CN110427542A

CN110427542A - Sorter network training and data mask method and device, equipment, medium

Info

Publication number: CN110427542A
Application number: CN201810385973.XA
Authority: CN
Inventors: 林子义; 邵婧; 赵宇航; 闫俊杰
Original assignee: Beijing Sensetime Technology Development Co Ltd
Current assignee: Beijing Sensetime Technology Development Co Ltd
Priority date: 2018-04-26
Filing date: 2018-04-26
Publication date: 2019-11-08

Abstract

The embodiment of the present disclosure discloses a kind of trained sorter network and data mask method and device, equipment, medium, wherein sorter network training method includes: the satellite information based on the associated multiple sample datas of target word, determines the mark label of the multiple sample data；Mark label based on the multiple sample data and the multiple sample data, training sorter network.The mark label for obtaining sample data based on multi-source based on disclosure above-described embodiment, based on the sample data training sorter network with mark label, for obtained sorter network when handling image data, accuracy is higher.

Description

Sorter network training and data mask method and device, equipment, medium

Technical field

This disclosure relates to computer vision technique, especially a kind of sorter network training and data mask method and device, Electronic equipment, computer storage medium.

Background technique

With the continuous development of deep learning, as long as theoretically possessing enough training datas, general picture classification is asked Topic can be obtained by very good solution.Web data is the data for being easier to obtain in a large amount of data, i.e., some social activities, User's uploading pictures on the websites such as camera shooting, the source of these data is numerous users of website, and amount is bigger, can also be led to The web crawlers of post-mature carries out batch acquisition；But the processing of web data is more complicated, difficulty in treatment is then its not lattice The characteristics of formula and a large amount of noises generated during data upload, acquisition etc..

Using general depth learning technology, specifically, picture classification needs genuine and believable reflection image content Label, these tag characterizations image content also provides indispensable input to deep learning algorithm.Existing data Acquisition modes mainly rely on mark person manually to mark, by people itself resolution capability to outgoing label.

Summary of the invention

A kind of sorter network training technique and label for labelling technology that the embodiment of the present disclosure provides.

According to the one aspect of the embodiment of the present disclosure, a kind of sorter network training method for providing, comprising:

Based on the satellite information of the associated multiple sample datas of target word, the mark mark of the multiple sample data is determined Label；

Mark label based on the multiple sample data and the multiple sample data, training sorter network.

Optionally, before the satellite information based on multiple sample datas, the mark label for determining multiple sample datas, also Include:

Network retrieval is carried out based on the target word, obtains the search result of the target word, the search result includes The satellite information of the multiple sample data and the multiple sample data.

Optionally, the satellite information based on the associated multiple sample datas of target word, determines the multiple sample number According to mark label, comprising:

Satellite information based on the sample data carries out Semantic mapping processing, and it is corresponding at least to obtain the sample data One classificating word；

Based at least one described classificating word, the mark label of the sample data is determined.

Optionally, described based at least one described classificating word, determine the mark label of the sample data, comprising:

Based on corresponding first weighted value of the target word and corresponding second weighted value of at least one described classificating word, really The mark label of the fixed sample data.

Optionally, described based at least one described classificating word, determine the mark label of the sample data, comprising: will At least one described classificating word is determined as the mark label of the sample data, wherein the mark label of the sample data is also Including the target word；

The mark label based on the multiple sample data and the multiple sample data, training sorter network, packet It includes:

Based on the corresponding first error of the target word and corresponding first weighted value of the first error and it is described extremely Few corresponding second error of a classificating word and corresponding second weighted value of second error, determine loss function value；

Based on the loss function value, processing is adjusted to the parameter value of the sorter network.

Optionally, the satellite information based on the sample data carries out Semantic mapping processing, obtains the sample number According at least one corresponding classificating word, comprising:

Keyword extraction processing is carried out to the satellite information of the sample data, it is corresponding at least to obtain the sample data One keyword；

At least one keyword corresponding to the sample data carries out Semantic mapping processing, obtains the sample data pair At least one classificating word answered.

Optionally, described at least one keyword corresponding to the sample data carries out Semantic mapping processing, obtains institute State at least one corresponding classificating word of sample data, comprising:

Using term vector model, the corresponding crucial vector of each keyword at least one described keyword is determined；

Based on the crucial vector of each keyword at least one described keyword, determined from multiple preset classificating words At least one corresponding described classificating word of at least one described keyword.

Optionally, the crucial vector based on each keyword at least one described keyword, from multiple preset At least one corresponding described classificating word of at least one described keyword is determined in classificating word, comprising:

By distance between corresponding class vector and the crucial vector of the keyword in the multiple preset classificating word Shortest classificating word is determined as the corresponding classificating word of the keyword.

Optionally, Semantic mapping processing is being carried out at least one corresponding keyword of the sample data, obtained described Before at least one corresponding classificating word of sample data, further includes:

At least one keyword corresponding to the sample data is filtered operation, and it is corresponding to obtain the sample data At least one target keyword；

Described at least one keyword corresponding to the sample data carries out Semantic mapping processing, obtains the sample number According at least one corresponding classificating word, comprising:

At least one target keyword corresponding to the sample data carries out Semantic mapping processing, obtains the sample number According at least one corresponding classificating word.

Optionally, at least one keyword corresponding to the sample data is filtered operation, obtains the sample number According at least one corresponding target keyword, comprising:

It is specially multiple keywords and the multiple pass in response at least one corresponding keyword of the sample data It include to be closed there are at least two first keywords of mutex relation and/or inclusion relation from described at least two first in keyword The target keyword is determined in keyword.

It is optionally, described that the target keyword is determined from least two first keyword, comprising:

It determines at least two first keyword between each first keyword and at least one second keyword The degree of association, wherein second keyword is the pass in the multiple keyword in addition at least two first keyword Keyword；

Based on each first keyword at least two first keyword and at least one described second keyword it Between the degree of association, the target keyword is determined from least two first keyword.

Optionally, it is described based on each first keyword at least two first keyword and it is described at least one the The degree of association between two keywords determines the target keyword from least two first keyword, comprising:

Most by the degree of association at least two first keyword between at least one described second keyword The first big keyword is as the target keyword.

Optionally, each first keyword and at least one second pass in the determination at least two first keyword The degree of association between keyword, comprising:

Based on neural network determine at least two first keyword each first keyword and it is described at least one The degree of association of second keyword, the neural network are got by the training of sample word, and the sample word collection includes at least Two sample words, each sample word are labeled with the degree of association with sample word described in other.

Optionally, the satellite information comprises at least one of the following:

Heading message, content description information, comment information, contextual information.

