CN108763246A

CN108763246A - Personnel's group technology and device, storage medium, electronic equipment

Info

Publication number: CN108763246A
Application number: CN201810273041.6A
Authority: CN
Inventors: 盛志超; 刘欣语
Original assignee: iFlytek Co Ltd
Current assignee: iFlytek Co Ltd
Priority date: 2018-03-29
Filing date: 2018-03-29
Publication date: 2018-11-06
Anticipated expiration: 2038-03-29
Also published as: CN108763246B

Abstract

A kind of personnel's group technology of disclosure offer and device, storage medium, electronic equipment.This method includes：The concern information is split as at least one phrase by the concern information for obtaining personnel to be allocated, and the vectorization for obtaining each phrase indicates；Phrase-based vectorization indicates progress clustering processing, obtains M₁A first category, each first category indicate a kind of dimensional information, M₁≥1；The corresponding weight of each dimensional information is set, using the dimensional information and the corresponding weight of the dimensional information, obtains the group result of the personnel to be allocated.Such scheme helps to improve accuracy, the reasonability of group result, and then improves satisfaction of the personnel to be allocated to group result.

Description

Personnel's group technology and device, storage medium, electronic equipment

Technical field

This disclosure relates to technical field of information processing, and in particular, to a kind of personnel's group technology and device, storage are situated between Matter, electronic equipment.

Background technology

In daily life, personnel are frequently encountered and are grouped problem.For example, school, enterprise distribute for student or employee Dormitory, class divide study group for student, and department divides work group etc. for employee, all refer to personnel's grouping and ask Topic.

In order to improve the reasonability of grouping, it will usually pre-set some and be grouped relevant dimensional information, further according to this A little dimensional informations carry out clustering, obtain final group result.By taking school distributes students' dormitory as an example, the possible dimension considered Degree information has：Grade, department, gender, age, daily schedule, hobby, can be taking human as the weighted value that each dimensional information is arranged, meter Mathematics gives birth to the weighted sum under each dimensional information, and carries out clustering processing to student according to weighted sum, and obtained cluster result is i.e. For the group result of dormitory distribution.

It is so assigned scheme, it will usually choose some general informations as dimensional information, dimensional information is relatively more fixed, and not Consider that personnel to be allocated compare the individual character dimensional information of concern, causes the accuracy of group result, reasonability relatively low, influence personnel To the satisfaction of group result.

Invention content

It is a general object of the present disclosure to provide a kind of personnel's group technology and device, storage medium, electronic equipments, contribute to Accuracy, the reasonability of group result are improved, and then improves satisfaction of the personnel to be allocated to group result.

To achieve the goals above, the disclosure provides a kind of personnel's group technology, the method includes：

The concern information is split as at least one phrase by the concern information for obtaining personnel to be allocated, is obtained each short The vectorization of language indicates；

Phrase-based vectorization indicates progress clustering processing, obtains M₁A first category, each first category indicate one Kind dimensional information, M₁≥1；

The corresponding weight of each dimensional information is set, using the dimensional information and the corresponding weight of the dimensional information, Obtain the group result of the personnel to be allocated.

Optionally, the vectorization for obtaining each phrase indicates, including：

The initial vectorization for obtaining each phrase indicates, and phrase-based initial vectorization indicates to carry out clustering processing, Obtain M₂A second category, M₂>1；

It will be from the M₂The first sample phrase selected in a second category, combination of two are first sample phrase pair, And the markup information of each first sample phrase pair is obtained, the markup information is similar or dissimilar；

Using the first sample phrase pair and the markup information of the first sample phrase pair, training obtains phrase classification Model, the phrase classification model include the phrase expression layer for carrying out vectorization processing；

The phrase that the concern information is split out exports new vectorization table through the phrase expression layer as input Show, the vectorization as the phrase indicates.

Optionally, after the new vectorization for obtaining the phrase indicates, the vectorization for obtaining each phrase indicates Further include：

Phrase-based new vectorization indicates progress clustering processing, obtains M₃A third classification, M₃>1；

Poor (the d of selected distance₂-d₁) minimum N₁A second sample phrase, d₁For the second sample phrase and nearest the The distance at the other class center of three classes, d₂It is the second sample phrase at a distance from the class center of time close third classification；

By the class center of the second sample phrase and nearest third classification and/or the second sample phrase with time The class center of close third classification, combination of two is the second sample phrase pair, and obtains the mark of each second sample phrase pair Information；

Using the second sample phrase pair and the markup information of the second sample phrase pair, the phrase classification is updated Model, until updated phrase classification model meets preset condition, the updated phrase classification model includes more Phrase expression layer after new；

Using the phrase that splits out of concern information as input, through the updated phrase expression layer output update to Quantization means, the vectorization as the phrase indicate.

Optionally, the method further includes：

Obtain vectorization of the personnel to be allocated for the feedback information and every feedback information of the group result It indicates；

Vectorization based on feedback information indicates progress clustering processing, obtains M₄A 4th classification, each 4th classification pair Answer a kind of satisfaction grade, M₄≥1；

Judge the M₄Whether include classification to be adjusted in a 4th classification, corresponding satisfaction of the classification to be adjusted etc. For indicating that the personnel to be adjusted for belonging to the classification to be adjusted are unsatisfied with the group result, the personnel to be adjusted belong to grade The personnel to be allocated；

If the M₄A 4th classification includes the classification to be adjusted, then obtains the grouping of the personnel to be adjusted more New information；

Using the grouping fresh information, the group result of the personnel to be adjusted is adjusted.

Optionally, the vectorization for obtaining every feedback information indicates, including：

The initial vectorization for obtaining every feedback information indicates, and is gathered based on the expression of the initial vectorization of feedback information Class processing, obtains M₅A 5th classification, M₅>1；

From the M₅First sample feedback information is chosen in a 5th classification, and marks each first sample feedback information Satisfaction grade；

Using the satisfaction grade of the first sample feedback information and the first sample feedback information, training is expired Meaning degree disaggregated model, the Satisfaction index class model include the text representation layer for carrying out vectorization processing；

Using the feedback information as input, exports new vectorization through the text representation layer and indicate, as described anti- The vectorization of feedforward information indicates.

Optionally, after the new vectorization for obtaining the feedback information indicates, it is described obtain every feedback information to Quantization means further include：

New vectorization based on feedback information indicates progress clustering processing, obtains M₆A 6th classification, M₆>1；

Poor (the d of selected distance₄-d₃) minimum N₂A second sample back information, and mark each second sample back information Satisfaction grade, d₃It is the second sample back information at a distance from the class center of the 6th nearest classification, d₄It is anti-for the second sample Feedforward information is at a distance from the class center of secondary the 6th close classification；

The satisfaction etc. of satisfaction grade, the second sample back information based on the first sample feedback information The satisfaction of grade and the remaining feedback information in addition to the first sample feedback information and the second sample back information Grade calculates update class center, and the satisfaction grade of the residue feedback information is by the classification belonging to the remaining feedback information It determines；

Based on the update class center, clustering processing is carried out to the feedback information again, determines every feedback information Affiliated more new category, each more new category correspond to a update class center；

Using the more new category belonging to the feedback information and the feedback information, the satisfaction classification mould is updated Type, until updated Satisfaction index class model meets predetermined condition, the updated Satisfaction index class model includes Updated text representation layer；

Using the feedback information as input, indicates, make through the updated text representation layer output renewal vectorization It is indicated for the vectorization of the feedback information.

