification model identifies text to be identified respectively, obtains the recognition result of each identification model；Weight based on each identification model with
And the recognition result of each identification model, obtain the recognition result of the text to be identified.
According to another aspect of the present disclosure, a kind of name entity recognition device is provided, described device includes: mark mould
Block obtains training corpus for being labeled according to default mark collection for text, and the default mark collection includes name entity
Type and notation methods and it is non-name entity notation methods；Training module, for according to the training corpus, training
Preset multiple identification models；Identification module obtains every for identifying text to be identified respectively using the multiple identification model
The recognition result of a identification model；Weighting block, the identification knot for weight and each identification model based on each identification model
Fruit obtains the recognition result of the text to be identified.
In the embodiments of the present disclosure, text is labeled to obtain training corpus according to default mark collection, according to training
Corpus, the preset multiple identification models of training, identifies text to be identified using multiple identification models respectively, obtains each identification mould
The recognition result of type combines the advantage of different identification models by the recognition result of the multiple identification models of Weighted Fusion, is promoted
The precision of name Entity recognition.
According to below with reference to the accompanying drawings to detailed description of illustrative embodiments, the other feature and aspect of the disclosure be will be apparent.
Detailed description of the invention
Comprising in the description and constituting the attached drawing of part of specification and specification together illustrates the disclosure
Exemplary embodiment, feature and aspect, and for explaining the principles of this disclosure.
Fig. 1 shows the flow chart of the name entity recognition method according to one embodiment of the disclosure.
Fig. 2 shows the identification model fusion architecture schematic diagrames according to the embodiment of the present disclosure.
Fig. 3 shows the flow chart of the name entity recognition method according to one embodiment of the disclosure.
Fig. 4 shows the flow chart of the step S12 according to one embodiment of the disclosure.
Fig. 5 shows the flow chart of the name entity recognition method according to one embodiment of the disclosure.
Fig. 6 shows the flow chart of the step S12 according to one embodiment of the disclosure.
Fig. 7 shows the block diagram of the name entity recognition device according to one embodiment of the disclosure.
Fig. 8 shows the block diagram of the name entity recognition device according to one embodiment of the disclosure.
Fig. 9 is shown according to an exemplary embodiment a kind of for naming the block diagram of entity recognition device 900.
Specific embodiment
Various exemplary embodiments, feature and the aspect of the disclosure are described in detail below with reference to attached drawing.It is identical in attached drawing
Appended drawing reference indicate element functionally identical or similar.Although the various aspects of embodiment are shown in the attached drawings, remove
It non-specifically points out, it is not necessary to attached drawing drawn to scale.
Dedicated word " exemplary " means " being used as example, embodiment or illustrative " herein.Here as " exemplary "
Illustrated any embodiment should not necessarily be construed as preferred or advantageous over other embodiments.
In addition, giving numerous details in specific embodiment below to better illustrate the disclosure.
It will be appreciated by those skilled in the art that without certain details, the disclosure equally be can be implemented.In some instances, for
Method, means, element and circuit well known to those skilled in the art are not described in detail, in order to highlight the purport of the disclosure.
Fig. 1 shows the flow chart of the name entity recognition method according to one embodiment of the disclosure.As shown in Figure 1, this method
Can include:
Step S11 is labeled text according to default mark collection to obtain training corpus, and the default mark collection includes
Name type, notation methods and the non-notation methods for naming entity of entity.
Step S12, according to the training corpus, the preset multiple identification models of training.
Step S13 identifies text to be identified using the multiple identification model respectively, obtains the identification of each identification model
As a result.
Step S14, the recognition result of weight and each identification model based on each identification model, obtains the text to be identified
This recognition result.
In the embodiments of the present disclosure, by the recognition result of the multiple identification models of Weighted Fusion, different identification moulds are combined
The advantage of type improves the precision of name Entity recognition.
Before model is trained and is tested, the text used when needing to training and test is labeled, and is instructed
Practice corpus.Default mark collection can be used for being labeled text.Default mark collection may include the type and mark of name entity
The notation methods of note mode and non-name entity.Wherein, the type of entity is named to can be used for identifying that different names is real
Body.The notation methods of name entity can serve to indicate that how to mark name entity all types of in text.Non- name entity is real
The notation methods of body can serve to indicate that the non-name entity how marked in text.
It in the embodiment of the present disclosure, is illustrated by taking medical electronics medical record data as an example, the name entity of the embodiment of the present disclosure
Recognition methods can be also used for the identification of other data.
It is illustrated for using medical electronics medical record data as text.By analyzing medical electronics medical record data, doctor
Diagnosis and treatment activity for patient may be summarized to be: (what do to check) performance (what disease of discovery disease by detection methods
Shape), it provides diagnosis (what disease), and be based on diagnosis, provides remedy measures (how treating).During from this
As can be seen that curative activity relates generally to 4 class important informations: inspection, symptom, disease and treatment.On this basis, mark is preset
Concentrate the type that may include 5 name entities, comprising: disease type, symptom type, checks class at medical diagnosis on disease classification type
Type and treatment type.
The notation methods for including according to default mark collection: the name entity of disease type can be labeled as disease；Disease
The name entity of diagnostic classification type can be labeled as disease_type；The name entity of symptom type can be labeled as
symptom；The name entity of inspect-type can be labeled as test；The name entity for treating type can be labeled as
treatment。
Table 1 shows the mark collection of the embodiment of the present disclosure and the example of training corpus.In one example, as shown in table 1,
Disease can be labeled as " hypertension, the diabetes " in text " old female suffers from, and denies hypertension, diabetic history ".
Default mark collection further comprises the notation methods of non-name entity, for example, non-name entity can be labeled as O.Its
In, non-name entity can indicate the part in corpus in addition to naming entity.
Table 1
In one possible implementation, the notation methods of entity is named to also can indicate that the first character of name entity
(or the last one divides for symbol (either first participle), intermediate character (either intermediate participle) and last character
Word).As shown in table 1, in one example, name the first character, intermediate character and last character of entity can be with
It is respectively labeled as B, I and E.For example, the first character of the name entity of disease type can be labeled as disease-B,
The character of the centre of the name entity of disease type can be labeled as disease-I, the name entity of disease type
Last character can be labeled as disease-E.
Can be according to the training corpus marked, the preset multiple identification models of training.Each identification model can be used for
Identification name entity.