Optionally, in the mark label based on the multiple sample data and the multiple sample data, training classification net Before network, further includes:

The sorter network is trained in advance based on multiple first sample data in the multiple sample data, obtains pre- instruction The sorter network after white silk；

Mark label based on the multiple sample data and the multiple sample data, to the classification after the pre-training Network is trained.

Optionally, the classification net is being trained in advance based on multiple first sample data in the multiple sample data Network, before the sorter network after obtaining pre-training, further includes:

Using the sample data in the multiple sample data with the maximum preset quantity of the target word degree of correlation as institute State multiple first sample data.

According to the other side of the embodiment of the present disclosure, a kind of data mask method for providing, comprising:

Obtain the satellite information of target word associated multiple sample datas and the multiple sample data；

Based on the satellite information of the multiple sample data, the mark label of the multiple sample data is obtained.

Optionally, the satellite information for obtaining target word associated multiple sample datas and the multiple sample data, Include:

Optionally, the mark label of the sample data includes multiple labels；

Described at least one classificating word based on described in, determines the mark label of the sample data, comprising:

At least one described classificating word and the target word are determined as to the mark label of the sample data.

Optionally, the satellite information comprises at least one of the following:

According to the other side of the embodiment of the present disclosure, a kind of sorter network training device for providing, comprising:

Label for labelling unit determines the multiple for the satellite information based on the associated multiple sample datas of target word The mark label of sample data；

Network training unit, for the mark label based on the multiple sample data and the multiple sample data, instruction Practice sorter network.

Optionally, further includes:

Target retrieval unit, for obtaining the search result of the target word based on target word progress network retrieval, The search result includes the satellite information of the multiple sample data and the multiple sample data.

Optionally, the label for labelling unit, comprising:

Semantic mapping module carries out Semantic mapping processing for the satellite information based on the sample data, obtains described At least one corresponding classificating word of sample data；

Classificating word module, for determining the mark label of the sample data based at least one described classificating word.

Optionally, the classificating word module, be specifically used for based on corresponding first weighted value of the target word and it is described extremely Few corresponding second weighted value of a classificating word, determines the mark label of the sample data.

Optionally, the classificating word module, specifically at least one described classificating word is determined as the sample data Mark label, wherein the mark label of the sample data further includes the target word；

The network training unit is specifically used for being based on the corresponding first error of the target word and the first error pair The first weighted value and corresponding second error of at least one described classificating word answered and corresponding second power of second error Weight values determine loss function value；Based on the loss function value, processing is adjusted to the parameter value of the sorter network.

Optionally, the semantic mapping module, comprising:

Keyword extracting module carries out keyword extraction processing for the satellite information to the sample data, obtains institute State at least one corresponding keyword of sample data；

Keyword mapping block carries out at Semantic mapping at least one keyword corresponding to the sample data Reason, obtains at least one corresponding classificating word of the sample data.

Optionally, the keyword mapping block, comprising:

Crucial vector module determines each keyword pair at least one described keyword for utilizing term vector model The crucial vector answered；

Classificating word determining module, for the crucial vector based on each keyword at least one described keyword, from more At least one corresponding described classificating word of at least one described keyword is determined in a preset classificating word.

Optionally, the classificating word determining module is specifically used for corresponding classification in the multiple preset classificating word The shortest classificating word of distance is determined as the corresponding classificating word of the keyword between vector and the crucial vector of the keyword.

Optionally, the semantic mapping module, further includes:

Keyword filtering module is obtained for being filtered operation at least one corresponding keyword of the sample data To at least one corresponding target keyword of the sample data；

The keyword mapping block is specifically used at least one target keyword corresponding to the sample data and carries out Semantic mapping processing, obtains at least one corresponding classificating word of the sample data.

Optionally, the keyword filtering module is specifically used at least one pass corresponding in response to the sample data Keyword is specially in multiple keywords and the multiple keyword comprising there are at least the two of mutex relation and/or inclusion relation A first keyword determines the target keyword from least two first keyword.

Optionally, the keyword filtering module, comprising:

Keyword degree of association module, for determine at least two first keyword each first keyword at least The degree of association between one the second keyword, wherein second keyword is in the multiple keyword except described at least two Keyword except a first keyword；

Target keyword module, for based on each first keyword at least two first keyword and it is described extremely The degree of association between few second keyword, determines the target keyword from least two first keyword.

Optionally, the target keyword module, be specifically used for by least two first keyword with it is described extremely Maximum first keyword of the degree of association between few second keyword is as the target keyword.

Optionally, the keyword degree of association module, specifically for determining described at least two first based on neural network The degree of association of each first keyword and at least one second keyword in keyword, the neural network pass through sample word Language training is got, and the sample word collection includes at least two sample words, and each sample word is labeled with and other The degree of association of the sample word.

Optionally, the satellite information comprises at least one of the following:

Optionally, further includes:

Pre-training unit, for based on described point of the training in advance of multiple first sample data in the multiple sample data Class network, the sorter network after obtaining pre-training；

The network training unit, specifically for the mark based on the multiple sample data and the multiple sample data Label is trained the sorter network after the pre-training.

Optionally, further includes:

Pretreatment unit, for by the multiple sample data with the maximum preset quantity of the target word degree of correlation Sample data is as the multiple first sample data.

According to the other side of the embodiment of the present disclosure, a kind of data annotation equipment for providing, comprising:

Information acquisition unit, for obtaining the attached of the associated multiple sample datas of target word and the multiple sample data Information；

Label for labelling unit obtains the multiple sample data for the satellite information based on the multiple sample data Mark label.

Optionally, the information acquisition unit is specifically used for carrying out network retrieval based on the target word, obtains the mesh The search result of word is marked, the search result includes the satellite information of the multiple sample data and the multiple sample data.

Optionally, the label for labelling unit, comprising:

Optionally, the mark label of the sample data includes multiple labels；

The classificating word module, specifically at least one described classificating word and the target word are determined as the sample The mark label of data.

Optionally, the semantic mapping module, comprising:

Optionally, the keyword mapping block, comprising:

Optionally, the semantic mapping module, further includes:

Optionally, the keyword filtering module, comprising:

Optionally, the satellite information comprises at least one of the following:

According to the other side of the embodiment of the present disclosure, a kind of electronic equipment provided, including processor, the processor Including data annotation equipment described in sorter network training device described in any one as above or any one as above.

According to the other side of the embodiment of the present disclosure, a kind of electronic equipment that provides, comprising: memory, for storing Executable instruction；

And processor, it is as above any one to complete that the executable instruction is executed for communicating with the memory The operation of the described sorter network training method, alternatively, for being communicated with the memory with execute the executable instruction from And complete the operation of data mask method described in any one as above.