Optionally, described to obtain M₆After a 6th classification, the vectorization expression for obtaining every feedback information further includes：

Determine M₇A available class center, the M₇A available class center and the M₆Between the class center of a 6th classification Distance be not less than pre-determined distance；

Based on new class center, clustering processing is carried out to the feedback information again, is determined belonging to every feedback information New classification, each new classification corresponds to a new class center, and the new class center includes the M₇In a available class The heart and the M₆The class center of a 6th classification；

Then, the second sample back information is chosen in the following way：Poor (the d of selected distance₆-d₅) minimum N₂A Two sample back information, d₅It is the second sample back information at a distance from nearest new class center, d₆Believe for the second sample back It ceases at a distance from secondary close new class center.

The disclosure provides a kind of personnel's apparatus for grouping, and described device includes：

Pay close attention to information split module, the concern information for obtaining personnel to be allocated, by the concern information be split as to A few phrase；

Phrase vectorization indicates acquisition module, and the vectorization for obtaining each phrase indicates；

Phrase clustering processing module indicates to carry out clustering processing for phrase-based vectorization, obtains M₁A first kind Not, each first category indicates a kind of dimensional information, M₁≥1；

Group result obtains module, for the corresponding weight of each dimensional information to be arranged, using the dimensional information and The corresponding weight of the dimensional information obtains the group result of the personnel to be allocated.

Optionally, the phrase vectorization indicates that acquisition module includes：

Primary vectorization indicates acquisition module, and the initial vectorization for obtaining each phrase indicates；

First clustering processing module indicates to carry out clustering processing for phrase-based initial vectorization, obtains M₂A Two classifications, M₂>1；

First sample phrase chooses module, and being used for will be from the M₂The first sample phrase selected in a second category, Combination of two is first sample phrase pair, and obtains the markup information of each first sample phrase pair, and the markup information is phase It is seemingly or dissimilar；

Phrase classification model training module, for utilizing the first sample phrase pair and the first sample phrase pair Markup information, training obtain phrase classification model, and the phrase classification model includes the phrase table for carrying out vectorization processing Show layer；

Secondary vectorization indicates output module, and the phrase for splitting out the concern information is as input, through described Phrase expression layer exports new vectorization and indicates, the vectorization as the phrase indicates.

Optionally, the phrase vectorization indicates that acquisition module further includes：

Second clustering processing module is phrase-based new after being indicated in the new vectorization for obtaining the phrase Vectorization indicates progress clustering processing, obtains M₃A third classification, M₃>1；

Second sample phrase chooses module, is used for the poor (d of selected distance₂-d₁) minimum N₁A second sample phrase, d₁For The second sample phrase is at a distance from the class center of nearest third classification, d₂For the second sample phrase and time close the The distance at the other class center of three classes；By the class center and/or described second of the second sample phrase and nearest third classification The class center of sample phrase and secondary close third classification, combination of two is the second sample phrase pair, and obtains each second sample The markup information of phrase pair；

Phrase classification model modification module, for utilizing the second sample phrase pair and the second sample phrase pair Markup information updates the phrase classification model, until updated phrase classification model meets preset condition, it is described more Phrase classification model after new includes updated phrase expression layer；

Third vectorization indicates output module, and the phrase for splitting out the concern information is as input, through described Updated phrase expression layer output renewal vectorization indicates that the vectorization as the phrase indicates.

Optionally, described device further includes：

Feedback information acquisition module is directed to the feedback information of the group result for obtaining the personnel to be allocated；

Feedback information vectorization indicates acquisition module, and the vectorization for obtaining every feedback information indicates；

Feedback information clustering processing module indicates to carry out clustering processing for the vectorization based on feedback information, obtains M₄ A 4th classification, each 4th classification correspond to a kind of satisfaction grade, M₄≥1；

Classification judgment module to be adjusted, for judging the M₄Whether include classification to be adjusted in a 4th classification, it is described The personnel to be adjusted that the corresponding satisfaction grade of classification to be adjusted is used to indicate to belong to the classification to be adjusted dissatisfied described divide Group is as a result, the personnel to be adjusted belong to the personnel to be allocated；

It is grouped fresh information acquisition module, in the M₄When a 4th classification includes the classification to be adjusted, obtain Take the grouping fresh information of the personnel to be adjusted；

Group result adjusts module, for utilizing the grouping fresh information, adjusts the grouping knot of the personnel to be adjusted Fruit.

Optionally, the feedback information vectorization indicates that acquisition module includes：

Four-way quantization means acquisition module, the initial vectorization for obtaining every feedback information indicate；

Third clustering processing module indicates to carry out clustering processing for the initial vectorization based on feedback information, obtains M₅ A 5th classification, M₅>1；

First sample feedback information chooses module, is used for from the M₅First sample feedback letter is chosen in a 5th classification Breath, and mark the satisfaction grade of each first sample feedback information；

Satisfaction index class model training module, for being fed back using the first sample feedback information and the first sample The satisfaction grade of information, training obtain Satisfaction index class model, and the Satisfaction index class model includes for carrying out vectorization The text representation layer of processing；

Five-way quantization means output module is used for using the feedback information as input, defeated through the text representation layer Go out new vectorization to indicate, the vectorization as the feedback information indicates.

Optionally, the feedback information vectorization indicates that acquisition module further includes：

4th clustering processing module is based on feedback letter after being indicated in the new vectorization for obtaining the feedback information The new vectorization of breath indicates progress clustering processing, obtains M₆A 6th classification, M₆>1；

Second sample back information chooses module, is used for the poor (d of selected distance₄-d₃) minimum N₂A second sample back letter Breath, and mark the satisfaction grade of each second sample back information, d₃For the second sample back information and the 6th nearest class The distance at other class center, d₄It is the second sample back information at a distance from the class center of the 6th time close classification；

Update class center calculation module, for based on the first sample feedback information satisfaction grade, described second The satisfaction grade of sample back information and in addition to the first sample feedback information and the second sample back information Remaining feedback information satisfaction grade, calculate update class center, the satisfaction grade of the residue feedback information is by described Classification belonging to remaining feedback information determines；

5th clustering processing module again carries out at cluster the feedback information for being based on the update class center Reason, determines the more new category belonging to every feedback information, and each more new category corresponds to a update class center；

Satisfaction index class model update module, for utilizing the update belonging to the feedback information and the feedback information Classification updates the Satisfaction index class model, until updated Satisfaction index class model meets predetermined condition, it is described more Satisfaction index class model after new includes updated text representation layer；

Six-way quantization means output module is used for using the feedback information as input, through the updated text Expression layer exports renewal vectorization and indicates, the vectorization as the feedback information indicates.

Class center determining module can be used, for obtaining M described₆After a 6th classification, M is determined₇A available class center, The M₇A available class center and the M₆The distance between class center of a 6th classification is not less than pre-determined distance；

6th clustering processing module, for based on new class center, carrying out clustering processing to the feedback information again, really The new classification belonging to every feedback information is made, each new classification corresponds to a new class center, the new class center Including the M₇A available class center and the M₆The class center of a 6th classification；

Then, the second sample back information chooses module, is used for the poor (d of selected distance₆-d₅) minimum N₂A second sample This feedback information, d₅It is the second sample back information at a distance from nearest new class center, d₆For the second sample back information with The distance at secondary close new class center.

The disclosure provides a kind of storage medium, wherein being stored with a plurality of instruction, described instruction is loaded by processor, in execution The step of stating personnel's group technology.

The disclosure provides a kind of electronic equipment, and the electronic equipment includes；

Above-mentioned storage medium；And

Processor, for executing the instruction in the storage medium.

In disclosure scheme, the concern information of personnel to be allocated can be analyzed, determines more embody in a manner of cluster and wait for The dimensional information of personnel demand is distributed, and then is grouped into administrative staff based on the dimensional information determined, compared with the existing technology only By commonly using the scheme that be grouped into administrative staff of dimensional information, disclosure scheme helps to improve the accuracy, reasonable of group result Property, and then improve satisfaction of the personnel to be allocated to group result.

Other feature and advantage of the disclosure will be described in detail in subsequent specific embodiment part.

