CN115330142B

CN115330142B - Training method of joint capacity model, capacity demand matching method and device

Info

Publication number: CN115330142B
Application number: CN202210878392.6A
Authority: CN
Inventors: 李满伟; 秦川; 马海平; 申大忠; 祝恒书; 张敬帅; 姚开春
Original assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Current assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Priority date: 2022-07-25
Filing date: 2022-07-25
Publication date: 2023-06-02
Anticipated expiration: 2042-07-25
Also published as: CN115330142A

Abstract

The disclosure provides a training method of a joint capacity model, a capacity demand matching method and a device, relates to the technical field of artificial intelligence, and particularly relates to the field of deep learning. The specific implementation scheme is as follows: processing the sample data by adopting a plurality of capability prediction models to obtain a plurality of capability information; processing the plurality of capability information and the requirement information by adopting a requirement matching model to obtain a requirement matching result, wherein the requirement information comprises capability information which needs to be met by a target scene; the demand matching model and the plurality of capacity prediction models are trained based on the demand matching results to update parameters of the demand matching model and the plurality of capacity prediction models. The method and the device have the advantages that the requirement matching model and the multiple capability prediction model are trained by utilizing multiple sample data, and the combined capability model obtained after training can meet comprehensive and multi-dimensional capability prediction and requirement matching, so that the method and the device are suitable for richer application scenes.

Description

Training method of joint capacity model, capacity demand matching method and device

Technical Field

The present disclosure relates to the field of artificial intelligence, and more particularly to the field of deep learning.

Background

Ability assessment can select the most appropriate candidate by assessing the consistency of candidate skills with job requirements, a critical task in talent recruitment. Traditional capability assessment involves multiple processes, employing different forms of assessment methods, often resulting in assessment conclusions that are decentralized, noisy, and unreliable. The decision can be made by expert decision making or natural language processing technology.

Disclosure of Invention

The disclosure provides a training method of a joint capacity model, a capacity demand matching method and a device.

According to an aspect of the present disclosure, there is provided a training method of a joint capacity model including a plurality of capacity prediction models and a demand matching model, the method including:

processing the sample data by adopting the multiple-capability prediction model to obtain multiple-capability information;

processing the plurality of capability information and the requirement information by adopting the requirement matching model to obtain a requirement matching result, wherein the requirement information comprises capability information which needs to be met by a target scene;

training the demand matching model and the plurality of capacity prediction models based on the demand matching result to update parameters of the demand matching model and the plurality of capacity prediction models.

According to another aspect of the present disclosure, there is provided a capability requirement matching method, including:

processing various data to be processed of the candidate by adopting various capability prediction models to obtain various capability information of the candidate;

and processing the plurality of capability information and the requirement information by adopting a requirement matching model to obtain a requirement matching result of the candidate, wherein the requirement information comprises capability information which needs to be met by a target scene.

According to another aspect of the present disclosure, there is provided a training apparatus of a joint capacity model including a plurality of capacity prediction models and a demand matching model, the apparatus including:

the capacity prediction module is used for processing the sample data by adopting the multiple capacity prediction models to obtain multiple capacity information;

the demand matching module is used for processing the plurality of capability information and the demand information by adopting the demand matching model to obtain a demand matching result, wherein the demand information comprises capability information which needs to be met by a target scene;

and the training module is used for training the demand matching model and the multiple capacity prediction models based on the demand matching result so as to update parameters of the demand matching model and the multiple capacity prediction models.

According to another aspect of the present disclosure, there is provided a capability requirement matching apparatus, comprising:

the capacity prediction module is used for processing various data to be processed of the candidate by adopting various capacity prediction models to obtain various capacity information of the candidate;

the demand matching module is used for processing the plurality of capability information and the demand information by adopting a demand matching model to obtain a demand matching result of the candidate, wherein the demand information comprises capability information which needs to be met by a target scene.

According to another aspect of the present disclosure, there is provided an electronic device including:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein, the liquid crystal display device comprises a liquid crystal display device,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of the embodiments of the present disclosure.

According to another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform a method according to any one of the embodiments of the present disclosure.

According to another aspect of the present disclosure, there is provided a computer program product comprising a computer program which, when executed by a processor, implements a method according to any of the embodiments of the present disclosure.