According to the other side of the embodiment of the present disclosure, a kind of computer storage medium provided, for storing computer The instruction that can be read, described instruction, which is performed, executes sorter network training method described in any one as above or as above any one The operation of the item data mask method.

According to the other side of the embodiment of the present disclosure, a kind of computer program product provided, including it is computer-readable Code, when the computer-readable code is run in equipment, the processor in the equipment is executed for realizing such as taking up an official post The instruction of data mask method described in a sorter network training method of anticipating or any one as above.

According to another aspect of the embodiment of the present disclosure, another computer program product provided is calculated for storing Machine readable instruction, described instruction is performed so that computer executes classification net described in any of the above-described possible implementation Network training method, or execute data mask method described in any possible implementation.

In an optional embodiment, the computer program product is specially computer storage medium, at another In optional embodiment, the computer program product is specially software product, such as SDK etc..

Another sorter network training and data mask method are additionally provided according to the embodiment of the present disclosure and device, electronics are set Standby, computer storage medium, computer program product, wherein the satellite information based on the associated multiple sample datas of target word, Determine the mark label of multiple sample datas；Mark label based on multiple sample datas and multiple sample datas, training classification Network.

It is set based on a kind of disclosure sorter network training provided by the above embodiment and data mask method and device, electronics Standby, computer storage medium, computer program product are determined based on the satellite information of the associated multiple sample datas of target word The mark label of multiple sample datas；Mark label based on multiple sample datas and multiple sample datas, training sorter network； The mark label that sample data is obtained based on multi-source is obtained based on the sample data training sorter network with mark label Sorter network when handling image data, accuracy is higher.

Below by drawings and examples, the technical solution of the disclosure is described in further detail.

Detailed description of the invention

The attached drawing for constituting part of specification describes embodiment of the disclosure, and together with description for explaining The principle of the disclosure.

The disclosure can be more clearly understood according to following detailed description referring to attached drawing, in which:

Fig. 1 is the flow chart for the sorter network training method that the embodiment of the present disclosure provides.

Fig. 2 is the structural schematic diagram for the sorter network training device that the embodiment of the present disclosure provides.

Fig. 3 is the flow chart for the data mask method that the embodiment of the present disclosure provides.

Fig. 4 is the structural schematic diagram for the data annotation equipment that the embodiment of the present disclosure provides.

Fig. 5 is the structural representation suitable for the electronic equipment of the terminal device or server that are used to realize the embodiment of the present disclosure Figure.

Specific embodiment

The various exemplary embodiments of the disclosure are described in detail now with reference to attached drawing.It should also be noted that unless in addition having Body explanation, the unlimited system of component and the positioned opposite of step, numerical expression and the numerical value otherwise illustrated in these embodiments is originally Scope of disclosure.

Simultaneously, it should be appreciated that for ease of description, the size of various pieces shown in attached drawing is not according to reality Proportionate relationship draw.

Be to the description only actually of at least one exemplary embodiment below it is illustrative, never as to the disclosure And its application or any restrictions used.

Technology, method and apparatus known to person of ordinary skill in the relevant may be not discussed in detail, but suitable In the case of, the technology, method and apparatus should be considered as part of specification.

It should also be noted that similar label and letter indicate similar terms in following attached drawing, therefore, once a certain Xiang Yi It is defined in a attached drawing, then in subsequent attached drawing does not need that it is further discussed.

The embodiment of the present disclosure can be applied to computer system/server, can be with numerous other general or specialized calculating System environments or configuration operate together.Suitable for be used together with computer system/server well-known computing system, ring The example of border and/or configuration includes but is not limited to: personal computer system, server computer system, thin client, thick client Machine, hand-held or laptop devices, microprocessor-based system, set-top box, programmable consumer electronics, NetPC Network PC, Minicomputer system, large computer system and distributed cloud computing technology environment including above-mentioned any system, etc..

Computer system/server can be in computer system executable instruction (such as journey executed by computer system Sequence module) general context under describe.In general, program module may include routine, program, target program, component, logic, number According to structure etc., they execute specific task or realize specific abstract data type.Computer system/server can be with Implement in distributed cloud computing environment, in distributed cloud computing environment, task is long-range by what is be linked through a communication network Manage what equipment executed.In distributed cloud computing environment, it includes the Local or Remote meter for storing equipment that program module, which can be located at, It calculates in system storage medium.

In implementing the present disclosure, publisher has found that the existing technology has at least the following problems: being by mark person The method that data manually mark label needs to consume a large amount of time, manpower, money.Though the method for existing automatic marking label Cost so has been saved, it, can error in precision compared with artificial mark.

Step 110, the satellite information based on the associated multiple sample datas of target word, determines the mark of multiple sample datas Label.

Optionally, target word can be term, carries out retrieval based on target word and obtains sample data and its satellite information, Such as retrieval or the web search, etc. of database are carried out based on target word, wherein in some optional embodiments, sample data It can be web data, but the embodiment of the present disclosure is without being limited thereto.

Optionally, satellite information can include but is not limited to one of following or a variety of: heading message, content description letter Breath, comment information, contextual information.Wherein heading message may include the corresponding title of sample data, such as web page title, interior Hold description that description information may include the content for sample data, such as abstract etc., comment information includes to sample data Comment, such as from network the modes such as other parts of webpage where other related web pages or web data obtain to the net The comment content of page data, contextual information may include the associated data or content of sample data, for example, it may be based on mark What topic information and/or content description information and/or comment information obtained has associated content with the sample data, these are attached The specific manifestation form for belonging to information includes but is not limited to literal expression.In some embodiments, can be appointed from above-mentioned group Meaning selection obtains satellite information or satellite information also and may include other kinds of information, and the embodiment of the present disclosure does not do this It limits.

In some embodiments, the mark label that sample data is determined by satellite information, specifically can be based on attached Information obtains some keywords, and obtains the label of sample data based on the keyword, for example, can carry out to satellite information Semantic mapping carries out Semantic mapping to the keyword obtained based on satellite information, obtains the label of sample data；Optionally, also It can be using target word as the label of some sample datas；Alternatively, can be with combining target word and the mark obtained based on satellite information Label, obtain the mark label of sample data, for example, target word and the label obtained based on satellite information can be weighted place Reason obtains mark label, for another example sample data can have multi-tag, wherein target word and is obtained based on satellite information Label is respectively as the component, etc. in multi-tag.In this way, the mark label of sample data can have a variety of different sources And method of determination, mark label more describes sample data, provides more letters for training sorter network Breath.

Step 120, the mark label based on multiple sample datas and multiple sample datas, training sorter network.