In one possible implementation, preset multiple identification models may include CRF model and/or BILSTM_
CRF model and/or IDCNN_CRF model.Wherein, CRF model is probabilistic model, in the embodiments of the present disclosure, the spy of CRF model
Sign includes the part of speech of analysis, and the contextual information being made of the context of the context and part of speech that segment；BILSTM_CRF
Model and IDCNN_CRF model are deep learning model, in the embodiments of the present disclosure, BILSTM_CRF model and IDCNN_CRF
The feature vector of model is the character feature vector that the word vector sum Character segmentation vector of character forms.It should be noted that this
Multiple identification models of open embodiment can also be other disaggregated models, with no restrictions to this disclosure.
Text to be identified can indicate the text for needing to identify name entity.In the instruction for completing multiple preset identification models
After white silk, server can identify the name entity in text to be identified using the identification model that training is completed, and provide and identify
Name entity recognition result.
Fig. 2 shows the identification model fusion architecture schematic diagrames according to the embodiment of the present disclosure.As shown in Fig. 2, can will be wait know
Other text inputs each identification model respectively, obtains the recognition result of each identification model, according to the weight of each identification model, to each
The recognition result of identification model is weighted summation, and final recognition result is determined according to summed result.
Wherein, the recognition result obtained by identification model is that each character of text to be identified is identified as naming entity
Probability.
In one possible implementation, when summed result is greater than specified threshold, determination belongs to name entity, when asking
When being less than or equal to specified threshold with result, determines and be not belonging to name entity.Wherein, specified threshold can according to need progress
It determines, such as can be 0.5.
In an example it is assumed that multiple identification models include model 1, model 2 and model 3, weight is respectively 0.1,0.3
With 0.7, specified threshold 0.5.After text input model 1 to be identified, model 2 and model 3, the to be identified of the output of model 1 is obtained
Recognition result (i.e. each character is identified as naming the probability of entity), the model 2 of each character export to be identified in text
The recognition result of each character in the text to be identified that the recognition result of each character and model 3 export in text.Wait know
For the first character of other text, it is assumed that it is 0.7 that the first character that model 1 exports, which is identified as naming the probability of entity,
It is 0.8 that the first character that model 2 exports, which is identified as naming the probability of entity, and the first character that model 3 exports is identified
Probability to name entity is 0.2, is weighted summation to these three probability, available summed result is 0.45, the summation
As a result it is less than specified threshold 0.5, accordingly, it is determined that the first character of text to be identified is not belonging to name entity.For to be identified
Other characters of text are referred to first character, and which is not described herein again.
It, can be with Loss (loss) minimum optimization aim of fused model, in reality in a kind of possible realization
During existing Loss value is the smallest, continuous iteration finally searches out optimal weight combination, so that it is determined that each identification model out
Weight.
By taking identification model is CRF model as an example.Fig. 3 shows the name entity recognition method according to the disclosure one embodiment
Flow chart.As shown in figure 3, step S11 is labeled text according to default mark collection and is instructed according to the training corpus
Practicing corpus may include step S111 and step S112.
Step S111 carries out word segmentation processing to the text, obtains multiple participles.
The participle for belonging to name entity is labeled as phase according to the notation methods that the default mark collection includes by step S112
The type answered, and the participle for belonging to non-name entity is labeled, obtain training corpus.
For each participle, it can determine whether the participle is noted as name entity in the text；If the participle is marked
Note is name entity, then can determine that the participle belongs to name entity, can be true if the participle is noted as non-name entity
The fixed participle belongs to non-name entity.
When participle belongs to name entity, the type of the corresponding name entity of the participle can be determined, according to default mark
The notation methods of the name entity of the type that collection includes, mark the participle.
When participle is not belonging to name entity, the mark side for the non-name entity for including can be concentrated according to default mark
Formula marks the participle.For example, being labeled as O.
In one example, the name entity in text " carried out gallbladder before 4 years and assist resection " is known, " gallbladder removal
Art " is the name entity for treating type, collects " cholecystectomy " according to default mark and is noted as treatment, forms training
Corpus.Word segmentation processing is carried out to training corpus, obtains following participle: " 4 ", " year ", " preceding ", " progress ", " gall-bladder " and " excision
Art ".Participle is matched with the name entity " cholecystectomy " for being noted as treatment, is obtained: segment " gall-bladder " and
" resection " belongs to the name entity of disease type, and participle " gall-bladder " is first participle of the name entity, participle " excision
Art " is the last one participle of the name entity.Therefore, " gall-bladder " can be labeled as to treatment-B, " resection " is marked
Note is treatment-E.It can be these participle addition O marks, to all points in text for other participles in text
After the completion of word mark, available training corpus.
In one possible implementation, according to the training corpus, the preset multiple identification models of training include: root
According to the training corpus, the training CRF model.On this basis, Fig. 4 shows the step S12 according to one embodiment of the disclosure
Flow chart.As shown in figure 4, step S12 can include:
Step S121 obtains the part of speech feature of each participle in the training corpus.
The training corpus tissue is matrix form, wherein first is classified as participle, middle column in matrix by step S122
For the part of speech feature of each participle, last is classified as the annotation results of each participle, and respectively segments in first row according to each participle
Position in the training corpus is ranked up.
Step S123 determines that the CRF template file that training process uses, the CRF template file are made of multiple template,
Each template is for specifying when extracting contextual information for current participle, and extracted contextual information is relative to current point
The absolute position of the line displacement of word and extracted contextual information column, the contextual information include the upper of participle
The contextual feature of following traits and part of speech.
Step S124 generates CRF model according to the matrix and the CRF template file.
The part of speech feature of participle can be used to indicate that the part of speech of participle.By analyzing medical electronics medical record data, discovery is very
More disease names are made of the connection of multiple nouns, i.e., the name entity of disease type is usually the participle of noun by multiple parts of speech
Composition.Referring to " the Chinese electronic health record standard of word segmentation ", part of speech mainly includes noun (n), verb (v), preposition (p), numeral-classifier compound
(m), symbol (x) and adverbial word (d) etc..
Table 2 shows the example of part of speech feature.As shown in table 2, the name entity " abdominal mass " of disease type is by two words
Property for noun participle " abdomen " and " mass " form.It should be noted that nr indicates proper noun, and proper noun can in table 2
Using as the specified participle in contextual feature.