According to the embodiment of the disclosure, the requirement matching model and the multiple capability prediction models are trained by utilizing multiple sample data, and the combined capability model obtained after training can meet more comprehensive and multidimensional capability prediction and requirement matching, so that the method is suitable for richer application scenes.

It should be understood that the description in this section is not intended to identify key or critical features of the embodiments of the disclosure, nor is it intended to be used to limit the scope of the disclosure. Other features of the present disclosure will become apparent from the following specification.

Drawings

The drawings are for a better understanding of the present solution and are not to be construed as limiting the present disclosure. Wherein:

FIG. 1 is a flow diagram of a method of training a joint capacity model according to an embodiment of the present disclosure;

FIG. 2 is a flow diagram of a method of training a joint capacity model according to another embodiment of the present disclosure;

FIG. 3 is a flow diagram of a method of training a joint capacity model according to another embodiment of the present disclosure;

FIG. 4 is a flow diagram of a method of training a joint capacity model according to another embodiment of the present disclosure;

FIG. 5 is a flow diagram of a method of training a joint capacity model according to another embodiment of the present disclosure;

FIG. 6 is a flow diagram of a method of training a joint capacity model according to another embodiment of the present disclosure;

FIG. 7 is a flow diagram of a method of training a joint capacity model according to another embodiment of the present disclosure;

FIG. 8 is a flow diagram of a method of training a joint capacity model according to another embodiment of the present disclosure;

FIG. 9 is a flow diagram of a capability requirement matching method according to an embodiment of the present disclosure;

FIG. 10 is a flow diagram of a capability requirement matching method according to an embodiment of the present disclosure;

FIG. 11 is a flow diagram of a capability requirement matching method according to an embodiment of the present disclosure;

FIG. 12 is a flow diagram of a capability requirement matching method according to an embodiment of the present disclosure;

FIG. 13 is a flow diagram of a capability requirement matching method according to an embodiment of the present disclosure;

FIG. 14 is a flow diagram of a capability requirement matching method according to an embodiment of the present disclosure;

FIG. 15 is a schematic structural view of a training device of the joint capacity model according to an embodiment of the present disclosure;

FIG. 16 is a schematic structural view of a training device of a joint capacity model according to another embodiment of the present disclosure;

FIG. 17 is a schematic diagram of a capability requirement matching device according to an embodiment of the present disclosure;

FIG. 18 is a schematic structural view of a capability requirement matching device according to another embodiment of the present disclosure;

FIG. 19 is a schematic diagram of a joint capacity diagnostic method framework in talent recruitment of the present disclosure;

FIG. 20 is a frame diagram of a try-on modeling of the present disclosure;

FIG. 21 is a framework diagram of resume modeling of the present disclosure;

fig. 22 is a block diagram of an electronic device used to implement the methods of embodiments of the present disclosure.

Detailed Description

Exemplary embodiments of the present disclosure are described below in conjunction with the accompanying drawings, which include various details of the embodiments of the present disclosure to facilitate understanding, and should be considered as merely exemplary. Accordingly, one of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present disclosure. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

FIG. 1 is a flow diagram of a method of training a joint capacity model according to an embodiment of the present disclosure. The joint capacity model may include a plurality of capacity prediction models and a demand matching model, and the training method may include:

s101, processing sample data by adopting a plurality of capability prediction models to obtain a plurality of capability information;

S102, processing the plurality of capability information and the requirement information by adopting a requirement matching model to obtain a requirement matching result, wherein the requirement information comprises capability information which needs to be met by a target scene;

and S103, training the demand matching model and the multiple capacity prediction models based on the demand matching result so as to update parameters of the demand matching model and the multiple capacity prediction models.

In the disclosed embodiments, the federated capability model may include a variety of capability prediction models and demand matching models. According to different application scenes, different capacity prediction models can be constructed. For example, in job hunting scenarios, the multiple capability prediction models may include a interview capability prediction model, a resume capability prediction model, an interview capability prediction model, and the like. For another example, in the context of an ascent, consultation, etc., the multiple capability prediction models may include a written trial capability prediction model, an interview capability prediction model, and the like.

In the embodiment of the disclosure, the requirement matching model can match various capability information with specific requirements to obtain a requirement matching result. For example, the demand matching results may include job entry probabilities, entrance probabilities, admission probabilities, and the like. For another example, the requirement matching result may also include a result of whether the requirement matching result matches a specific post or school.