The specific training process of sorter network can realize by any sorter network training method, such as: it will be multiple Sample data inputs sorter network, the corresponding classification results of sample data is obtained based on sorter network, based on classification results and mark Note label determines loss function value, is adjusted by reversed gradient method to the parameter in sorter network based on loss function value It is whole.

Based on disclosure sorter network training method provided by the above embodiment, it is based on the associated multiple sample numbers of target word According to satellite information, determine the mark label of multiple sample datas；Mark based on multiple sample datas and multiple sample datas Label, training sorter network are conducive to the accuracy for improving sorter network.

It can also include: that net is carried out based on target word prior to step 110 in one or more optional embodiments Network retrieval, obtains the search result of target word, search result includes the satellite information of multiple sample datas and multiple sample datas.

It is retrieved from network based on target word, obtains sample data relevant to target word and sample data is corresponding attached Information, specifically, the sample data of acquisition can include but is not limited to Web page picture, and a feature of Web page picture is exactly to include A large amount of satellite information, social network sites user can take some information in uploading pictures, and obtain from these satellite informations It can be used for the mark label of deep learning training.One based on target word and search acquisition web data is advantageous in that, data Although information do not format but richer, can more fully excavate information therein；Be conducive to train sorter network, Make trained sorter network that there is stronger robustness.

In one or more optional embodiments, step 110 may include:

Satellite information based on sample data carries out Semantic mapping processing, obtains at least one corresponding classification of sample data Word.

It is alternatively possible to directly carry out Semantic mapping processing to satellite information, alternatively, satellite information can be based on, obtain At least one keyword, and Semantic mapping processing, etc. is carried out to keyword, the embodiment of the present disclosure does not limit this.

It is alternatively possible to which the satellite information to sample data carries out keyword extraction processing, it is corresponding to obtain sample data At least one keyword, and Semantic mapping processing is carried out at least one corresponding keyword of sample data, obtain sample data At least one corresponding classificating word.

Keyword extraction is carried out to these satellite informations by keyword extraction techniques, it is corresponding at least to obtain satellite information One keyword, wherein keyword extraction techniques can be any appropriate keyword extracting method, such as: it is extracted based on word frequency The method TF-IDF of the keyword or method Topic-model, etc. that keyword is extracted using topic model, the disclosure is to key The specific method that word extracts is with no restrictions.

In the embodiments of the present disclosure, Semantic mapping can be carried out in several ways.Optionally, corresponding to sample data At least one keyword carries out Semantic mapping processing, obtains at least one corresponding classificating word of sample data, comprising:

Using term vector model, the corresponding crucial vector of each keyword at least one keyword is determined；

Based on the crucial vector of each keyword at least one keyword, determined at least from multiple preset classificating words At least one corresponding classificating word of one keyword.

Optionally, Semantic mapping can using the term vector word2vec model after training, by keyword according to it is semantic with At least one preset classificating word establishes mapping relations, be specifically, during the foundation of term vector model mapping based on to What amount was established, therefore, the corresponding crucial vector sum of keyword need to be obtained and preset the corresponding class vector of classificating word, it specifically, will The shortest classificating word of distance is determined as between corresponding class vector and the crucial vector of keyword in multiple preset classificating words The corresponding classificating word of keyword.The distance between class vector and crucial the vector shorter corresponding classificating word of explanation and corresponding pass The keyword meaning of a word is more close, by crucial vector and its apart from shortest class vector, that is, can determine that its corresponding classificating word is closest The corresponding classificating word of the class vector.

Term vector word2vec model is the tool that word is converted into vector form.It can be to content of text The vector operation being reduced in vector space is handled, the similarity in vector space is calculated, to indicate the phase on text semantic Like degree.

Optionally, Semantic mapping processing is being carried out at least one corresponding keyword of sample data, is obtaining sample data Before at least one corresponding classificating word, further includes:

At least one keyword corresponding to sample data is filtered operation, obtain sample data it is corresponding at least one Target keyword；

At least one keyword corresponding to sample data carries out Semantic mapping processing, and it is corresponding at least to obtain sample data One classificating word, comprising:

At least one target keyword corresponding to sample data carries out Semantic mapping processing, and it is corresponding to obtain sample data At least one classificating word.

The present embodiment passes through filter operation, and the noise data at least one keyword is found out and is classified as correctly Classification is based on carrying out model training with noise data in other modes relatively, or whole noises in data is all deleted After carry out model training, the embodiment of the present disclosure is by being referred to correct classification for noise data, to noise and without mark When data are classified, obvious improves classification results.

Optionally, at least one keyword corresponding to sample data is filtered operation, and it is corresponding to obtain sample data At least one target keyword, comprising:

It is specially to be wrapped in multiple keywords and multiple keywords in response at least one corresponding keyword of sample data Containing there are at least two first keywords of mutex relation and/or inclusion relation, mesh is determined from least two first keywords Mark keyword.

Since at least one keyword is all based on what multiple sample datas obtained, and multiple sample datas are based on target What word and search obtained, therefore, when comprising there are when at least two keyword of mutex relation, illustrating wherein extremely in these keywords There is mistake in a rare keyword, is corrected；And when comprising saying there are when at least two keyword of inclusion relation Bright range definition inaccuracy, also needs to be corrected, and from these, there are at least two the of mutex relation and/or inclusion relation Determination can illustrate the target keyword of sample data in one keyword.

Mutex relation self-explanatory characters' part A and event B will not occur simultaneously in any primary test, then claim event A and event B Mutual exclusion.Two keywords illustrate that the two keywords do not occur generally simultaneously there are mutex relation, such as: indoor and outdoors are just It is the keyword of a pair of of mutual exclusion, when there are when multiple keywords of mutual exclusion, illustrate to deposit in the keyword obtained for the same data of correspondence In problem, the Partial key word in the keyword by mutual exclusion is needed to delete, specifically deletes which or which keyword in combination with right Should data other keywords determine, such as: the corresponding keyword of a data includes: indoor and outdoor, sky, white clouds, grass Ground." interior " this keyword should be deleted by so combining sky, white clouds, these three keywords of meadow to can determine, and retain " room Outside ".

Inclusion relation is the subordinate relation between set and set, is also subset relation.Using between sub- word, there is packet Two words containing relationship are properly termed as comprising word and comprising word, that is, are indicated that the range referred to comprising word is certain It within comprising word, and include the range that is referred to of word not necessarily by within comprising word.Such as: " animal " and " cat " The two words i.e. there are inclusion relation, " animal " be comprising word, " cat " be comprising word, be " cat " must be " dynamic Object ", but " animal " is not necessarily " cat ".There are the keyword of inclusion relation not contradictions for corresponding two of same picture, but in order to More accurate label is obtained, is needed to judge there are which is more acurrate between the keyword of inclusion relation, and the process judged is base In comprising keyword and included what incidence relation between keyword and other keywords determined.