Table 2
| Word | Part of speech | Mark |
| Cause | p | O |
| It was found that | v | O |
| Abdomen | n | disease-B |
| Mass | n | disease-E |
| 1 | m | O |
| Year | m | O |
| It is remaining | m | O |
| It is admitted to hospital | n | O |
| 。 | x | O |
| Physical examination | nr | O |
| : | x | O |
| Cardiopulmonary | n | test-B |
| Auscultation | v | test-E |
| Nothing | v | O |
| It is abnormal | d | O |
It can be matrix form by training corpus tissue, wherein first be classified as participle in matrix, centre is classified as each participle
Part of speech feature, last is classified as the annotation results of each participle, and respectively segmented in first row according to it is each participle in the training
Position in corpus is ranked up.As shown in table 2, first row is that text " is admitted to hospital for abdominal mass more than 1 year because of discovery.Physical examination: cardiopulmonary
Auscultate without exception " each participle for including, secondary series is the part of speech feature of each participle, and third column are the mark knots of each participle
Fruit.Therefore, table 2 can indicate the matrix that training corpus is organized into, wherein first row shown in table 2 can be used as the of matrix
One arranges, and secondary series shown in table 2 can be used as the secondary series of matrix, and the column of third shown in table 2 can be used as the third column of matrix.
Wherein, contextual information includes the contextual feature of participle and the contextual feature of part of speech.
The contextual feature of the part of speech of participle indicates the feature of the context of participle.For example, the symptom of disease generally occurs within
After verb, i.e. the name entity of symptom type typically occurs in after the participle that part of speech is verb.For example, such as table 2
Shown, the name entity " heart and lung auscultation " of inspect-type appears in after the participle " physical examination " that part of speech is verb.
The contextual feature of participle can be used to indicate that the feature for the participle that the context of participle includes.The context of participle
Feature can be generated by the feature templates used when training CRF model.It can be in the feature templates used in training CRF model
It is specified participle comprising the participle before or after specified participle.For example, " physical examination: " can be used as the upper of " heart and lung auscultation "
Following traits, " physical examination: " back very maximum probability will appear the name entity of inspect-type.The character modules used in CRF model
May include in plate connected participle " physical examination " and segment ": " after participle be inspect-type name entity.
CRF template file is made of multiple template, and each template extracts context for current participle for specifying how
Information.In CRF algorithm, when extracting contextual information for current participle, extracted contextual information is relative to current
The absolute position of the line displacement of participle and extracted contextual information column.
Contextual information can be expressed as, wherein the initial position of row and column is all 0.Table 3 shows CRF template file
One example.Every a line in CRF template file shown in table 3 is a template, and each template is come by %x [row, col]
The specified contextual information extracted.Wherein, row indicates that the relative displacement of row, col indicate the absolute position of column.Wherein, row
When for negative, biased forwards being indicated, indicating to deviate backward when being positive number, current participle is indicated when being 0.For example, referring to table 2 and table
3, it is illustrated so that current participle is " abdomen " as an example.In table 3 %x [- 2,0] be in table 2 row it is opposite with " abdomen " offset be-
2 row is classified as the content of the 0th column, i.e. participle " because ".%x [0,1] is to go to be in table 2 relative to the offset of " abdomen " in table 3
0 row is classified as the content of the 1st column, that is, segments the part of speech feature " n " of " abdomen ".Table 4 is shown, when current participle is " abdomen ", table
The corresponding contextual information of each template shown in 3.
Table 3
| #Unigram |
| U01:%x [- 2,0] |
| U02:%x [- 1,0] |
| U03:%x [0,0] |
| U04:%x [1,0] |
| U05:%x [2,0] |
| U06:%x [- 1,0]/%x [0,0] |
| U07:%x [0,0]/%x [1,0] |
| U08:%x [- 1,0]/%x [0,0]/%x [1,0] |
| U09:%x [- 2,0]/%x [- 1,0]/%x [0,0] |
| U10:%x [0,0]/%x [1,0]/%x [2,0] |
| U11:%x [0,1] |
| U12:%x [0,0]/%x [0,1] |
| #Bigram |
| B |
Table 4
Corresponding each template, CRF algorithm can generate a series of characteristic function set, the feelings to response training corpus
Condition, and then generate corresponding CRF model.Characteristic function includes transfer function and function of state in CRF algorithm, wherein transfer function
To include the characteristic function before and after current location, function of state be include characteristic function on current location.In general, feature
Function value is 1 or 1.
In one example, table 1 shows 21 kinds of annotation results, according to template U01%x [- 2,0] available following shape
State characteristic function:
1 else of Func1=if (output=' disease-B ' and feature=' U01: because ') return
return 0
1 else of Func2=if (output=' disease-I ' and feature=' U01: because ') return
return 0
1 else of Func3=if (output=' disease-E ' and feature=' U01: because ') return
return 0
……
1 else return 0 of Func21=if (output=' O ' and feature=' U01: because ') return
Wherein, output is the observation label of current participle " abdomen ", and each template can be defeated all possible label
It all lists out, the weight of each characteristic function is then determined by training.Time that reasonable label occurs in training sample
Number is more, and the weight of corresponding characteristic function is higher, otherwise fewer.
According to the matrix and the CRF template file, CRF model can be generated.Matrix and CRF template file are inputted
After CRF algorithm, CRF algorithm can train automatically and obtain CRF model.
In one example, order " crf_learn template train.txt model " training CRF can be passed through
Model.Wherein, template representing matrix, train.txt indicate CRF template file.The order can generate model after having executed
File, the model file are CRF model.
By taking BILSTM_CRF model or IDCNN_CRF model as an example.Fig. 5 shows the name according to one embodiment of the disclosure
The flow chart of entity recognition method.As shown in figure 5, step S11 collects according to default mark for text according to the training corpus
It is labeled to obtain training corpus to may include step S113 and step S114.
The text is divided into multiple characters by step S113.
Step S114, the notation methods for including are integrated according to the default mark will belong to the character label of name entity as phase
The type answered, and the character for belonging to non-name entity is labeled, obtain training corpus.
For each character, it can determine whether the character is noted as name entity in training corpus；If the character
It is noted as name entity, then can determine that the character belongs to name entity；If the character is noted as non-name entity, can
To determine that the character belongs to non-name entity.