Description of the drawings

Attached drawing is for providing further understanding of the disclosure, and a part for constitution instruction, with following tool Body embodiment is used to explain the disclosure together, but does not constitute the limitation to the disclosure.In the accompanying drawings：

Fig. 1 is the flow diagram of disclosure copywriter's group technology embodiment 1；

Fig. 2 is the flow diagram for the vectorization expression embodiment 1 that phrase is obtained in disclosure scheme；

Fig. 3 is the network diagram of phrase classification model in disclosure scheme；

Fig. 4 is the flow diagram for the vectorization expression embodiment 2 that phrase is obtained in disclosure scheme；

Fig. 5 is the flow diagram of disclosure copywriter's group technology embodiment 2；

Fig. 6 is the flow diagram for the vectorization expression embodiment 1 that feedback information is obtained in disclosure scheme；

Fig. 7 is the network diagram of Satisfaction index class model in disclosure scheme；

Fig. 8 is the corresponding node schematic diagram of intermediate node in disclosure scheme；

Fig. 9 is the flow diagram for the vectorization expression embodiment 2 that feedback information is obtained in disclosure scheme；

Figure 10 is the composition schematic diagram of disclosure copywriter's apparatus for grouping；

Figure 11 is structural schematic diagram of the disclosure scheme for the electronic equipment of personnel's grouping.

Specific implementation mode

The specific implementation mode of the disclosure is described in detail below in conjunction with attached drawing.It should be understood that this place is retouched The specific implementation mode stated is only used for describing and explaining the disclosure, is not limited to the disclosure.

Referring to Fig. 1, the flow diagram of disclosure personnel's group technology embodiment 1 is shown.It may comprise steps of：

S101 obtains the concern information of personnel to be allocated, and the concern information is split as at least one phrase, obtains every The vectorization of a phrase indicates.

In order to improve accuracy, the reasonability of group result, the letter that disclosure scheme can be paid close attention to based on personnel to be allocated Breath carries out dimensional analysis, determines the dimensional information that can more embody personnel demand to be allocated.It is to be appreciated that disclosure scheme The dimensional information determined be compared with the prior art in common dimensional information for, not represent single personnel to be allocated Property demand dimensional information, that is, general character that the dimensional information that disclosure scheme is determined, which is also personnel to be allocated, to be had needs It asks, only the prior art may be as the factor for influencing group result.

As an example, the concern information of personnel to be allocated, disclosure side can by way of questionnaire survey, be obtained Case can be not specifically limited this.After obtaining concern information, deconsolidation process can be carried out to concern information, obtained at least one short Language finds the dimensional information for embodying personnel demand to be allocated by analyzing these phrases.It is to be appreciated that actually answering With in the process, when carrying out phrase fractionation, some stop words for itself having no meaning can be filtered out, such as " ", " " Deng disclosure scheme can be not specifically limited this.

S102, phrase-based vectorization indicate progress clustering processing, obtain M₁A first category, each first category table Show a kind of dimensional information, M₁≥1。

Disclosure scheme, based on this progress clustering processing, can obtain M after the vectorization for obtaining phrase indicates₁A One classification, that is, find M₁A dimensional information.For example, it can be carried out at cluster by k-means algorithms, k nearest neighbor algorithm etc. Reason, disclosure scheme can be not specifically limited this.

In actual application, there may be a large amount of synonyms for the phrase split out from concern information, in order to carry The cluster accuracy of high synonymous phrase can be based on word embedding (Chinese：Word is embedded in) mode, obtain the vector of phrase Change and indicate, and then phrase-based vectorization indicates to carry out clustering processing.

As an example, the vectorization that each word that phrase includes can be first obtained based on general corpus is indicated； The mean value that the vectorization of each word is indicated again, the initial vectorization for being determined as phrase indicate, and then phrase-based initial Vectorization indicates progress clustering processing, obtains M₁A first category.

In addition, it is contemplated that having centainly between the corpus that the phrase that general corpus and concern information are split out is formed Difference, in order to improve the accuracy of phrase vectorization expression, disclosure scheme also provides a kind of vectorization of new acquisition phrase The method of expression wouldn't be described in detail herein for details, reference can be made to being introduced at following FIG. 2, Fig. 4.

The corresponding weight of each dimensional information is arranged in S103, corresponding using the dimensional information and the dimensional information Weight obtains the group result of the personnel to be allocated.

Based on the phrase that concern information is split out, M is obtained₁After a dimensional information, each dimensional information can be set and corresponded to Weight, and then be based on dimensional information and the corresponding weight calculation weighted sum of dimensional information, and treat assigner according to weighted sum Member carries out clustering processing, and obtained cluster result is the group result of personnel to be allocated.It as an example, can be taking human as setting The weighted value of each dimensional information is set, disclosure scheme can be not specifically limited this.

In actual application, the dimensional information that can be based only upon discovery is grouped into administrative staff, or can integrate hair Existing dimensional information and existing common dimensional information are grouped into administrative staff；It, can be with base for the dimensional information of discovery In M₁A dimensional information is grouped into administrative staff, or can be based on from M₁The partial dimensional information selected in a dimensional information into Administrative staff is grouped.Disclosure scheme using dimensional information into the specific implementation that administrative staff is grouped to that can not be limited, specifically It is determined in combination with practical application request.

To sum up, disclosure scheme determines M by the concern information of analysis personnel to be allocated in a manner of cluster₁It is a more The dimensional information of personnel demand to be allocated is embodied, the side being only grouped compared with the existing technology into administrative staff by commonly using dimensional information Case, disclosure scheme help to improve the accuracy of group result, reasonability, and then improve personnel to be allocated to group result Satisfaction.

Referring to Fig. 2, show that the disclosure obtains the flow diagram of the vectorization expression embodiment 1 of phrase.May include Following steps：

S201, the initial vectorization for obtaining each phrase indicate, and phrase-based initial vectorization expression is clustered Processing, obtains M₂A second category, M₂>1。

S202, will be from the M₂The first sample phrase selected in a second category, combination of two are that first sample is short Language pair, and the markup information of each first sample phrase pair is obtained, the markup information is similar or dissimilar.

According to introduction made above, obtained based on general corpus phrase initial vectorization indicate after, can be first with The initial vectorization of phrase indicates clustering processing of progress, obtains M₂A second category, it is contemplated that the initial vector of phrase The accuracy of expression is relatively low, thus obtained M₂The accuracy of a second category is relatively low.In view of this, the disclosure Scheme can train a phrase classification model for including phrase expression layer, by phrase expression layer to being split out in concern information Phrase carry out vectorization processing, can improve vectorization expression accuracy, and then improve cluster accuracy.

Specifically, the sample data of training phrase classification model may include：

(1) first sample phrase pair.It specifically, can be first from M₂First sample phrase is selected in a second category, then First sample phrase combination of two is obtained into first sample phrase pair.

(2) markup information of first sample phrase pair.Specifically, markup information can be presented as similar or dissimilar, It can be taking human as the markup information of setting first sample phrase pair.

It as an example, can be at random from M₂First sample phrase is chosen in a second category；Alternatively, can from positioned at The boundary of second category is easy in the phrase of aliasing, chooses first sample phrase.Specifically, each phrase can be calculated separately With the class center distance d of nearest second category_M21, each phrase and time close second category class center distance d_M22；It is logical Often, the corresponding range difference (d of phrase_M22-d_M21) bigger, it is less susceptible to occur aliasing between class, therefore can be according to (d_M22-d_M21) to phrase It is ranked up, chooses the phrase of a certain number of range difference minimums as first sample phrase.

Disclosure scheme is to choosing the mode of first sample phrase, the quantity of first sample phrase, first sample phrase pair Quantity can be not specifically limited, in order to ensure the data balancing of training sample, first that phrase in class forms can be made Sample phrase is suitable between the quantity for the first sample phrase pair that phrase, class forms, so that the phrase classification model that training obtains With preferable distinction.