Optionally, target keyword is determined from least two first keywords, comprising:

Determine being associated between each first keyword and at least one second keyword at least two first keywords Degree, wherein the second keyword is the keyword in multiple keywords in addition at least two first keywords；

Based on being associated between each first keyword and at least one second keyword at least two first keywords Degree determines target keyword from least two first keywords.

Specifically, by the degree of association at least two first keywords between at least one second keyword maximum One keyword is as target keyword.

Determination can illustrate sample data from there are at least two first keywords of mutex relation and/or inclusion relation Target keyword is determined based on the degree of association between the first keyword and the second keyword, when at least two first keys There is the degree of association of first keyword and other the second keywords to be greater than other first keywords and other second keys in word The degree of association of word, it is believed that first keyword is target keyword；Alternatively, when having one at least two first keywords When the degree of association of first keyword and other the second keywords is greater than a setting value, it is believed that first keyword is target Keyword.

Optionally, based on neural network determine at least two first keywords each first keyword and at least one the The degree of association of two keywords, neural network are got by the training of sample word, and sample word collection includes at least two sample words Language, each sample word are labeled with the degree of association with other sample words.

The degree of association between the first keyword and the second keyword is determined by trained neural network, the nerve net The training of network is carried out based on the sample word for being labeled with the degree of association, and therefore, which can accurately obtain at least The degree of association between two words, the degree of association obtained based on the neural network can determine that first keyword for target pass Keyword.

Step 110 can also include: the mark label that sample data is determined based at least one classificating word.

Since other notation methods can not utilize effective informations more in sample data (such as: satellite information), this public affairs It opens embodiment and can use natural language processing tool and the satellite information of sample data is excavated, and these information are also added Enter into the training of sorter network.

Optionally, it is based on corresponding first weighted value of target word the second weighted value corresponding at least one classificating word, really The mark label of random sample notebook data.

In some embodiments, the ratio-dependent based on the first weighted value and at least one the second weighted value marks label, The mark label includes the target word of the first ratio and at least one classificating word of the second ratio, wherein the first ratio and second Ratio can be preset, alternatively, the first ratio a and the second ratio b can be obtained respectively by formula (1) and (2):

A=m/ (m+ ∑ n_i) formula (1)

B=n_jj/(m+∑_i≠jn_i) formula (2)

Wherein, m indicates the first weighted value, n_iIndicate corresponding second weighted value of i-th of classificating word, wherein i be greater than etc. It is less than k in 1, k value represents the quantity of classificating word.

At this point, optionally, when being trained based on mark label to sorter network, loss function value is based on predicted value Subtract what the first ratio was determined multiplied by target word and the second ratio multiplied by least one classificating word.

Alternatively, optionally, mark label can be specially multi-tag, i.e., at least one classificating word and target word are determined as The mark label of sample data.At this point, it is alternatively possible to the cum rights training method using multi-tag instructs sorter network Practice.

Step 120 may include:

Based on the corresponding first error of target word and corresponding first weighted value of first error and at least one classificating word Corresponding second error and corresponding second weighted value of the second error, determine loss function value；

Based on loss function value, processing is adjusted to the parameter value of sorter network.

It include target word and at least one classificating word in labeled data in the present embodiment, and the first weighted value and second Weighted value is applied when calculating loss function, optionally, obtains first error, base as mark label based on target word The second error is obtained as mark label in classificating word, first error is subtracted multiplied by the first weighted value and all the by predicted value Two errors determine loss function value multiplied by the second weighted value.

In one or more optional embodiments, pre-training can also be carried out, before above-mentioned training to improve training Efficiency.For example, before step 120, can also include:

Sorter network is trained in advance based on multiple first sample data in multiple sample datas, point after obtaining pre-training Class network；

Mark label based on multiple sample datas and multiple sample datas, training sorter network, comprising:

Mark label based on multiple sample datas and multiple sample datas, instructs the sorter network after pre-training Practice.

In this way, network convergence rate can be accelerated by being trained to the network that pre-training obtains, network training is improved Efficiency.

Optionally, sorter network is being trained in advance based on multiple first sample data in multiple sample datas, obtaining pre- Before sorter network after training, can also include:

Using in multiple sample datas with the sample data of the maximum preset quantity of the target word degree of correlation as multiple first samples Notebook data.

Sample of some embodiments using part sorted order forward data and its corresponding target word as pre-training Data, the sorter network after an available relatively good pre-training.It is being tied on the basis of sorter network after pre-training It closes mark label, the first weighted value and the second weighted value to be trained, can further promote the classification accuracy of sorter network. Using target word as the pre-training label of the data with the maximum preset quantity of the target word degree of correlation.Since data are based on target What word was searched for, therefore more forward data mesh (higher with the target word degree of correlation) in the data that search can be obtained Mark word is trained as confidence label.By only selecting the higher partial data of the degree of correlation as pre-training, available one A relatively good pre-training model.It is higher as the confidence level of label using term because sorted order is forward.

Those of ordinary skill in the art will appreciate that: realize that all or part of the steps of above method embodiment can pass through The relevant hardware of program instruction is completed, and program above-mentioned can be stored in a computer readable storage medium, the program When being executed, step including the steps of the foregoing method embodiments is executed；And storage medium above-mentioned includes: ROM, RAM, magnetic disk or light The various media that can store program code such as disk.

Fig. 2 is the structural schematic diagram for the sorter network training device that the embodiment of the present disclosure provides.The device of the embodiment can For realizing the above-mentioned each method embodiment of the disclosure.As shown in Fig. 2, the device of the embodiment includes:

Label for labelling unit 21 determines multiple samples for the satellite information based on the associated multiple sample datas of target word The mark label of notebook data.

Network training unit 22, for the mark label based on multiple sample datas and multiple sample datas, training classification Network.

Based on disclosure sorter network training device provided by the above embodiment, it is based on the associated multiple sample numbers of target word According to satellite information, determine the mark label of multiple sample datas；Mark based on multiple sample datas and multiple sample datas Label, training sorter network are conducive to the accuracy for improving sorter network.

In one or more optional embodiments, further includes:

Target retrieval unit obtains the search result of target word, search result for carrying out network retrieval based on target word Satellite information including multiple sample datas and multiple sample datas.