When character belongs to name entity, the type of the corresponding name entity of the character can be determined, according to default mark
The notation methods of the name entity of the type that collection includes, mark the character.
When character is not belonging to name entity, the mark side for the non-name entity for including can be concentrated according to default mark
Formula marks the character.For example, being labeled as O.
In one example, the name entity in text " carried out gallbladder before 4 years and assist resection " is known, " gallbladder removal
Art " is the name entity for treating type, collects " cholecystectomy " according to default mark and is noted as treatment, forms training
Corpus.Training corpus is split as character, obtains character " 4 ", " year ", " preceding ", " into ", " row ", " gallbladder ", " assisting ", " cutting ", " removing "
" art ".Character is matched with the name entity " cholecystectomy " for being noted as treatment, wherein character " gallbladder ",
" assisting ", " cutting ", " removing " and " art " belongs to the name entity of disease type, and character " gallbladder " is the first character of the name entity,
Character " assisting ", " cutting " and " removing " is the character of the centre of the name entity, and character " art " is the last character of the name entity
Symbol.Therefore, it can be labeled as treatment-B for character " gallbladder ", character " assisting ", " cutting " and " removing " is respectively labeled as
Treatment-I, character " art " are labeled as treatment-E.For other participles in training corpus, server can be this
A little participles are labeled as O, after the completion of character labels all in text, available training corpus.
In one possible implementation, according to the training corpus, the preset multiple identification models of training include: root
According to the training corpus, the training BILSTM_CRF model or the IDCNN_CRF model.On this basis, Fig. 6 is shown
According to the flow chart of the step S12 of one embodiment of the disclosure.As shown in fig. 6, step S12 can include:
Step S125 obtains the character feature vector of each character in the training corpus, the character feature vector packet
Include word vector sum Character segmentation vector.
Step S126 is tied according to the corresponding mark of each character of character feature vector sum of character each in the training corpus
Fruit, the training BILSTM_CRF model or the IDCNN_CRF model.
Character feature vector can indicate the feature vector of character, and character feature vector may include word vector sum Character segmentation
Vector.
Wherein, word vector is the vector that can indicate the feature of character, and word vector can represent one per one-dimensional value
With feature that is certain semantic and grammatically explaining.Wherein, feature can be used for the fundamental of word (such as radical, radical,
Stroke, meaning etc.) the various information that are characterized.Can according to external word vector table (such as pass through word2vec training complete
Word vector table) to each character carry out id coding, obtain the word id of each character.It, can be outer using word id matching when training pattern
Portion's word vector table, obtains word vector.
Character segmentation vector can be used for identifying the character for belonging to the same participle.Training corpus is being split into character
In the process, numeric coding, such as " Kazakhstan " " that " " shore " three character categories can be carried out to the character for belonging to the same name entity
In a name entity, numeric coding can be expressed as " 1 " " 2 " " 3 ", and " north " " capital " two characters belong to a name entity,
Numeric coding can be expressed as " 1 " " 3 ", and " Hubei Province " " that " " more " " this " four characters belong to a name entity, and numeric coding can
To be expressed as " 1 " " 2 " " 2 " " 3 ".During training corpus is split into character, to the character for being not belonging to name entity, number
Value coding can be labeled as 0.For example, the numeric coding of " I is Harbin people " can be expressed as " 001230 ".It can be right
After each character code of training corpus, for each character, random initializtion can be carried out to the character, generate the first of the character
Beginning vector.It in training pattern, is continued to optimize by back-propagation algorithm, finally trains an optimal character for the character
Split vector.
Table 5 shows the example of word id and numeric coding.Table 6 shows the word vector sum character point obtained on the basis of table 5
Cut the example of vector.Wherein, the dimension of word vector is 100, and the dimension of Character segmentation vector is 20.
Table 5
| Training corpus | [because sending out, existing, abdomen, portion, packet, block ... ...] |
| Word id | [230,16,511,14,1052,363 ... ...] |
| Numeric coding | [0,0,0,1,2,2,3 ... ...] |
Table 6
| Training corpus | Word vector | Character segmentation vector |
| Cause | [0.5,0.6 ... 0.25,0.5]1*100 | [0.5,0.6 ... 0.25]1*20 |
| Hair | [0.7,0.8 ... 0.32,0.3]1*100 | [0.5,0.6 ... 0.25]1*20 |
| It is existing | [0.9,0.5 ... 0.72,0.8]1*100 | [0.5,0.6 ... 0.25]1*20 |
| Abdomen | [0.9,0.8 ... 0.75,0.9]1*100 | [0.9,0.8 ... 0.75]1*20 |
| Portion | [0.2,0.6 ... 0.22,0.6]1*100 | [0.2,0.6 ... 0.22]1*20 |
| Packet | [0.5,0.3 ... 0.52,0.7]1*100 | [0.2,0.6 ... 0.22]1*20 |
| Block | [0.4,0.7 ... 0.23,0.9]1*100 | [0.4,0.7 ... 0.23]1*20 |
| …… | …… | … |
According to the corresponding annotation results of each character of character feature vector sum of character each in the training corpus, described in training
BILSTM_CRF model or the IDCNN_CRF model.Wherein, the character feature vector of each character can be used as
The input of BILSTM_CRF model or IDCNN_CRF model, the annotation results of each character can be used for supervising BILSTM_CRF
The output result of model and IDCNN_CRF model.
With the multiple identification model include CRF model and/or BILSTM_CRF model and/or IDCNN_CRF model is
Example.
Wherein, text to be identified is identified using the CRF model, obtains recognition result can include: to the text to be identified
This progress word segmentation processing obtains multiple participles；Determine the part of speech feature of each participle；The multiple of the part of speech feature will be carried
The segmentation sequence of participle composition inputs in the CRF model, obtains the recognition result of the CRF model.
Text to be identified is identified using the BILSTM_CRF model or the IDCNN_CRF model, obtains recognition result
Can include: the text to be identified is split into character, obtains multiple characters；Determine the character feature vector of each character；It will
The sequence of the character feature vector composition of multiple characters is inputted respectively in BILSTM_CRF model or IDCNN_CRF model, is obtained
The recognition result of BILSTM_CRF model or the recognition result of IDCNN_CRF model.
Text to be identified can be segmented, and determine the part of speech feature of each participle, point of part of speech feature will be carried
The segmentation sequence of word composition is input in the CRF model of training completion, obtains the recognition result of CRF model.