As an example, phrase-based initial vectorization indicates to carry out clustering processing, can be embodied as automatic Cluster, i.e., do not limit the quantity of second category；Alternatively, in order to improve cluster efficiency, it can also be in such a way that super ginseng be set, in advance The quantity of second category is first determined, then is clustered.Disclosure scheme can not limit the specific implementation of clustering processing It is fixed.

S203, using the first sample phrase pair and the markup information of the first sample phrase pair, training obtains short Language disaggregated model, the phrase classification model include the phrase expression layer for carrying out vectorization processing.

As an example, network training phrase classification model shown in Fig. 3 can be utilized.Wherein, phrase expression layer includes Left and right two parts, this two parts network layer having the same, the weight between each network layer are also consistent, i.e. left and right two parts The input of shared parameter, model is first sample phrase pair, and the output of model is the markup information of first sample phrase pair.Citing For, the topological structure of phrase expression layer can be presented as CNN (English：Convolutional Neural Network, in Text：Convolutional neural networks), RNN (English：Recurrent Neural Network, Chinese：Recognition with Recurrent Neural Network), recurrence from Recursive Autoencoder etc. are encoded, disclosure scheme can be not specifically limited this.

Specifically, two phrases of first sample phrase centering input left and right two parts of phrase expression layer respectively, through volume After lamination, pond layer, full articulamentum, the vectorization for obtaining phrase indicates；Then, two exported phrase expression layer by series connection layer A vectorization indicates that result carries out splicing merging, then the markup information of the first sample phrase pair is exported through the layer that feedovers.It can manage Xie Di, when the markup information of the first sample phrase pair of phrase classification model output, when being consistent with the markup information being artificially arranged, It is considered that model training is completed.

S204, the phrase that the concern information is split out export new vector as input through the phrase expression layer Change and indicate, the vectorization as the phrase indicates.

Relative to the phrase classification model that general corpus, the phrase split out using concern information are trained, carry out The accuracy higher of vectorization processing, therefore the new vectorization that can export phrase expression layer indicates, the vector as phrase Change and indicates.Accordingly, the new vectorization that S102 can be phrase-based indicates progress clustering processing, obtains M₁A first category.

Referring to Fig. 4, show that the disclosure obtains the flow diagram of the vectorization expression embodiment 2 of phrase.May include Following steps：

S301, the initial vectorization for obtaining each phrase indicate, and phrase-based initial vectorization expression is clustered Processing, obtains M₂A second category, M₂>1。

S302, will be from the M₂The first sample phrase selected in a second category, combination of two are that first sample is short Language pair, and the markup information of each first sample phrase pair is obtained, the markup information is similar or dissimilar.

S303, using the first sample phrase pair and the markup information of the first sample phrase pair, training obtains short Language disaggregated model, the phrase classification model include the phrase expression layer for carrying out vectorization processing.

S304, the phrase that the concern information is split out export new vector as input through the phrase expression layer Change and indicates.

The realization process of step S301~S304, can refer to and introduced at S201~S204 above, details are not described herein again.

S305, phrase-based new vectorization indicate progress clustering processing, obtain M₃A third classification, M₃>1。

S306, the poor (d of selected distance₂-d₁) minimum N₁A second sample phrase, d₁For the second sample phrase and most The distance at the class center of close third classification, d₂For the second sample phrase and the class center of time close third classification away from From.

S307, by the class center and/or the second sample phrase of the second sample phrase and nearest third classification With the class center of secondary close third classification, combination of two is the second sample phrase pair, and obtains each second sample phrase pair Markup information.

In order to further increase the accuracy of phrase vectorization expression, disclosure scheme also provides a kind of optimization phrase classification The scheme of model.Specifically, the sample data of optimization phrase classification model may include：

(1) second sample phrase pair.

Specifically, new vectorization expression that can be first phrase-based carries out clustering processing, obtains M₃A third classification；So Afterwards from the phrase positioned at the boundary of third classification, easy aliasing, the second sample phrase is selected；Again by the second sample phrase With corresponding class center combination of two, the second sample phrase pair is obtained.

As an example, choosing the mode of the second sample phrase can be presented as：Calculate separately each phrase with it is nearest The distance d at the class center of third classification₁, each phrase and time close third classification class center distance d₂；In general, phrase corresponds to Range difference (d₂-d₁) bigger, it is less susceptible to occur aliasing between class, therefore can be according to (d₂-d₁) phrase is ranked up, it chooses most Small preceding N₁A phrase is as the second sample phrase.It in actual application, can be preferentially from addition to first sample phrase Phrase in choose the second sample phrase, disclosure scheme can be not specifically limited this.

It is described above to refer to the second sample phrase and corresponding class center combination of two by the second sample phrase and most The class center of close third classification, and/or by the class center of the second sample phrase and secondary close third classification, combination of two obtains Second sample phrase pair.

Furthermore, it is necessary to explanation, M₃A third classification can be carried out by automatic cluster, or the mode of the super ginseng of setting Clustering processing, disclosure scheme can be not specifically limited this.

The markup information of (2) second sample phrases pair.Specifically, markup information can be presented as similar or dissimilar, It can be taking human as the markup information of the second sample phrase pair of setting.

S308, using the second sample phrase pair and the markup information of the second sample phrase pair, update is described short Language disaggregated model, until updated phrase classification model meets preset condition, the updated phrase classification model Including updated phrase expression layer.

In disclosure scheme, meeting preset condition can be presented as：The number of phrase classification model modification iteration is not less than Preset times；Alternatively, indicating to carry out clustering processing using the renewal vectorization of updated phrase expression layer output, meet following Condition：(d₂-d₁) poor not less than pre-determined distance；Alternatively, (d₂-d₁) it is not less than the phrase quantity of pre-determined distance difference not less than default Quantity, etc., disclosure scheme do not limit preset condition, preset times, the poor, preset quantity of pre-determined distance etc., specifically may be used It is arranged in conjunction with practical application request.

S309, the phrase that the concern information is split out are exported as input through the updated phrase expression layer Renewal vectorization indicates that the vectorization as the phrase indicates.

Using the second sample phrase to optimizing updated phrase classification model, vectorization expression can be further increased Accuracy, therefore the renewal vectorization that can export updated phrase expression layer indicates, the vectorization as phrase indicates.It is right Ying Di, S102 can be phrase-based renewal vectorization indicate carry out clustering processing, obtain M₁A first category.

As an example, in order to further increase accuracy, the reasonability of group result, disclosure scheme also provides one Feedback information of the kind based on personnel to be allocated, adjusts the scheme of group result.Referring to Fig. 5, disclosure personnel grouping side is shown The flow diagram of method embodiment 2.It may comprise steps of：

S401 obtains the personnel to be allocated feedback information and every feedback information for the group result Vectorization indicates.

After obtaining the group result of personnel to be allocated according to method shown in Fig. 1, personnel to be allocated can also be obtained and be directed to and divided The feedback information of group result, and then group result is optimized based on feedback information, personnel to be allocated are further increased to group result Satisfaction.

As an example, the feedback letter of personnel to be allocated can be obtained by modes such as questionnaire survey, regular return visits Breath, disclosure scheme can be not specifically limited this.

S402, the vectorization based on feedback information indicate progress clustering processing, obtain M₄A 4th classification, each 4th class A kind of satisfaction grade, M are not corresponded to₄≥1。

Disclosure scheme, based on this progress clustering processing, can obtain M after the vectorization for obtaining feedback information indicates₄ A 4th classification, can be taking human as the corresponding satisfaction grade of each 4th classification be marked, for example, satisfaction grade can be full It anticipates, is general, dissatisfied, disclosure scheme can be not specifically limited the quantity of satisfaction grade.For example, k- can be passed through Means algorithms, k nearest neighbor algorithm etc. carry out clustering processing, and disclosure scheme can be not specifically limited this.