In one or more optional embodiments, label for labelling unit 21, comprising:

Semantic mapping module carries out Semantic mapping processing for the satellite information based on sample data, obtains sample data At least one corresponding classificating word；

Optionally, semantic mapping module, comprising:

Keyword extracting module carries out keyword extraction processing for the satellite information to sample data, obtains sample number According at least one corresponding keyword；

Keyword mapping block carries out Semantic mapping processing at least one keyword corresponding to sample data, obtains To at least one corresponding classificating word of sample data.

In the embodiments of the present disclosure, Semantic mapping can be carried out in several ways.Optionally, keyword mapping block, Include:

Crucial vector module determines that each keyword is corresponding at least one keyword for utilizing term vector model Crucial vector；

Classificating word determining module, for the crucial vector based on each keyword at least one keyword, from multiple pre- If classificating word in determine corresponding at least one classificating word of at least one keyword.

Optionally, classificating word determining module is specifically used for corresponding class vector in multiple preset classificating words and closes The shortest classificating word of distance is determined as the corresponding classificating word of keyword between the crucial vector of keyword.

Optionally, keyword mapping block, comprising: keyword filtering module is used for corresponding to sample data at least one A keyword is filtered operation, obtains at least one corresponding target keyword of sample data；

Keyword mapping block is specifically used at least one target keyword corresponding to sample data and carries out Semantic mapping Processing, obtains at least one corresponding classificating word of sample data.

Optionally, keyword filtering module is specifically used for specific in response at least one corresponding keyword of sample data For in multiple keywords and multiple keywords comprising that there are at least two first of mutex relation and/or inclusion relation is crucial Word determines target keyword from least two first keywords.

Optionally, keyword filtering module, comprising:

Keyword degree of association module, for determining at least two first keywords each first keyword and at least one The degree of association between second keyword, wherein the second keyword be multiple keywords in addition at least two first keywords Keyword；

Optionally, keyword degree of association module, specifically for being determined at least two first keywords based on neural network The degree of association of each first keyword and at least one the second keyword, neural network are got by the training of sample word, sample This word collection includes at least two sample words, and each sample word is labeled with the degree of association with other sample words.

Target keyword module, for based on each first keyword at least two first keywords and at least one the The degree of association between two keywords determines target keyword from least two first keywords.

Optionally, target keyword module, be specifically used for by least two first keywords at least one second close Maximum first keyword of the degree of association between keyword is as target keyword.

Label for labelling unit 21 can also include: classificating word module, for being based at least one classificating word, determine sample number According to mark label.

Optionally, classificating word module is specifically used for based on corresponding first weighted value of target word and at least one classificating word Corresponding second weighted value determines the mark label of sample data.

In some embodiments, the ratio-dependent based on the first weighted value and at least one the second weighted value marks label, The mark label includes the target word of the first ratio and at least one classificating word of the second ratio, wherein the first ratio and second Ratio can be preset, alternatively, the first ratio a and the second ratio b can be obtained respectively by above formula (1) and (2)；This When, optionally, when being trained based on mark label to sorter network, loss function value is to subtract the first ratio based on predicted value Example is determined multiplied by target word and the second ratio multiplied by least one classificating word.

Alternatively, optionally, classificating word module, specifically at least one classificating word to be determined as to the mark mark of sample data Label, wherein the mark label of sample data further includes target word；

Network training unit 22 is specifically used for based on corresponding first power of the corresponding first error of target word and first error Weight values and corresponding second error of at least one classificating word and corresponding second weighted value of the second error, determine loss function Value；Based on loss function value, processing is adjusted to the parameter value of sorter network.

In one or more optional embodiments, further includes:

Pre-training unit, for training sorter network in advance based on multiple first sample data in multiple sample datas, Sorter network after obtaining pre-training；

Network training unit 22, specifically for the mark label based on multiple sample datas and multiple sample datas, to pre- Sorter network after training is trained.

Optionally, further includes:

Pretreatment unit, for by the sample data in multiple sample datas with the maximum preset quantity of the target word degree of correlation As multiple first sample data.

Fig. 3 is one exemplary flow chart of data mask method that the embodiment of the present disclosure provides.

Step 310, the satellite information of target word associated multiple sample datas and multiple sample datas is obtained.

Optionally, target word can be term, carries out retrieval based on target word and obtains sample data and its satellite information, Wherein sample data can be web data, and at this time optionally, satellite information can include but is not limited to following at least one: mark Inscribe information, content description information, comment information, contextual information.Wherein heading message can be the corresponding title of webpage, content Pair that description information can be the description content for web page contents, comment information is the acquisitions such as other related web pages from network The comment content of the web data and contextual information can be based on heading message and/or content description information, and/or comment There is associated content with the web data by acquisition of information, the specific manifestation form of these satellite informations includes but is not limited to Literal expression.

In one or more optional embodiments, network retrieval is carried out based on target word, obtains the retrieval knot of target word Fruit, search result include the satellite information of multiple sample datas and multiple sample datas.

It is retrieved from network based on target word, obtains sample data relevant to target word and sample data is corresponding attached Information, specifically, the sample data of acquisition can include but is not limited to Web page picture, and a feature of Web page picture is exactly to include A large amount of satellite information, social network sites user can take some information in uploading pictures, and the present embodiment is from these attached letters The mark label that can be used for deep learning training is obtained in breath.A benefit for obtaining web data based on target word and search exists In can more fully excavate information therein although the information of data does not format richer；Be conducive to train Sorter network makes trained sorter network have stronger robustness.

Step 320, the satellite information based on multiple sample datas obtains the mark label of multiple sample datas.

The mark label that sample data is determined by satellite information specifically can be attached by what is expressed from written form Some keywords are obtained in information optionally can also be using target word as some as some mark labels of sample data The mark label of sample data；Each sample data is set to correspond at least one mark label by marking label, it can be to sample number According to more being described, more information is provided for training sorter network.

Based on a kind of disclosure data mask method provided by the above embodiment, the associated multiple sample numbers of target word are obtained The mark of multiple sample datas is obtained based on the satellite information of multiple sample datas according to the satellite information with multiple sample datas Label；The mark label that sample data is obtained based on multi-source trains sorter network based on the sample data with mark label, For obtained sorter network when handling image data, accuracy is higher.

In one or more optional embodiments, step 310 may include:

Optionally, keyword extraction processing is carried out to the satellite information of sample data, it is corresponding at least obtains sample data One keyword；

At least one keyword corresponding to sample data carries out Semantic mapping processing, and it is corresponding at least to obtain sample data One classificating word.

Satellite information is usually text information, carries out keyword to these satellite informations by keyword extraction techniques and mentions It takes, obtains at least one corresponding keyword of satellite information, wherein keyword extraction techniques can be any one keyword and mention Method is taken, such as: the method extracted the method TF-IDF of keyword based on word frequency or extract keyword using topic model Topic-model；The disclosure to the specific method of keyword extraction with no restrictions.