The text to be identified can be split into character, the character feature vector of each character be obtained, by each character
The sequence of character feature vector composition is separately input in the BILSTM_CRF model and IDCNN_CRF model of training completion, is obtained
To the recognition result of BILSTM_CRF model and the recognition result of IDCNN_CRF model.
Table 7 shows an example of recognition result.As shown in table 6, for some character in text to be identified, three kinds of moulds
The probability that type determines that the character belongs to name entity is respectively 0.7,0.8 and 0.2, with 0.5 for for classification thresholds, three kinds of models
Predict that classification is respectively 1,1,0.Wherein, it 1 indicates, which belongs to name entity, and 0 indicates that the character is not belonging to name entity.
After being weighted, the probability for determining that the character belongs to name entity is 0.45, can determine that final recognition result is 0, the i.e. word
Symbol is not belonging to name entity.
Table 7
Fig. 7 shows the block diagram of the name entity recognition device according to one embodiment of the disclosure.As shown in fig. 7, the device 30
May include:
Labeling module 31, for being labeled to obtain training corpus, the pre- bidding for text according to default mark collection
The notation methods of type of the note collection comprising name entity and notation methods and non-name entity；
Training module 32, for according to the training corpus, the preset multiple identification models of training；
Identification module 33 obtains each identification mould for identifying text to be identified respectively using the multiple identification model
The recognition result of type；
Weighting block 34 obtains described for the recognition result of weight and each identification model based on each identification model
The recognition result of text to be identified.
In the embodiments of the present disclosure, by the recognition result of the multiple identification models of Weighted Fusion, different identification moulds are combined
The advantage of type improves the precision of name Entity recognition.
In one possible implementation, the labeling module 31 is particularly used in:
When the identification model is CRF model, word segmentation processing is carried out to the text, obtains multiple participles；
The participle for belonging to name entity is labeled as corresponding type according to the notation methods that the default mark collection includes,
And the participle for belonging to non-name entity is labeled, obtain training corpus.
Fig. 8 shows the block diagram of the name entity recognition device according to one embodiment of the disclosure.As shown in figure 8, one kind can
In the implementation of energy, the training module 32 can include:
First training submodule 321, for according to the training corpus, the training CRF model；
The first training submodule 321 is particularly used in:
Obtain the part of speech feature of each participle in the training corpus；
It is matrix form by the training corpus tissue, wherein first be classified as participle in matrix, centre is classified as each participle
Part of speech feature, last is classified as the annotation results of each participle, and respectively segmented in first row according to it is each participle in the training
Position in corpus is ranked up；
Determine that the CRF template file that training process uses, the CRF template file are made of multiple template, each template
For specifying when extracting contextual information for current participle, extracted contextual information is inclined relative to the row currently segmented
It moves and the absolute position of extracted contextual information column, the contextual information includes the contextual feature of participle
With the contextual feature of part of speech；
According to the matrix and the CRF template file, CRF model is generated.
In one possible implementation, the labeling module 31 is particularly used in:
When the identification model is BILSTM_CRF model or IDCNN_CRF model, the text is divided into multiple
Character；
The character label for belonging to name entity is corresponding type by the notation methods for including according to the default mark collection,
And the character for belonging to non-name entity is labeled, obtain training corpus.
In one possible implementation, the training module 32 can include:
Second training submodule 322, for according to the training corpus, the training BILSTM_CRF model or described
IDCNN_CRF model；
The second training submodule 322 is particularly used in:
The character feature vector of each character in the training corpus is obtained, the character feature vector includes word vector sum
Character segmentation vector；
According to the corresponding annotation results of each character of character feature vector sum of character each in the training corpus, described in training
BILSTM_CRF model or the IDCNN_CRF model.
In one possible implementation, the multiple identification model includes CRF model and/or BILSTM_CRF model
And/or the IDCNN_CRF model；The identification module 33 can include:
First identification submodule 331 obtains recognition result for identifying text to be identified using the CRF model；
Second identification submodule 332, for using the BILSTM_CRF model or the IDCNN_CRF model identification to
It identifies text, obtains recognition result；
Wherein, the first identification submodule 331 is particularly used in:
Word segmentation processing is carried out to the text to be identified, obtains multiple participles；
Determine the part of speech feature of each participle；
The segmentation sequence for the multiple participles composition for carrying the part of speech feature is inputted in the CRF model, institute is obtained
State the recognition result of CRF model；
The second identification submodule 332 is particularly used in:
The text to be identified is split into character, obtains multiple characters；
Determine the character feature vector of each character；
The sequence that the character feature vector of multiple characters forms is inputted into BILSTM_CRF model or IDCNN_CRF mould respectively
In type, the recognition result of BILSTM_CRF model or the recognition result of IDCNN_CRF model are obtained.
Fig. 9 is shown according to an exemplary embodiment a kind of for naming the block diagram of entity recognition device 900.Referring to figure
9, which may include processor 901, the machine readable storage medium 902 for being stored with machine-executable instruction.Processor
901 can communicate with machine readable storage medium 902 via system bus 903.Also, processor 901 is readable by read machine
Machine-executable instruction corresponding with name Entity recognition logic is known in storage medium 902 with executing name entity described above
Other method.
Machine readable storage medium 902 referred to herein can be any electronics, magnetism, optics or other physical stores
Device may include or store information, such as executable instruction, data, etc..For example, machine readable storage medium may is that
RAM (Radom Access Memory, random access memory), volatile memory, nonvolatile memory, flash memory, storage are driven
Dynamic device (such as hard disk drive), solid state hard disk, any kind of storage dish (such as CD, dvd) or similar storage are situated between
Matter or their combination.
The presently disclosed embodiments is described above, above description is exemplary, and non-exclusive, and
It is not limited to disclosed each embodiment.Without departing from the scope and spirit of illustrated each embodiment, for this skill
Many modifications and changes are obvious for the those of ordinary skill in art field.The selection of term used herein, purport
In the principle, practical application or technological improvement to the technology in market for best explaining each embodiment, or lead this technology
Other those of ordinary skill in domain can understand each embodiment disclosed herein.