As an example, can the vectorization table for each word that feedback information includes first be obtained based on general corpus Show；The mean value that the vectorization of each word is indicated again, the initial vectorization for being determined as feedback information indicates, and then is based on feedback The initial vectorization of information indicates progress clustering processing, obtains M₄A 4th classification.

In addition, it is contemplated that there is certain difference between the corpus that general corpus and feedback information are formed, in order to carry The accuracy that high feedback information vectorization indicates, disclosure scheme also provide a kind of vectorization expression of new acquisition feedback information Method wouldn't be described in detail herein for details, reference can be made to being introduced at following FIG. 6, Fig. 9.

S403 judges the M₄Whether include classification to be adjusted in a 4th classification, the classification to be adjusted is corresponding full Meaning degree grade is for indicating that the personnel to be adjusted for belonging to the classification to be adjusted are unsatisfied with the group result, the people to be adjusted Member belongs to the personnel to be allocated.

Based on feedback information, M is obtained₄It, can be with after a 4th classification and the corresponding satisfaction grade of each 4th classification Judge whether include wherein classification to be adjusted.If there is no classification to be adjusted, illustrate that all personnel to be allocated are to shown in Fig. 1 The group result satisfaction that method obtains；Otherwise the group result that declaratives personnel to be allocated obtain method shown in Fig. 1 is discontented , that is, there are personnel to be adjusted in meaning.

For example, if cluster obtains 2 the 4th classifications, corresponding satisfaction grade is respectively：It is satisfied, dissatisfied, 4th classification of meaning with thumb down can be determined as classification to be adjusted.For example, if cluster obtains 4 the 4th classifications, Corresponding satisfaction grade is respectively：It is satisfied, general, be unsatisfied with, be not satisfied at all, can according to demand by meaning with thumb down and/ Or the 4th classification being not satisfied at all is determined as classification to be adjusted.Disclosure scheme can to the corresponding satisfaction grade of classification to be adjusted It does not limit, can specifically be determined by practical application request.

S404, if the M₄A 4th classification includes the classification to be adjusted, then obtains the personnel's to be adjusted It is grouped fresh information.

S405 adjusts the group result of the personnel to be adjusted using the grouping fresh information.

Through S403 judgements, there are the groupings that when classification to be adjusted, can obtain the personnel to be adjusted for belonging to the category to update letter It ceases, and then adjusts the group result of each personnel to be adjusted according to grouping fresh information, to improve personnel to be adjusted to group result Satisfaction.As an example, the grouping fresh information of personnel to be adjusted can be inputted by external staff, disclosure scheme pair This can be not specifically limited.

Referring to Fig. 6, show that the disclosure obtains the flow diagram of the vectorization expression embodiment 1 of feedback information.It can be with Include the following steps：

S501, the initial vectorization for obtaining every feedback information indicates, and is indicated based on the initial vectorization of feedback information Clustering processing is carried out, M is obtained₅A 5th classification, M₅>1。

S502, from the M₅First sample feedback information is chosen in a 5th classification, and marks each first sample feedback The satisfaction grade of information.

According to introduction made above, after the initial vectorization expression that feedback information is obtained based on general corpus, Ke Yixian It indicates to carry out a clustering processing using the initial vectorization of feedback information, obtains M₅A 5th classification, it is contemplated that feedback information Initial vectorization indicate accuracy it is relatively low, thus obtained M₅The accuracy of a 5th classification is relatively low.Needle For this, disclosure scheme can train a Satisfaction index class model for including text representation layer, right by text representation layer Feedback information carries out vectorization processing, can improve the accuracy of vectorization expression, and then improve cluster accuracy.

Specifically, the sample data of training Satisfaction index class model may include：

(1) first sample feedback information.It specifically, can be from M₅First sample feedback letter is selected in a 5th classification Breath.

(2) the satisfaction grade of first sample feedback information.Specifically, satisfaction grade can be presented as it is satisfied, general, At least one of be unsatisfied with, be not satisfied at all, it can be taking human as the satisfaction grade of setting first sample feedback information.

It as an example, can be at random from M₅First sample feedback information is chosen in a 5th classification；Alternatively, can be from In the boundary of the 5th classification, the feedback information of easy aliasing, first sample feedback information is chosen.Specifically, Ke Yifen The class center distance d of each feedback information and the 5th nearest classification is not calculated_M51, each feedback information and the 5th time close classification Class center distance d_M52；In general, the corresponding range difference (d of feedback information_M52-d_M51) bigger, it is less susceptible to occur mixing between class It is folded, therefore can be according to (d_M52-d_M51) feedback information is ranked up, choose the feedback information conduct of a certain number of range difference minimums First sample feedback information.

To choosing, the mode of first sample feedback information, the quantity of first sample feedback information can not be done disclosure scheme has Body limits, and in order to ensure the data balancing of training sample, can make the first sample feedback information of different satisfaction grades Quantity it is suitable so that the obtained Satisfaction index class model of training has preferable distinction.

As an example, M₅A 5th classification can be clustered by automatic cluster, or the mode of the super ginseng of setting Processing, disclosure scheme can be not specifically limited this.

S503 utilizes the satisfaction grade of the first sample feedback information and the first sample feedback information, training Satisfaction index class model is obtained, the Satisfaction index class model includes the text representation layer for carrying out vectorization processing.

As an example, can utilize the trained Satisfaction index class model of phrase structure grammar tree shown in Fig. 7, model it is defeated Enter for first sample feedback information, the output of model is the satisfaction grade of first sample feedback information.First sample feedback letter After breath input, the structure grammar tree of text representation can be automatically generated through text expression layer, Fig. 7 show two points of phrase structure texts Method tree, therefore the input of each intermediate node is divided into left and right two.The specific implementation process of phrase structure grammar tree, reference can be made to shore State Chinese treebank (Chinese Treebank：https://catalog.ldc.upenn.edu/LDC2016T13) in correlation It explains, disclosure scheme can be not detailed this.In general, the intermediate node number in grammar tree is mainly by first sample feedback letter Cease the influence of the factors such as word quantity, the content of first sample feedback information split out.

In actual application, the intermediate node in Fig. 7 can be embodied as node diagram shown in Fig. 8, corresponding public affairs Formula transformation, which can be found in, to be introduced below, wherein capitalization representing matrix, lowercase letter vector：

The calculation for indicating input gate (i), for believing the input of present node Breath is converted, wherein

The calculation for indicating out gate (o), for controlling present node to father The information of node is transmitted；

For to input two child nodes and the information that currently inputs into Row converts and is merged into u_jIn；

f_k=tanh (W^fx_j+U^fh_k+b^f), indicate the calculation for forgeing door (f), for the filtering of sub- nodal information or Person converts；

c_j=i_j*u_j+∑_k∈C(j)f_k*c_k, the calculation of mnemon (c) is indicated, for by forgeing door, input gate Corresponding information input is controlled, suitable information is selected to retain and is passed in descendant node；

h_j=tanh (c_j)*o_j, indicate the calculation of present node hidden layer (h).

For example, the topological structure of text representation layer can be presented as Tree-LSTM, Bi-LSTM, CNN, Recursive Autoencoder etc., disclosure scheme can be not specifically limited this.It is to be appreciated that when satisfaction is classified The satisfaction grade of the first sample feedback information of model output, when being consistent with the satisfaction grade being artificially arranged, it is believed that Model training is completed.

S504 exports new vectorization through the text representation layer and indicates, as institute using the feedback information as input The vectorization for stating feedback information indicates.