Optionally, at least one keyword corresponding to sample data carries out Semantic mapping processing, obtains sample data pair At least one classificating word answered, comprising:

In the present embodiment, Semantic mapping can using training after term vector word2vec model, by keyword according to It is semantic to establish mapping relations at least one preset classificating word, specifically, be during the foundation mapping of term vector model It is established based on vector, therefore, the corresponding crucial vector sum of keyword need to be obtained and preset the corresponding class vector of classificating word, specifically Ground, by the shortest classificating word of distance is true between corresponding class vector and the crucial vector of keyword in multiple preset classificating words It is set to the corresponding classificating word of keyword.Class vector classificating word corresponding with the shorter explanation of the distance between crucial vector and corresponding The keyword meaning of a word it is more close, by crucial vector and its apart from shortest class vector, that is, can determine its corresponding classificating word most Close to the corresponding classificating word of the class vector.

The present embodiment passes through filter operation, and the noise data at least one keyword is found out and is classified as correctly Classification is based on carrying out model training with noise data in other modes relatively, or whole noises in data is all deleted After carry out model training, the present embodiment is by being referred to correct classification for noise data, to noise and without the data of mark When being classified, obvious improves classification results.

Step 310 can also include: the mark label that sample data is determined based at least one classificating word.

In the embodiment, the ratio-dependent based on the first weighted value and at least one the second weighted value marks label, the mark The composition for infusing label includes the target word of the first ratio and the classificating word of at least one the second ratio, wherein the first ratio a can be with It is obtained by above-mentioned formula (1), and the second ratio b can be obtained by above-mentioned formula (2), the mark label based on the embodiment When being trained to sorter network, loss function be equal to predicted value subtract the first ratio multiplied by target word and at least one second Ratio is multiplied by classificating word.

Alternatively, optionally, the mark label of sample data includes multiple labels；

Step 320 may include:

At least one classificating word and target word are determined as to the mark label of sample data.

Fig. 4 is the structural schematic diagram for the data annotation equipment that the embodiment of the present disclosure provides.The device of the embodiment can be used for Realize the above-mentioned each method embodiment of the disclosure.As shown in figure 4, the device of the embodiment includes:

Information acquisition unit 41, for obtaining the attached letter of target word associated multiple sample datas and multiple sample datas Breath；

In one or more optional embodiments, information acquisition unit 41 can be used for carrying out network based on target word Retrieval, obtains the search result of target word, search result includes the satellite information of multiple sample datas and multiple sample datas.

Label for labelling unit 21 obtains the mark of multiple sample datas for the satellite information based on multiple sample datas Label.

Based on a kind of disclosure data annotation equipment provided by the above embodiment, the associated multiple sample numbers of target word are obtained The mark of multiple sample datas is obtained based on the satellite information of multiple sample datas according to the satellite information with multiple sample datas Label；The mark label that sample data is obtained based on multi-source trains sorter network based on the sample data with mark label, For obtained sorter network when handling image data, accuracy is higher.

In one or more optional embodiments, label for labelling unit 21, comprising:

Optionally, semantic mapping module, comprising:

Optionally, keyword filtering module, comprising:

Alternatively, optionally, classificating word module, specifically at least one classificating word to be determined as to the mark mark of sample data Label, wherein the mark label of sample data further includes target word

According to the other side of the embodiment of the present disclosure, a kind of electronic equipment provided, including processor, processor include Data mark described in sorter network training device described in any of the above-described embodiment of the disclosure or any of the above-described embodiment of the disclosure Dispensing device.

And processor, for being communicated with memory to execute executable instruction to complete any of the above-described implementation of the disclosure Sorter network training method described in example, or for being communicated with memory to execute executable instruction to complete the disclosure State data mask method described in any embodiment.

According to the other side of the embodiment of the present disclosure, a kind of computer storage medium provided, for storing computer The instruction that can be read, when instruction is executed by processor, which executes any of the above-described classification as described in the examples of the disclosure Network training method or any of the above-described data mask method as described in the examples.

According to the other side of the embodiment of the present disclosure, a kind of computer program product provided, including it is computer-readable Code, when computer-readable code is run in equipment, the processor in equipment is executed in any of the above-described embodiment of the disclosure The sorter network training method or any of the above-described data mask method as described in the examples.

According to another aspect of the embodiment of the present disclosure, another computer program product provided is calculated for storing Machine readable instruction, described instruction is performed so that computer executes classification net described in any of the above-described possible implementation The operation of network training method or data mask method.

In one or more optional embodiments, the embodiment of the present disclosure additionally provides a kind of computer program program production Product, for storing computer-readable instruction, described instruction is performed so that computer executes described in any of the above-described embodiment Sorter network training method or any of the above-described data mask method as described in the examples operation.

The computer program product can be realized especially by hardware, software or its mode combined.In an alternative embodiment In son, the computer program product is embodied as computer storage medium, in another optional example, the computer Program product is embodied as software product, such as software development kit (Software Development Kit, SDK) etc..

Another method for tracking target and its corresponding device and electronic equipment, meter are additionally provided according to the embodiment of the present disclosure Calculation machine storage medium and computer program product, wherein this method comprises: obtaining the spy of multiple reference pictures of target image Sign；Based on the feature of multiple reference pictures, multiple initial predicted positions that target is tracked in target image are determined；Based on multiple first Beginning predicted position determines the final position that target is tracked in target image.

In some embodiments, target following instruction can be specially call instruction, and first device can pass through calling Mode indicate second device performance objective track, accordingly, in response to call instruction is received, second device can be executed State the step and/or process in any embodiment in method for tracking target.

It should be understood that the terms such as " first " in the embodiment of the present disclosure, " second " are used for the purpose of distinguishing, and be not construed as Restriction to the embodiment of the present disclosure.

It should also be understood that in the disclosure, " multiple " can refer to two or more, "at least one" can refer to one, Two or more.

It should also be understood that clearly being limited or no preceding for the either component, data or the structure that are referred in the disclosure In the case where opposite enlightenment given hereinlater, one or more may be generally understood to.

It should also be understood that the disclosure highlights the difference between each embodiment to the description of each embodiment, Same or similar place can be referred to mutually, for sake of simplicity, no longer repeating one by one.