Claims (12)
1. a kind of name entity recognition method, which is characterized in that the described method includes:
Text is labeled according to default mark collection to obtain training corpus, the default class of the mark collection comprising name entity
The notation methods of type and notation methods and non-name entity；
According to the training corpus, the preset multiple identification models of training；
Text to be identified is identified using the multiple identification model respectively, obtains the recognition result of each identification model；
The recognition result of weight and each identification model based on each identification model, obtains the identification knot of the text to be identified
Fruit.
2. the method according to claim 1, wherein the basis is default when the identification model is CRF model
Mark collection is labeled text to obtain training corpus, comprising:
Word segmentation processing is carried out to the text, obtains multiple participles；
The participle for belonging to name entity is labeled as corresponding type according to the notation methods that the default mark collection includes, and
The participle for belonging to non-name entity is labeled, training corpus is obtained.
3. according to the method described in claim 2, it is characterized in that, training preset multiple identifications according to the training corpus
Model includes:
According to the training corpus, the training CRF model, comprising:
Obtain the part of speech feature of each participle in the training corpus；
It is matrix form by the training corpus tissue, wherein first is classified as participle, the intermediate word for being classified as each participle in matrix
Property feature, last is classified as the annotation results of each participle, and respectively segmented in first row according to it is each participle in the training corpus
In position be ranked up；
Determine that the CRF template file that training process uses, the CRF template file are made of multiple template, each template is used for
Specify when extracting contextual information for current participle, extracted contextual information relative to the line displacement currently segmented,
And the absolute position of extracted contextual information column, the contextual information include the contextual feature and word of participle
The contextual feature of property；
According to the matrix and the CRF template file, CRF model is generated.
4. the method according to claim 1, wherein the identification model be BILSTM_CRF model or
When IDCNN_CRF model, the default mark collection of the basis is labeled text to obtain training corpus, comprising:
The text is divided into multiple characters；
The character label for belonging to name entity is corresponding type by the notation methods for including according to the default mark collection, and
The character for belonging to non-name entity is labeled, training corpus is obtained.
5. according to the method described in claim 4, it is characterized in that, training preset multiple identifications according to the training corpus
Model includes:
According to the training corpus, the training BILSTM_CRF model or the IDCNN_CRF model, comprising:
The character feature vector of each character in the training corpus is obtained, the character feature vector includes word vector sum character
Split vector；
According to the corresponding annotation results of each character of character feature vector sum of character each in the training corpus, described in training
BILSTM_CRF model or the IDCNN_CRF model.
6. the method according to claim 3 or 5, which is characterized in that the multiple identification model include CRF model and/or
BILSTM_CRF model and/or IDCNN_CRF model；
Text to be identified is identified using the CRF model, obtaining recognition result includes:
Word segmentation processing is carried out to the text to be identified, obtains multiple participles；
Determine the part of speech feature of each participle；
The segmentation sequence for the multiple participles composition for carrying the part of speech feature is inputted in the CRF model, the CRF is obtained
The recognition result of model；
Text to be identified is identified using the BILSTM_CRF model or the IDCNN_CRF model, obtaining recognition result includes:
The text to be identified is split into character, obtains multiple characters；
Determine the character feature vector of each character；
The sequence that the character feature vector of multiple characters forms is inputted into BILSTM_CRF model or IDCNN_CRF model respectively
In, obtain the recognition result of BILSTM_CRF model or the recognition result of IDCNN_CRF model.
7. a kind of name entity recognition device, which is characterized in that described device includes:
Labeling module, for being labeled to obtain training corpus, the default mark Ji Bao for text according to default mark collection
The notation methods of the type and notation methods of the entity containing name and non-name entity；
Training module, for according to the training corpus, the preset multiple identification models of training；
Identification module obtains the knowledge of each identification model for identifying text to be identified respectively using the multiple identification model
Other result；
Weighting block obtains described to be identified for the recognition result of weight and each identification model based on each identification model
The recognition result of text.
8. device according to claim 7, which is characterized in that the labeling module is specifically used for:
When the identification model is CRF model, word segmentation processing is carried out to the text, obtains multiple participles；
The participle for belonging to name entity is labeled as corresponding type according to the notation methods that the default mark collection includes, and
The participle for belonging to non-name entity is labeled, training corpus is obtained.
9. device according to claim 8, which is characterized in that the training module includes:
First training submodule, for according to the training corpus, the training CRF model；
The first training submodule is specifically used for:
Obtain the part of speech feature of each participle in the training corpus；
It is matrix form by the training corpus tissue, wherein first is classified as participle, the intermediate word for being classified as each participle in matrix
Property feature, last is classified as the annotation results of each participle, and respectively segmented in first row according to it is each participle in the training corpus
In position be ranked up；
Determine that the CRF template file that training process uses, the CRF template file are made of multiple template, each template is used for
Specify when extracting contextual information for current participle, extracted contextual information relative to the line displacement currently segmented,
And the absolute position of extracted contextual information column, the contextual information include the contextual feature and word of participle
The contextual feature of property；
According to the matrix and the CRF template file, CRF model is generated.
10. device according to claim 7, which is characterized in that the labeling module is specifically used for:
When the identification model is BILSTM_CRF model or IDCNN_CRF model, the text is divided into multiple characters；
The character label for belonging to name entity is corresponding type by the notation methods for including according to the default mark collection, and
The character for belonging to non-name entity is labeled, training corpus is obtained.
11. device according to claim 10, which is characterized in that the training module includes:
Second training submodule, for according to the training corpus, the training BILSTM_CRF model or the IDCNN_
CRF model；
The second training submodule is specifically used for:
The character feature vector of each character in the training corpus is obtained, the character feature vector includes word vector sum character
Split vector；
According to the corresponding annotation results of each character of character feature vector sum of character each in the training corpus, described in training
BILSTM_CRF model or the IDCNN_CRF model.