Relative to general corpus, the Satisfaction index class model trained using the feedback information of personnel to be allocated, into The accuracy higher of row vectorization processing, therefore the new vectorization that can export text representation layer indicates, as feedback information Vectorization indicate.Accordingly, S402 can indicate to carry out clustering processing based on the new vectorization of feedback information, obtain M₄It is a 4th classification.

Referring to Fig. 9, show that the disclosure obtains the flow diagram of the vectorization expression embodiment 2 of feedback information.It can be with Include the following steps：

S601, the initial vectorization for obtaining every feedback information indicates, and is indicated based on the initial vectorization of feedback information Clustering processing is carried out, M is obtained₅A 5th classification, M₅>1。

S602, from the M₅First sample feedback information is chosen in a 5th classification, and marks each first sample feedback The satisfaction grade of information.

S603 utilizes the satisfaction grade of the first sample feedback information and the first sample feedback information, training Satisfaction index class model is obtained, the Satisfaction index class model includes the text representation layer for carrying out vectorization processing.

S604 exports new vectorization through the text representation layer and indicates using the feedback information as input.

The realization process of step S601~S604, can refer to and introduced at S501~S504 above, details are not described herein again.

S605, the new vectorization based on feedback information indicate progress clustering processing, obtain M₆A 6th classification, M₆>1。

S606, the poor (d of selected distance₄-d₃) minimum N₂A second sample back information, and it is anti-to mark each second sample The satisfaction grade of feedforward information, d₃It is the second sample back information at a distance from the class center of the 6th nearest classification, d₄It is second Sample back information is at a distance from the class center of secondary the 6th close classification.

In order to further increase the accuracy of feedback information vectorization expression, disclosure scheme also provides a kind of optimization satisfaction Spend the scheme of disaggregated model.Specifically, the sample data of optimization Satisfaction index class model may include：

(1) second sample back information.

Specifically, it first the new vectorization based on feedback information can indicate to carry out clustering processing, obtain M₆A 6th class Not；Then from the feedback information positioned at the boundary of the 6th classification, easy aliasing, the second sample back information is selected.

As an example, choosing the mode of the second sample back information can be presented as：Calculate separately each feedback information With the class center distance d of the 6th nearest classification₃, each feedback information and the 6th time close classification class center distance d₄；It is logical Often, the corresponding range difference (d of feedback information₄-d₃) bigger, it is less susceptible to occur aliasing between class, therefore can be according to (d₄-d₃) to feedback Information is ranked up, and chooses minimum preceding N₂A feedback information is as the second sample back information.In actual application, may be used Preferentially to choose the second sample back information from the feedback information in addition to first sample feedback information, disclosure scheme is to this It can be not specifically limited.

It should be noted that M₆A 6th classification can be clustered by automatic cluster, or the mode of the super ginseng of setting Processing, disclosure scheme can be not specifically limited this.

The satisfaction grade of (2) second sample back information.Specifically, satisfaction grade can be presented as it is satisfied, general, At least one of be unsatisfied with, be not satisfied at all, it can be taking human as the satisfaction grade of the second sample back information of setting.

As an example, the M clustered₆A 6th classification possibly can not cover all satisfaction grades, corresponding In this, disclosure scheme can also provide a kind of new clustering method, be desirably to obtain more classifications, cover all expire as possible Meaning degree grade.

Specifically, M can be first determined₇A available class center, M₇A available class center and M₆The class center of a 6th classification The distance between be not less than pre-determined distance；Then by M₇A available class center and M₆The class center of a 6th classification, is referred to as new Class center, based on new class center, the feedback information for treating distribution personnel again carries out clustering processing, determines every feedback letter New classification belonging to breath, each new classification correspond to a new class center.

As an example, M can be determined in the following way₇A available class center：

Mode one, directly selects M₇A feedback information conduct can use class center, it is ensured that selected feedback information and M₆A 6th class The distance between other class center is not less than pre-determined distance.

Mode two, first selects M₇A feedback information is as initially available class center, then at each initially available class center Nearby choose a certain number of available feedback information, calculate initial available class center, available feedback information mean value, obtain M₇It is a Class center can be used.Disclosure scheme can not limit the quantity of available feedback information, specific true in combination with practical application request It is fixed.

It is to be appreciated that pre-determined distance is bigger, the distance between class center and the class center of the 6th classification can be used remoter, hair The possibility of now new satisfaction grade is bigger, and disclosure scheme can not limit the specific value of pre-determined distance, preset away from With a distance from the typically larger than class center to boundary of the 6th classification.

As an example, M₆The class center of a 6th classification can be all feedback informations that the 6th classification includes, meter The class center of calculating；Alternatively, can also be the partial feedback information that the 6th classification includes, calculated class center, disclosure side Case can not limit this, specific to be determined in combination with practical application request.

To sum up, (M is obtained₆+M₇) behind a new class center, can a clustering processing be carried out to feedback information again.Specifically Ground can determine the new classification that feedback information is belonged to according to feedback information and new the distance between class center, in general, Feedback information belongs to the nearest corresponding new classification in new class center of its distance.

Corresponding to this, the second sample can be chosen from the feedback information positioned at the boundary of new classification, easy aliasing Feedback information.Specifically, feedback information and nearest new class center distance d can first be calculated₅, feedback information with it is time close The distance d at new class center₆, then according to range difference (d₆-d₅) select N₂A second sample back information.

S607, the satisfaction of satisfaction grade, the second sample back information based on the first sample feedback information Degree grade and remaining feedback information in addition to the first sample feedback information and the second sample back information are expired Meaning degree grade calculates update class center, and the satisfaction grade of the residue feedback information is belonging to the remaining feedback information Classification determines.

The feedback information of personnel to be allocated can be divided into three parts by disclosure scheme：First sample feedback information, Two sample back information, remaining feedback information.Wherein, the satisfaction grade of first sample feedback information, the second sample back letter The satisfaction grade of breath can be artificially arranged, and the satisfaction grade of remaining feedback information can be according to remaining feedback information What affiliated classification determined.

For example, if from the feedback information positioned at the boundary of the 6th classification, easy aliasing, the second sample is chosen Feedback information then can determine the satisfaction grade of residue feedback information according to the 6th classification belonging to remaining feedback information；If From the feedback information positioned at the boundary of new classification, easy aliasing, choose the second sample back information, then it can be according to residue New classification belonging to feedback information determines the satisfaction grade of residue feedback information.

As an example, update class center can be calculated according to following formula：

Wherein, μ_kClass center is updated for k-th；r_nk、w_nkFor 0/1 matrix of N × K dimensions, it is used to indicate n-th of feedback information Satisfaction grade, when n-th of feedback information is first sample feedback information or the second sample back information, if its be satisfied with Degree grade is noted as K classes, then w_nk=1, otherwise w_nk=0；When n-th of feedback information is remaining feedback information, if it is full Meaning degree grade is noted as K classes, then r_nk=1, otherwise r_nk=0；N is the total number of feedback information, and K is the sum for updating class Mesh；α is the weight of remaining feedback information, 0 < α < 1；f(S_n) it is S layers of outputs in network shown in Fig. 7.

S608 is based on the update class center, carries out clustering processing to the feedback information again, determines every feedback More new category belonging to information, each more new category correspond to a update class center.

After obtaining update class center, can a clustering processing be carried out to feedback information again.It specifically, can be according to anti- The distance between feedforward information and update class center, determine the more new category that feedback information is belonged to, in general, feedback information belongs to It is apart from the nearest corresponding more new category in update class center.It is to be appreciated that carrying out clustering processing, phase to feedback information again When in the satisfaction grade r for updating n-th of feedback information_nk、w_nk。

S609 updates the Satisfaction index using the more new category belonging to the feedback information and the feedback information Class model, until updated Satisfaction index class model meets predetermined condition, the updated Satisfaction index class model Including updated text representation layer.