In the embodiment of the disclosure, in the training process, multiple kinds of capability prediction models can be adopted to respectively and correspondingly process multiple kinds of sample data, so as to obtain multiple kinds of capability information. The plurality of sample data may include a lot of the plurality of sample data, and in particular may be classified according to the type of the capacity prediction model, for example, a lot of 3A sample data includes a sample pen test data, a sample resume data, and a sample interview data. A is a positive integer. Of course, the number of sample pen test data, sample resume data, and sample interview data may also be different.

In the embodiment of the disclosure, the requirement information includes capability information that needs to be satisfied by the target scene. The demand information for different target scenarios may be different. For example, the demand information for the application scenario includes a post-related skill specifically required for a post, which may also be referred to as skill demand information for the post, such as: familiar computer languages, professionals, and the like. For another example, the requirement information of the entrance scene includes learning-related skills specifically required by a certain school, and may also be referred to as learning-related skill requirement information, such as english level, discipline score, competition rewards, and the like. And then, matching the capability information predicted by each capability prediction model with specific requirement information such as skill requirement information of posts by adopting a requirement matching model to obtain a requirement matching result. And then training the demand matching model and each capacity prediction model based on the demand matching result so as to update the parameters of the demand matching model and each capacity prediction model until the integral loss function converges, and stopping training. And carrying out demand matching prediction on the data related to the capacities of the candidates of the multiple data sources by utilizing multiple capacity prediction models and demand matching models in the combined capacity model obtained through training.

In the embodiment of the disclosure, the requirement matching model and the multiple capability prediction models are trained by utilizing multiple sample data, and the combined capability model obtained after training can meet more comprehensive and multidimensional capability prediction and requirement matching, so that the method is applicable to richer application scenes. Further, the processing efficiency of some application scenes can be improved, for example, recruitment efficiency in job hunting scenes is improved. Furthermore, subjective factors of the prediction result can be reduced, and the method has strong expandability.

FIG. 2 is a flow diagram of a method of training a joint capacity model according to another embodiment of the present disclosure. The method of this embodiment includes one or more features of the training method embodiments described above. In one possible implementation, the sample data is processed by using a plurality of capability prediction models to obtain a plurality of capability information, including at least two steps as follows:

s201, processing sample trial data by adopting a trial capacity prediction model to obtain trial capacity information;

s202, processing sample interview data by adopting an interview capability prediction model to obtain interview capability information;

s203, processing the sample resume data by adopting a resume capability prediction model to obtain resume capability information.

In embodiments of the present disclosure, sample try-on data may be selected and generated from the try-on data for several candidates. For example, the candidate's pen test data may include text generated from the candidate's job entry test, entrance test, etc. papers. The pen test data may include pen test questions and answer results of candidates for each question. Each of the pen questions may have a corresponding number, and the answer result of the candidate for each question may be represented by a specific numerical value. For example 0 indicates correct and 1 indicates incorrect. For another example, 10 represents 100 minutes, 9 represents 90 minutes, 8 represents 80 minutes, and so on. In the embodiments of the present disclosure, the questions may also be referred to as problems, exercises, questions, or the like.

In embodiments of the present disclosure, sample interview data may be selected and generated from interview data for a number of candidates. The interview data of the candidate may include the interview's text of the candidate's assessment of various skills in the interview process, may also include text related to the candidate's skills extracted from the interview audio-visual data, and so on. For example, the skills of the candidate may include, but are not limited to, english level, computer software usage skills, graduation institution professional-related skills, and the like.

In embodiments of the present disclosure, sample resume data may be selected and generated from resume data for a number of candidates. The resume data of the candidate may include information on whether the candidate has a certain skill or not.

At least two kinds of capability information of the candidate can be predicted and obtained through at least two kinds of the trial capability prediction model, the interview capability prediction model and the resume capability prediction model, and further the method of adopting the combined training based on the comprehensive capability of the candidate is facilitated to update the combined capability model, so that more comprehensive capability prediction and demand matching are realized.

In one possible implementation, as shown in fig. 3, the processing of the sample test data by using the test capability prediction model to obtain test capability information includes:

s301, extracting candidate features from the sample pen test data;

and S302, obtaining the trial capability information of the candidate based on the candidate characteristics.