12. the device according to claim 9 or 11, which is characterized in that the multiple identification model include CRF model and/
Or BILSTM_CRF model and/or IDCNN_CRF model；The identification module includes:
First identification submodule obtains recognition result for identifying text to be identified using the CRF model；
Second identification submodule, for identifying text to be identified using the BILSTM_CRF model or the IDCNN_CRF model
This, obtains recognition result；
Wherein, the first identification submodule is specifically used for:
Word segmentation processing is carried out to the text to be identified, obtains multiple participles；
Determine the part of speech feature of each participle；
The segmentation sequence for the multiple participles composition for carrying the part of speech feature is inputted in the CRF model, the CRF is obtained
The recognition result of model；
The second identification submodule is specifically used for:
The text to be identified is split into character, obtains multiple characters；
Determine the character feature vector of each character；
The sequence that the character feature vector of multiple characters forms is inputted into BILSTM_CRF model or IDCNN_CRF model respectively
In, obtain the recognition result of BILSTM_CRF model or the recognition result of IDCNN_CRF model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811518491.3A CN109710925A (en) | 2018-12-12 | 2018-12-12 | Name entity recognition method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811518491.3A CN109710925A (en) | 2018-12-12 | 2018-12-12 | Name entity recognition method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109710925A true CN109710925A (en) | 2019-05-03 |
Family
ID=66255634
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811518491.3A Pending CN109710925A (en) | 2018-12-12 | 2018-12-12 | Name entity recognition method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109710925A (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110162795A (en) * | 2019-05-30 | 2019-08-23 | 重庆大学 | A kind of adaptive cross-cutting name entity recognition method and system |
| CN110210023A (en) * | 2019-05-23 | 2019-09-06 | 竹间智能科技（上海）有限公司 | A kind of calculation method of practical and effective name Entity recognition |
| CN110321566A (en) * | 2019-07-10 | 2019-10-11 | 北京邮电大学 | Chinese name entity recognition method, device, computer equipment and storage medium |
| CN110472062A (en) * | 2019-07-11 | 2019-11-19 | 新华三大数据技术有限公司 | The method and device of identification name entity |
| CN110472665A (en) * | 2019-07-17 | 2019-11-19 | 新华三大数据技术有限公司 | Model training method, file classification method and relevant apparatus |
| CN110489727A (en) * | 2019-07-12 | 2019-11-22 | 深圳追一科技有限公司 | Name recognition methods and relevant apparatus |
| CN110489739A (en) * | 2019-07-03 | 2019-11-22 | 东莞数汇大数据有限公司 | A kind of the name extracting method and its device of public security case and confession text based on CRF algorithm |
| CN110705300A (en) * | 2019-09-27 | 2020-01-17 | 上海烨睿信息科技有限公司 | Emotion analysis method, emotion analysis system, computer terminal and storage medium |
| CN110807324A (en) * | 2019-10-09 | 2020-02-18 | 四川长虹电器股份有限公司 | Video entity identification method based on IDCNN-crf and knowledge graph |
| CN111160035A (en) * | 2019-12-31 | 2020-05-15 | 北京明朝万达科技股份有限公司 | Text corpus processing method and device |
| CN111222337A (en) * | 2020-01-08 | 2020-06-02 | 山东旗帜信息有限公司 | Training method and device for entity recognition model |
| CN111651991A (en) * | 2020-04-15 | 2020-09-11 | 天津科技大学 | Medical named entity identification method utilizing multi-model fusion strategy |
| WO2020252950A1 (en) * | 2019-06-17 | 2020-12-24 | 五邑大学 | Named entity recognition method for medical texts based on pre-training model and fine turning technology |
| CN112784015A (en) * | 2021-01-25 | 2021-05-11 | 北京金堤科技有限公司 | Information recognition method and apparatus, device, medium, and program |
| CN112988979A (en) * | 2021-04-29 | 2021-06-18 | 腾讯科技（深圳）有限公司 | Entity identification method, entity identification device, computer readable medium and electronic equipment |
| CN115270799A (en) * | 2022-09-27 | 2022-11-01 | 北京云迹科技股份有限公司 | Named entity identification method and device |
| CN116911305A (en) * | 2023-09-13 | 2023-10-20 | 中博信息技术研究院有限公司 | Chinese address recognition method based on fusion model |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104615589A (en) * | 2015-02-15 | 2015-05-13 | 百度在线网络技术（北京）有限公司 | Named-entity recognition model training method and named-entity recognition method and device |
| CN106611375A (en) * | 2015-10-22 | 2017-05-03 | 北京大学 | Text analysis-based credit risk assessment method and apparatus |
| CN107169573A (en) * | 2017-05-05 | 2017-09-15 | 第四范式（北京）技术有限公司 | Using composite machine learning model come the method and system of perform prediction |
| CN107741927A (en) * | 2017-09-25 | 2018-02-27 | 沈阳航空航天大学 | Had complementary advantages tactful prepositional phrase recognition methods based on multi-model |
| CN107818083A (en) * | 2017-09-29 | 2018-03-20 | 华南师范大学 | Disease data name entity recognition method and system based on three layers of condition random field |
| CN108255816A (en) * | 2018-03-12 | 2018-07-06 | 北京神州泰岳软件股份有限公司 | A kind of name entity recognition method, apparatus and system |
| CN108763218A (en) * | 2018-06-04 | 2018-11-06 | 四川长虹电器股份有限公司 | A kind of video display retrieval entity recognition method based on CRF |
| CN108965245A (en) * | 2018-05-31 | 2018-12-07 | 国家计算机网络与信息安全管理中心 | Detection method for phishing site and system based on the more disaggregated models of adaptive isomery |
-
2018
- 2018-12-12 CN CN201811518491.3A patent/CN109710925A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104615589A (en) * | 2015-02-15 | 2015-05-13 | 百度在线网络技术（北京）有限公司 | Named-entity recognition model training method and named-entity recognition method and device |
| CN106611375A (en) * | 2015-10-22 | 2017-05-03 | 北京大学 | Text analysis-based credit risk assessment method and apparatus |
| CN107169573A (en) * | 2017-05-05 | 2017-09-15 | 第四范式（北京）技术有限公司 | Using composite machine learning model come the method and system of perform prediction |
| CN107741927A (en) * | 2017-09-25 | 2018-02-27 | 沈阳航空航天大学 | Had complementary advantages tactful prepositional phrase recognition methods based on multi-model |
| CN107818083A (en) * | 2017-09-29 | 2018-03-20 | 华南师范大学 | Disease data name entity recognition method and system based on three layers of condition random field |