Based on more new category, after obtaining the satisfaction grade of feedback information and feedback information, backpropagation can be used Text representation network in training update Fig. 7, i.e. f (s_n), n=1,2, L, N, until convergence.Wherein, loss function can embody For：

Loss function J_semiIncluding following three：First item indicates the cost of remaining feedback information；Section 2 indicates first The cost of sample back information and the second sample back information；The effect of Section 3 is when in feedback information and two update classes The heart it is closely located when, promote it close to affiliated update class center, far from other update class centers；α is remaining feedback information Weight, 0 < α < 1；L is the super parameter for indicating closely located degree.

S610 exports renewal vector table using the feedback information as input through the updated text representation layer Show, the vectorization as the feedback information indicates.

Using the updated Satisfaction index class model of the second sample back Advance data quality, vectorization table can be further increased The accuracy shown, therefore the renewal vectorization that can export updated text representation layer indicates, the vector as feedback information Change and indicates.Accordingly, S402 can indicate to carry out clustering processing based on the renewal vectorization of feedback information, obtain M₄A 4th class Not.

Referring to Figure 10, the composition schematic diagram of disclosure personnel's apparatus for grouping is shown.Described device may include：

It pays close attention to information and splits module 701, the concern information is split as by the concern information for obtaining personnel to be allocated At least one phrase；

Phrase vectorization indicates acquisition module 702, and the vectorization for obtaining each phrase indicates；

Phrase clustering processing module 703 indicates to carry out clustering processing for phrase-based vectorization, obtains M₁A first Classification, each first category indicate a kind of dimensional information, M₁≥1；

Group result obtains module 704, for the corresponding weight of each dimensional information to be arranged, using the dimensional information with And the corresponding weight of the dimensional information, obtain the group result of the personnel to be allocated.

Optionally, described device further includes：

About the device in above-described embodiment, wherein modules execute the concrete mode of operation in related this method Embodiment in be described in detail, explanation will be not set forth in detail herein.

Referring to Figure 11, structural schematic diagram of the disclosure for the electronic equipment 800 of personnel's grouping is shown.Electronic equipment 800 at least may include processor 801 and storage medium 802, and as an example, processor 801 and storage medium 802 can be with It is connected by bus or other means, shown in Figure 11 for being connected by bus.The quantity of processor 801 can be one or Person is multiple, shown in Figure 11 by taking a processor as an example.Storage device resource representated by storage medium 802, can be by for storing The instruction that processor 801 executes, such as application program.In addition, processor 801 can be configured as the finger in load store medium It enables, to execute above-noted persons' group technology.

The preferred embodiment of the disclosure is described in detail above in association with attached drawing, still, the disclosure is not limited to above-mentioned reality The detail in mode is applied, in the range of the technology design of the disclosure, a variety of letters can be carried out to the technical solution of the disclosure Monotropic type, these simple variants belong to the protection domain of the disclosure.

It is further to note that specific technical features described in the above specific embodiments, in not lance In the case of shield, can be combined by any suitable means, in order to avoid unnecessary repetition, the disclosure to it is various can The combination of energy no longer separately illustrates.

In addition, arbitrary combination can also be carried out between a variety of different embodiments of the disclosure, as long as it is without prejudice to originally Disclosed thought equally should be considered as disclosure disclosure of that.

Claims

1. a kind of personnel's group technology, which is characterized in that the method includes：

The concern information is split as at least one phrase, obtains each phrase by the concern information for obtaining personnel to be allocated Vectorization indicates；

Phrase-based vectorization indicates progress clustering processing, obtains M₁A first category, each first category indicate a kind of dimension Information, M₁≥1；

The corresponding weight of each dimensional information is set, using the dimensional information and the corresponding weight of the dimensional information, is obtained The group result of the personnel to be allocated.

2. according to the method described in claim 1, it is characterized in that, the vectorization for obtaining each phrase indicates, including：

The initial vectorization for obtaining each phrase indicates, and phrase-based initial vectorization indicates to carry out clustering processing, obtains M₂ A second category, M₂>1；

It will be from the M₂The first sample phrase selected in a second category, combination of two is first sample phrase pair, and is obtained The markup information of each first sample phrase pair, the markup information are similar or dissimilar；

Using the first sample phrase pair and the markup information of the first sample phrase pair, training obtains phrase classification mould Type, the phrase classification model include the phrase expression layer for carrying out vectorization processing；

The phrase that the concern information is split out exports new vectorization as input, through the phrase expression layer and indicates, makees It is indicated for the vectorization of the phrase.

3. according to the method described in claim 2, it is characterized in that, after the new vectorization for obtaining the phrase indicates, institute State obtain each phrase vectorization expression further include：

Poor (the d of selected distance₂-d₁) minimum N₁A second sample phrase, d₁For the second sample phrase and nearest third class The distance at other class center, d₂It is the second sample phrase at a distance from the class center of time close third classification；

By the class center of the second sample phrase and nearest third classification and/or the second sample phrase with it is time close The class center of third classification, combination of two is the second sample phrase pair, and obtains the markup information of each second sample phrase pair；

Using the second sample phrase pair and the markup information of the second sample phrase pair, the phrase classification mould is updated Type, until updated phrase classification model meets preset condition, the updated phrase classification model includes update Phrase expression layer afterwards；

The phrase that the concern information is split out exports renewal vector as input, through the updated phrase expression layer It indicates, the vectorization as the phrase indicates.

4. method according to any one of claims 1 to 3, which is characterized in that the method further includes：

The personnel to be allocated are obtained for the feedback information of the group result and the vectorization table of every feedback information Show；

Vectorization based on feedback information indicates progress clustering processing, obtains M₄A 4th classification, each 4th classification correspond to a kind of Satisfaction grade, M₄≥1；

Judge the M₄Whether include classification to be adjusted in a 4th classification, the corresponding satisfaction grade of the classification to be adjusted is used The personnel to be adjusted for belonging to the classification to be adjusted in expression are unsatisfied with the group result, and the personnel to be adjusted belong to described Personnel to be allocated；

If the M₄A 4th classification includes the classification to be adjusted, then obtains the grouping update letter of the personnel to be adjusted Breath；

5. according to the method described in claim 4, it is characterized in that, the vectorization expression for obtaining every feedback information, packet It includes：

The initial vectorization for obtaining every feedback information indicates, and is carried out at cluster based on the expression of the initial vectorization of feedback information Reason, obtains M₅A 5th classification, M₅>1；

From the M₅First sample feedback information is chosen in a 5th classification, and marks the satisfaction of each first sample feedback information Spend grade；

Using the satisfaction grade of the first sample feedback information and the first sample feedback information, training obtains satisfaction Disaggregated model, the Satisfaction index class model include the text representation layer for carrying out vectorization processing；

Using the feedback information as input, exports new vectorization through the text representation layer and indicate, as the feedback letter The vectorization of breath indicates.

6. according to the method described in claim 5, it is characterized in that, being indicated in the new vectorization for obtaining the feedback information Afterwards, the vectorization expression for obtaining every feedback information further includes：

Poor (the d of selected distance₄-d₃) minimum N₂A second sample back information, and mark expiring for each second sample back information Meaning degree grade, d₃It is the second sample back information at a distance from the class center of the 6th nearest classification, d₄Believe for the second sample back It ceases at a distance from the class center of secondary the 6th close classification；

The satisfaction grade of satisfaction grade, the second sample back information based on the first sample feedback information, with And the satisfaction grade of the remaining feedback information in addition to the first sample feedback information and the second sample back information, Update class center is calculated, the satisfaction grade of the residue feedback information is determined by the classification belonging to the remaining feedback information；

Based on the update class center, clustering processing is carried out to the feedback information again, is determined belonging to every feedback information More new category, each more new category correspond to a update class center；

Using the more new category belonging to the feedback information and the feedback information, the Satisfaction index class model is updated, directly Until updated Satisfaction index class model meets predetermined condition, the updated Satisfaction index class model includes after updating Text representation layer；

Using the feedback information as input, indicated through the updated text representation layer output renewal vectorization, as institute The vectorization for stating feedback information indicates.