The embodiment of the present disclosure additionally provides a kind of electronic equipment, such as can be mobile terminal, personal computer (PC), puts down Plate computer, server etc..Below with reference to Fig. 5, it illustrates the terminal device or the services that are suitable for being used to realize the embodiment of the present application The structural schematic diagram of the electronic equipment 500 of device: as shown in figure 5, electronic equipment 500 includes one or more processors, communication unit For example Deng, one or more of processors: one or more central processing unit (CPU) 501, and/or one or more figures As processor (GPU) 513 etc., processor can according to the executable instruction being stored in read-only memory (ROM) 502 or from Executable instruction that storage section 508 is loaded into random access storage device (RAM) 503 and execute various movements appropriate and place Reason.Communication unit 512 may include but be not limited to network interface card, and the network interface card may include but be not limited to IB (Infiniband) network interface card.

Processor can with communicate in read-only memory 502 and/or random access storage device 503 to execute executable instruction, It is connected by bus 504 with communication unit 512 and is communicated through communication unit 512 with other target devices, to completes the application implementation The corresponding operation of any one method that example provides, for example, the satellite information based on the associated multiple sample datas of target word, determines The mark label of multiple sample datas；Mark label based on multiple sample datas and multiple sample datas, training sorter network.

In addition, in RAM 503, various programs and data needed for being also stored with device operation.CPU501,ROM502 And RAM503 is connected with each other by bus 504.In the case where there is RAM503, ROM502 is optional module.RAM503 storage Executable instruction, or executable instruction is written into ROM502 at runtime, executable instruction executes central processing unit 501 The corresponding operation of above-mentioned communication means.Input/output (I/O) interface 505 is also connected to bus 504.Communication unit 512 can integrate Setting, may be set to be with multiple submodule (such as multiple IB network interface cards), and in bus link.

I/O interface 505 is connected to lower component: the importation 506 including keyboard, mouse etc.；It is penetrated including such as cathode The output par, c 507 of spool (CRT), liquid crystal display (LCD) etc. and loudspeaker etc.；Storage section 508 including hard disk etc.； And the communications portion 509 of the network interface card including LAN card, modem etc..Communications portion 509 via such as because The network of spy's net executes communication process.Driver 510 is also connected to I/O interface 505 as needed.Detachable media 511, such as Disk, CD, magneto-optic disk, semiconductor memory etc. are mounted on as needed on driver 510, in order to read from thereon Computer program be mounted into storage section 508 as needed.

It should be noted that framework as shown in Figure 5 is only a kind of optional implementation, it, can root during concrete practice The component count amount and type of above-mentioned Fig. 5 are selected, are deleted, increased or replaced according to actual needs；It is set in different function component It sets, separately positioned or integrally disposed and other implementations, such as the separable setting of GPU513 and CPU501 or can also be used GPU513 is integrated on CPU501, the separable setting of communication unit 512 can also be integrally disposed on CPU501 or GPU513, etc. Deng.These interchangeable embodiments each fall within protection scope disclosed in the disclosure.

Particularly, in accordance with an embodiment of the present disclosure, it may be implemented as computer above with reference to the process of flow chart description Software program.For example, embodiment of the disclosure includes a kind of computer program product comprising be tangibly embodied in machine readable Computer program on medium, computer program include the program code for method shown in execution flow chart, program code It may include the corresponding instruction of corresponding execution method and step provided by the embodiments of the present application, for example, associated multiple based on target word The satellite information of sample data determines the mark label of multiple sample datas；Based on multiple sample datas and multiple sample datas Mark label, training sorter network.In such embodiments, which can be by communications portion 509 from net It is downloaded and installed on network, and/or is mounted from detachable media 511.In the computer program by central processing unit (CPU) When 501 execution, the operation for the above-mentioned function of limiting in the present processes is executed.

Disclosed method and device, equipment may be achieved in many ways.For example, software, hardware, firmware can be passed through Or any combination of software, hardware, firmware realizes disclosed method and device, equipment.The step of for method Sequence is stated merely to be illustrated, the step of disclosed method is not limited to sequence described in detail above, unless with other Mode illustrates.In addition, in some embodiments, the disclosure can be also embodied as recording program in the recording medium, this A little programs include for realizing according to the machine readable instructions of disclosed method.Thus, the disclosure also covers storage for holding The recording medium gone according to the program of disclosed method.

The description of the disclosure is given for the purpose of illustration and description, and is not exhaustively or by the disclosure It is limited to disclosed form.Many modifications and variations are obvious for the ordinary skill in the art.It selects and retouches Embodiment is stated and be the principle and practical application in order to more preferably illustrate the disclosure, and those skilled in the art is enable to manage The solution disclosure is to design various embodiments suitable for specific applications with various modifications.

Claims

1. a kind of sorter network training method characterized by comprising

Based on the satellite information of the associated multiple sample datas of target word, the mark label of the multiple sample data is determined；

2. the method according to claim 1, wherein determination is more in the satellite information based on multiple sample datas Before the mark label of a sample data, further includes:

Network retrieval is carried out based on the target word, obtains the search result of the target word, the search result includes described The satellite information of multiple sample datas and the multiple sample data.

3. method according to claim 1 or 2, which is characterized in that described to be based on the associated multiple sample datas of target word Satellite information, determine the mark label of the multiple sample data, comprising:

Satellite information based on the sample data carries out Semantic mapping processing, obtain the sample data it is corresponding at least one Classificating word；

4. a kind of data mask method characterized by comprising

5. a kind of sorter network training device characterized by comprising

Label for labelling unit determines the multiple sample for the satellite information based on the associated multiple sample datas of target word The mark label of data；

Network training unit, for the mark label based on the multiple sample data and the multiple sample data, training point Class network.

6. a kind of data annotation equipment characterized by comprising

Information acquisition unit, for obtaining the attached letter of target word associated multiple sample datas and the multiple sample data Breath；

Label for labelling unit obtains the mark of the multiple sample data for the satellite information based on the multiple sample data Infuse label.

7. a kind of electronic equipment, which is characterized in that including processor, the processor includes classification net described in claim 5 Network training device or data annotation equipment as claimed in claim 6.

8. a kind of electronic equipment characterized by comprising memory, for storing executable instruction；

And processor, it completes claims 1 to 3 to execute the executable instruction for being communicated with the memory and appoints The operation for a sorter network training method of anticipating, alternatively, for being communicated with the memory to execute the executable finger Enable completing the operation of data mask method described in claim 4.

9. a kind of computer storage medium, for storing computer-readable instruction, which is characterized in that described instruction is performed When perform claim require the behaviour of data mask method described in sorter network training method or claim 4 described in 1 to 3 any one Make.

10. a kind of computer program product, including computer-readable code, which is characterized in that when the computer-readable code When running in equipment, the processor in the equipment is executed for realizing sorter network described in claims 1 to 3 any one The instruction of data mask method described in training method or claim 4.