| CN108255816A (en) * | 2018-03-12 | 2018-07-06 | 北京神州泰岳软件股份有限公司 | A kind of name entity recognition method, apparatus and system |
| CN108965245A (en) * | 2018-05-31 | 2018-12-07 | 国家计算机网络与信息安全管理中心 | Detection method for phishing site and system based on the more disaggregated models of adaptive isomery |
| CN108763218A (en) * | 2018-06-04 | 2018-11-06 | 四川长虹电器股份有限公司 | A kind of video display retrieval entity recognition method based on CRF |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110210023A (en) * | 2019-05-23 | 2019-09-06 | 竹间智能科技（上海）有限公司 | A kind of calculation method of practical and effective name Entity recognition |
| CN110162795A (en) * | 2019-05-30 | 2019-08-23 | 重庆大学 | A kind of adaptive cross-cutting name entity recognition method and system |
| WO2020252950A1 (en) * | 2019-06-17 | 2020-12-24 | 五邑大学 | Named entity recognition method for medical texts based on pre-training model and fine turning technology |
| CN110489739B (en) * | 2019-07-03 | 2023-06-20 | 东莞数汇大数据有限公司 | Naming extraction method and device for public security cases and oral text based on CRF algorithm |
| CN110489739A (en) * | 2019-07-03 | 2019-11-22 | 东莞数汇大数据有限公司 | A kind of the name extracting method and its device of public security case and confession text based on CRF algorithm |
| CN110321566A (en) * | 2019-07-10 | 2019-10-11 | 北京邮电大学 | Chinese name entity recognition method, device, computer equipment and storage medium |
| CN110321566B (en) * | 2019-07-10 | 2020-11-13 | 北京邮电大学 | Chinese named entity recognition method and device, computer equipment and storage medium |
| CN110472062B (en) * | 2019-07-11 | 2020-11-10 | 新华三大数据技术有限公司 | Method and device for identifying named entity |
| CN110472062A (en) * | 2019-07-11 | 2019-11-19 | 新华三大数据技术有限公司 | The method and device of identification name entity |
| CN110489727B (en) * | 2019-07-12 | 2023-07-07 | 深圳追一科技有限公司 | Person name recognition method and related device |
| CN110489727A (en) * | 2019-07-12 | 2019-11-22 | 深圳追一科技有限公司 | Name recognition methods and relevant apparatus |
| CN110472665A (en) * | 2019-07-17 | 2019-11-19 | 新华三大数据技术有限公司 | Model training method, file classification method and relevant apparatus |
| CN110705300A (en) * | 2019-09-27 | 2020-01-17 | 上海烨睿信息科技有限公司 | Emotion analysis method, emotion analysis system, computer terminal and storage medium |
| CN110807324A (en) * | 2019-10-09 | 2020-02-18 | 四川长虹电器股份有限公司 | Video entity identification method based on IDCNN-crf and knowledge graph |
| CN111160035B (en) * | 2019-12-31 | 2023-06-20 | 北京明朝万达科技股份有限公司 | Text corpus processing method and device |
| CN111160035A (en) * | 2019-12-31 | 2020-05-15 | 北京明朝万达科技股份有限公司 | Text corpus processing method and device |
| CN111222337A (en) * | 2020-01-08 | 2020-06-02 | 山东旗帜信息有限公司 | Training method and device for entity recognition model |
| CN111651991B (en) * | 2020-04-15 | 2022-08-26 | 天津科技大学 | Medical named entity identification method utilizing multi-model fusion strategy |
| CN111651991A (en) * | 2020-04-15 | 2020-09-11 | 天津科技大学 | Medical named entity identification method utilizing multi-model fusion strategy |
| CN112784015A (en) * | 2021-01-25 | 2021-05-11 | 北京金堤科技有限公司 | Information recognition method and apparatus, device, medium, and program |
| CN112784015B (en) * | 2021-01-25 | 2024-03-12 | 北京金堤科技有限公司 | Information identification method and device, apparatus, medium, and program |
| CN112988979A (en) * | 2021-04-29 | 2021-06-18 | 腾讯科技（深圳）有限公司 | Entity identification method, entity identification device, computer readable medium and electronic equipment |
| CN115270799A (en) * | 2022-09-27 | 2022-11-01 | 北京云迹科技股份有限公司 | Named entity identification method and device |
| CN116911305A (en) * | 2023-09-13 | 2023-10-20 | 中博信息技术研究院有限公司 | Chinese address recognition method based on fusion model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109710925A (en) | Name entity recognition method and device | |
| Wang et al. | Crossweigh: Training named entity tagger from imperfect annotations | |
| WO2021139424A1 (en) | Text content quality evaluation method, apparatus and device, and storage medium | |
| CN110705293A (en) | Electronic medical record text named entity recognition method based on pre-training language model | |
| CN112597774B (en) | Chinese medical named entity recognition method, system, storage medium and equipment | |
| CN108628824A (en) | A kind of entity recognition method based on Chinese electronic health record | |
| CN112800766B (en) | Active learning-based Chinese medical entity identification labeling method and system | |
| CN112464662B (en) | Medical phrase matching method, device, equipment and storage medium | |
| CN108959566B (en) | A kind of medical text based on Stacking integrated study goes privacy methods and system | |
| CN111949759A (en) | Method and system for retrieving medical record text similarity and computer equipment | |
| CN110096572B (en) | Sample generation method, device and computer readable medium | |
| He | Towards Visual Question Answering on Pathology Images. | |
| CN110931128A (en) | Method, system and device for automatically identifying unsupervised symptoms of unstructured medical texts | |
| CN113393916B (en) | Method and device for extracting structural relationship of coronary artery medical report | |
| CN111651991A (en) | Medical named entity identification method utilizing multi-model fusion strategy | |
| Attardi et al. | Annotation and Extraction of Relations from Italian Medical Records. | |
| CN107122582B (en) | diagnosis and treatment entity identification method and device facing multiple data sources | |
| Michalopoulos et al. | ICDBigBird: a contextual embedding model for ICD code classification | |
| Mitkov et al. | Methods for extracting and classifying pairs of cognates and false friends | |
| Syeda-Mahmood et al. | Extracting and learning fine-grained labels from chest radiographs | |
| CN117422074A (en) | Method, device, equipment and medium for standardizing clinical information text | |
| CN112101030A (en) | Method, device and equipment for establishing term mapping model and realizing standard word mapping | |
| CN109635123A (en) | A kind of Chinese medicine text concept recognition methods of increment type | |
| CN112765353B (en) | Scientific research text-based biomedical subject classification method and device | |
| CN115545021A (en) | Clinical term identification method and device based on deep 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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190503 |