In the embodiment of the disclosure, the candidate characteristics in the try data can comprise macro characterization of the candidate about the try, the candidate characteristics can be expressed in the form of vectors and the like, and the specific capability of the candidate in the try can be accurately obtained through the candidate characteristics in the try data. For example, the candidate feature extracted from the sample pen test data may be candidate u _i Is a one-hot vector

The candidate's test ability information may be the macroscopic characterization of the candidate with respect to the test +.>

May be a d-dimensional vector. />

Wherein W is ^U May be an embedding matrix, W ^U May have an initial value. W can be aligned in the subsequent training process ^U Learning and updating are performed, and further, the written test capability information is updated. After training, W is used for predicting the trained W in the model by using the pen test capability ^U Candidate characteristics in the real test pencil data of a certain candidate can be accurately predicted, and test pencil capacity information of the candidate can be accurately predicted.

In one possible implementation, the loss function of the test capability prediction model is constructed based on the probability of correct answer predicted by the test capability prediction model and the actual answer result; the probability of correct answer is predicted by the test capacity prediction model based on the interaction information of the candidate obtained by the sample test data and the questions. Parameters of the pilot capacity prediction model can be updated by using the loss function of the pilot capacity prediction model, so that a more accurate pilot capacity prediction model is obtained.

For example, the loss function of the trial capacity prediction model is loss ^E ：

Wherein, the liquid crystal display device comprises a liquid crystal display device,

as candidate u _i Questions e _j Is (are) true>

Predicting candidate u for model _i Questions e _j Is the case in (a). Minimizing the loss function based on the try-on data allows learning the ability of each candidate +.>

U represents the candidate set. />

Representing a set of topics.

In one possible implementation, as shown in fig. 4, the step of obtaining the interaction information of the candidate and the topic includes:

s401, extracting candidate features and topic features from the sample pen test data;

s402, acquiring the pen test capability information based on the candidate characteristics;

s403, acquiring the pilot skill characterization information based on the pilot capability information and the overall skill characterization information;

s404, obtaining topic difficulty information and topic distinguishing degree information based on the topic features;

s405, based on the pen test skill characterization information, the topic difficulty information, the topic distinguishing degree information and the topic association skill information, the interaction information of the candidate and the topic is obtained.

In one example, assume that N candidates, M lanes, exist for the systemProblem, K skills, candidate e.g. number u _i ∈R ¹ ^×N Problem number e _j ∈R ^1×M . Wherein R is ^1×N Representing a 1 xn vector. R is R ^1×M Representing a 1×m vector, other similar expressions have similar meanings.

The candidate features extracted from the sample trial data may include candidate u _i Is a one-hot vector

The title features may include title e _j Is>

The obtained test ability information based on the candidate features may be a macroscopic characterization of the candidate with respect to the test

The overall skill characterization information may be a skill characterization matrix h ^S ∈R ^K×d . The test skill characterization information derived based on the test capability information and the overall skill characterization information may be a characterization vector of the candidate with respect to the test on a particular skill

sigmoid represents an activation function. The superscript T denotes a transpose.

The question difficulty information obtained based on the question characteristics can be a question difficulty parameter

The topic distinction degree information obtained based on topic features can be topic distinction degree parameters

For example, questionsThe difficulty parameter of the title may be a vector and the distinguishing parameter of the title may be a scalar. W (W) ^diff 、W ^disc May be a learnable embedding matrix.

The topic associated skill information may be a topic associated skill vector

Representing the skill of the subject study. For example, if subject e _j Let s look at skill s, then the s-th position is 1, otherwise 0. Based on the pen test skill characterization information, the question difficulty information, the question distinguishing degree information and the question association skill information, the obtained interaction information of the candidate and the question can be referred to as the following formula:

Wherein x represents interaction information;

representing the candidate as representing vector on specific skill, namely representing information of the skill of the pen test; />

The question difficulty parameter is question difficulty information; />

The method is characterized in that topic distinguishing parameters, namely topic distinguishing degree information; />

The topic associated skill vector, i.e., topic associated skill information, is represented. Sign->

Representing para-multiplication, i.e.

And->

And multiplying the positions.

In an embodiment of the present disclosure, the trial capacity prediction model may include a neural network full connectivity layer (Fully Connected Layers, FCL) where candidates interact with topics. The input information of the neural network FCL may include interaction information x of the candidate with the topic. The output information of the neural network FCL may include the probability of correct predictive answer

The candidate's trial ability information may also be derived in the neural network based on the candidate features and a learnable embedding matrix. The test capability information may be a test capability vector.