7. according to the method described in claim 6, it is characterized in that, described obtain M₆It is described to obtain every instead after a 6th classification The vectorization of feedforward information indicates：

Determine M₇A available class center, the M₇A available class center and the M₆Between the class center of a 6th classification away from From not less than pre-determined distance；

Based on new class center, clustering processing is carried out to the feedback information again, is determined new belonging to every feedback information Classification, each new classification corresponds to a new class center, and the new class center includes the M₇A available class center and The M₆The class center of a 6th classification；

Then, the second sample back information is chosen in the following way：Poor (the d of selected distance₆-d₅) minimum N₂A second sample This feedback information, d₅It is the second sample back information at a distance from nearest new class center, d₆For the second sample back information with The distance at secondary close new class center.

8. a kind of personnel's apparatus for grouping, which is characterized in that described device includes：

It pays close attention to information and splits module, the concern information is split as at least one by the concern information for obtaining personnel to be allocated A phrase；

Phrase clustering processing module indicates to carry out clustering processing for phrase-based vectorization, obtains M₁A first category, often A first category indicates a kind of dimensional information, M₁≥1；

Group result obtains module, for the corresponding weight of each dimensional information to be arranged, utilizes the dimensional information and the dimension The corresponding weight of information is spent, the group result of the personnel to be allocated is obtained.

9. device according to claim 8, which is characterized in that the phrase vectorization indicates that acquisition module includes：

First clustering processing module indicates to carry out clustering processing for phrase-based initial vectorization, obtains M₂A second class Not, M₂>1；

First sample phrase chooses module, and being used for will be from the M₂The first sample phrase selected in a second category, two-by-two group Be combined into first sample phrase pair, and obtain the markup information of each first sample phrase pair, the markup information be it is similar or It is dissimilar；

Phrase classification model training module, for the mark using the first sample phrase pair and the first sample phrase pair Information, training obtain phrase classification model, and the phrase classification model includes the phrase expression layer for carrying out vectorization processing；

Secondary vectorization indicates output module, and the phrase for splitting out the concern information is as input, through the phrase Expression layer exports new vectorization and indicates, the vectorization as the phrase indicates.

10. device according to claim 9, which is characterized in that the phrase vectorization indicates that acquisition module further includes：

Second clustering processing module, after being indicated in the new vectorization for obtaining the phrase, phrase-based new vector Change and indicate to carry out clustering processing, obtains M₃A third classification, M₃>1；

Second sample phrase chooses module, is used for the poor (d of selected distance₂-d₁) minimum N₁A second sample phrase, d₁It is described Two sample phrases are at a distance from the class center of nearest third classification, d₂For the second sample phrase and secondary close third classification Class center distance；The class center and/or second sample of the second sample phrase and nearest third classification is short The class center of language and secondary close third classification, combination of two is the second sample phrase pair, and obtains each second sample phrase pair Markup information；

Phrase classification model modification module, for the mark using the second sample phrase pair and the second sample phrase pair Information updates the phrase classification model, until updated phrase classification model meets preset condition, after the update Phrase classification model include updated phrase expression layer；

Third vectorization indicates output module, and the phrase for splitting out the concern information is as input, through the update Phrase expression layer output renewal vectorization afterwards indicates that the vectorization as the phrase indicates.

11. according to claim 8 to 10 any one of them device, which is characterized in that described device further includes：

Feedback information clustering processing module indicates to carry out clustering processing for the vectorization based on feedback information, obtains M₄A 4th Classification, each 4th classification correspond to a kind of satisfaction grade, M₄≥1；

Classification judgment module to be adjusted, for judging the M₄Whether include classification to be adjusted in a 4th classification, it is described to be adjusted The corresponding satisfaction grade of classification for indicating that the personnel to be adjusted for belonging to the classification to be adjusted are unsatisfied with the group result, The personnel to be adjusted belong to the personnel to be allocated；

It is grouped fresh information acquisition module, in the M₄When a 4th classification includes the classification to be adjusted, described in acquisition The grouping fresh information of personnel to be adjusted；

Group result adjusts module, for utilizing the grouping fresh information, adjusts the group result of the personnel to be adjusted.

12. according to the devices described in claim 11, which is characterized in that the feedback information vectorization indicates acquisition module packet It includes：

Third clustering processing module indicates to carry out clustering processing for the initial vectorization based on feedback information, obtains M₅A 5th Classification, M₅>1；

First sample feedback information chooses module, is used for from the M₅First sample feedback information is chosen in a 5th classification, and is marked Note the satisfaction grade of each first sample feedback information；

Satisfaction index class model training module, for utilizing the first sample feedback information and the first sample feedback information Satisfaction grade, training obtain Satisfaction index class model, the Satisfaction index class model includes for carrying out vectorization processing Text representation layer；

Five-way quantization means output module, for using the feedback information as input, being exported through the text representation layer new Vectorization indicate, as the feedback information vectorization indicate.

13. device according to claim 12, which is characterized in that the feedback information vectorization indicates that acquisition module also wraps It includes：

4th clustering processing module, after being indicated in the new vectorization for obtaining the feedback information, based on feedback information New vectorization indicates progress clustering processing, obtains M₆A 6th classification, M₆>1；

Second sample back information chooses module, is used for the poor (d of selected distance₄-d₃) minimum N₂A second sample back information, And mark the satisfaction grade of each second sample back information, d₃For the second sample back information and the 6th nearest classification The distance at class center, d₄It is the second sample back information at a distance from the class center of the 6th time close classification；

Class center calculation module is updated, for the satisfaction grade based on the first sample feedback information, second sample The satisfaction grade of feedback information and remaining in addition to the first sample feedback information and the second sample back information The satisfaction grade of remaining feedback information calculates update class center, and the satisfaction grade of the residue feedback information is by the residue Classification belonging to feedback information determines；

5th clustering processing module carries out clustering processing, really to the feedback information again for being based on the update class center The more new category belonging to every feedback information is made, each more new category corresponds to a update class center；

Satisfaction index class model update module, for utilizing the update class belonging to the feedback information and the feedback information Not, the Satisfaction index class model is updated, until updated Satisfaction index class model meets predetermined condition, the update Satisfaction index class model afterwards includes updated text representation layer；

Six-way quantization means output module is used for using the feedback information as input, through the updated text representation Layer output renewal vectorization indicates that the vectorization as the feedback information indicates.

14. device according to claim 13, which is characterized in that the feedback information vectorization indicates that acquisition module also wraps It includes：

Class center determining module can be used, for obtaining M described₆After a 6th classification, M is determined₇A available class center, it is described M₇A available class center and the M₆The distance between class center of a 6th classification is not less than pre-determined distance；

6th clustering processing module, for based on new class center, carrying out clustering processing to the feedback information again, determining New classification belonging to every feedback information, each new classification correspond to a new class center, and the new class center includes The M₇A available class center and the M₆The class center of a 6th classification；

Then, the second sample back information chooses module, is used for the poor (d of selected distance₆-d₅) minimum N₂A second sample is anti- Feedforward information, d₅It is the second sample back information at a distance from nearest new class center, d₆For the second sample back information with it is time close New class center distance.

15. a kind of storage medium, wherein being stored with a plurality of instruction, which is characterized in that described instruction is loaded by processor, right of execution Profit requires the step of any one of 1 to 7 the method.

16. a kind of electronic equipment, which is characterized in that the electronic equipment includes；

Storage medium described in claim 15；And

Processor, for executing the instruction in the storage medium.