The interaction information of the candidate and the questions can be obtained based on the project reaction theoretical formula of educational psychology, so that the interaction condition of the candidate and each question in the written test data can be more scientifically and reasonably determined, and further a more suitable loss function is obtained.

In one possible implementation, as shown in fig. 5, the interview capability prediction model is used to process the sample interview data to obtain interview capability information, including:

s501, extracting initial interview characterization information from the sample interview data;

s502, obtaining a first mean value and a first variance based on the interview characterization information;

s503, obtaining first characterization information conforming to Gaussian distribution based on the first mean value and the first variance, and sampling from the first characterization information to obtain first distribution potential variables;

s504, obtaining first theme distribution information based on the first distribution potential variable;

s505, obtaining a first topic collection based on the first topic representation information and the overall skill representation information;

s506, obtaining predicted interview characterization information based on the first topic distribution information and the first topic set;

s507, obtaining interview capability information based on the first theme distribution information and the first theme characterization information.

In one example, the initial interview characterization information may be an input vector of an interview capability prediction model

The input vector includes skills extracted from the interview text. For example, if there is a skill s in the interview text, s-th (s-th position) is 1, otherwise is 0.

Vector

Respectively passing through two fully connected neural networks (or called fully connected neural networks) to obtain a first average value mu ^A And a first difference sigma ^A . Next, the mixture is subjected to a Gaussian distribution N (mu) ^A ,(σ ^A ) ² ) Is sampled from the features of (a) to obtain a first distributed latent variable such as Gaussian random vector +.>

Then the first topic distribution information, such as topic distribution vector +.>

Characterizing a first topic, such as topic characterization matrix t ^A And overall skill characterization information such as skill characterization matrix h ^S Cross-multiplying and passing the softmax layer to obtain a first topic set beta ^A (the set may be a matrix). Wherein beta is ^A May include the distribution of topics within skill words ++>

Representing the proficiency of each topic k in the corresponding skill,/->

Finally, the first topic distribution information is->

With the first topic set beta ^A Multiplication results in a predicted value, i.e. predictive interview characterization information, e.g

The specific values in the predicted interview characterization information and the initial interview characterization information are typically different. In addition, if training needs to be continued, the interview characterization information predicted at this time can be used as the initial interview characterization information of the next time. Interview ability information->

May be a vector. In addition, t ^A May include->

A specific vector representing the topic k.

The first topic distribution information described above may correspond to a probability of each topic, indicating a grasp under each topic (skill set).

The first topic representation information described above may be a matrix that indicates that there are multiple topics in the system with respect to interviewed skills, where each topic represents a set of closer skill words. Each topic may have a token vector, i.e., a topic specific vector.

Initial interview characterization information, predicted interview characterization information, topic distribution, topic collection and the like related to interviews are obtained through sample interview data, subsequent training is facilitated, and more accurate interview capability information can be obtained.

In one possible implementation, the loss function of the interview capability prediction model is constructed based on the first topic distribution information, the first distribution latent variable, the predicted interview characterization information, and the first topic set. Parameters of the interview capacity prediction model can be updated by using the loss function of the interview capacity prediction model, so that a more accurate interview capacity prediction model is obtained.

For example, an example of a formula for a loss function of the interview capability prediction model is as follows:

is interview a _i Subject distribution of->

Is a distributed latent variable,/->

Is interview a _i Characterization of beta ^A Is the subject set of interviews. Wherein "|" represents the prior probability problem; "||" indicates juxtaposition. />

Representing a priori distribution +.>

And posterior distribution->

Is a distance of (3).

In one possible implementation, as shown in fig. 6, the processing the sample resume data using the resume capability prediction model to obtain resume capability information includes:

s601, extracting initial resume characterization information from the sample resume data;

s602, obtaining a second mean value and a second variance based on the resume characterization information;

s603, obtaining second characterization information conforming to Gaussian distribution based on the second mean and the second variance, and sampling from the second characterization information to obtain second distribution potential variables;

s604, obtaining second topic distribution information based on the second distribution potential variable;

s605, obtaining a second topic collection based on the second topic representation information and the overall skill representation information;

s606, obtaining predicted resume characterization information based on the second theme distribution information and the second theme set;

s607, obtaining resume capability information based on the second theme distribution information and the second theme characterization information.

In one example, the resume characterization information may be an input vector of a resume capability prediction model

The input vector includes skills extracted from the profile text. For example, if skills s exist in the resume text, the s-th position is 1, otherwise it is 0.

Vector

Respectively passing through two fully connected neural networks (or called fully connected neural networks) to obtain a second average value mu ^R And a second variance sigma ^R . Next, the mixture is subjected to a Gaussian distribution N (mu) ^R ,(σ ^R ) ² ) Is sampled in the features of (a) to obtain a second distributed latent variable such as Gaussian random vector +.>

Then a second topic distribution information, such as topic distribution vector +.>

Second topic representation information such as topic representation matrix t ^R And overall skill characterization information such as skill characterization matrix h ^S Cross-multiplying and passing the softmax layer to obtain a second topic set beta ^R (the set may be a matrix). Wherein beta is ^R May include the distribution (vector) of topics among skill words>

Representing the proficiency of each topic k in the corresponding skill,/->

Finally, the second topic distribution information->

And a second theme set beta ^R Multiplication to obtain a predicted value, i.e. predicted resume characterization information +.>

Furthermore, resume ability information +.>

May be a vector. In addition, t ^R May include- >

A specific vector representing the topic k.

The second topic distribution information described above may correspond to a probability of each topic, indicating a grasp under each topic (skill set).

The second topic representation information described above may be a matrix that indicates that there are multiple topics in the system for the skills of the resume, where each topic represents a set of closer skill words. Each topic may have a token vector, i.e., a topic specific vector.

The resume related theme distribution, theme collection and the like are obtained through the sample resume data, the predicted resume characterization information is obtained based on the initial resume characterization information prediction, subsequent training is facilitated, and more accurate resume capability information can be obtained.

In one possible implementation, the loss function of the resume capability prediction model is constructed based on the second topic distribution information, the second distribution latent variable, the predicted resume characterization information, and the second topic set. Parameters of the resume capacity prediction model can be updated by using the loss function of the resume capacity prediction model, so that a more accurate resume capacity prediction model is obtained.

For example, an example of a formula for a loss function for a resume capability prediction model is as follows:

is resume r _i Subject distribution of->

Is a distributed latent variable,/->

Is resume r _i Characterization of beta ^R Is the topic set of the resume. Wherein "|" represents the prior probability problem; "||" indicates juxtaposition. />

Representing a priori distribution +.>

And posterior distribution->

Is a distance of (3).

In one possible implementation, as shown in fig. 7, the processing the multiple capability information and the requirement information by using a requirement matching model to obtain a requirement matching result includes:

s701, obtaining whole capability information based on the plurality of capability information and the attention weight;

s702, obtaining specific skill information based on the overall capability information and skill characterization information;

s703, obtaining a predicted demand matching result based on the specific skill information and the post skill demand information.

For example, the test capability information is

The resume ability information is +.>

Interview ability information is +.>

The whole macroscopic capability of the candidate, namely the whole capability information, is obtained by adopting an attention mechanism and is h _i ∈R ^1×d Wherein->

And h _i May be a vector.

Wherein a is _E 、a _R 、a _A To pay attention to the weights, these weights may be updated during the training process.

The matching degree between the candidate and the post can be identified through the requirement matching model, and the appropriate post can be recommended for the candidate or the appropriate candidate can be recommended for the post based on the overall capability information of the candidate and the specific skill requirement information of the post.

In one possible implementation, the loss function of the demand matching model is constructed based on the predicted demand matching result and the actual demand matching result. Parameters of the demand matching model can be updated by using the loss function of the demand matching model, so that a more accurate demand matching model is obtained.

For example, after obtaining the macroscopic capabilities of the candidate, the technique characterization matrix h ^S Cross-multiplying to obtain candidates such as job seekers' mastery of a particular skill _i ∈R ^1×K ，α _i May be a vector.

α _i ＝sigmoid(h _i ×(h ^S ) ^T )

Skill requirement gamma for post _p ∈R ^1×K Representing post p _p Demand for specific skills, gamma _p May be a vector, the formula of which is exemplified as follows:

one-hot vector, W, which is the position number ^J Is a learnable training matrix.

Finally, the mastery condition vector alpha of the candidate in the specific skill _i Skill demand vector gamma for post _p Para-multiplication, probability of success of predicted application:

wherein alpha is _i,s A mastery condition vector, gamma, representing skills s (s-th skill) _p,s Representation post p _p Is a requirement vector for skill s.

According to the above example, an example of a loss function of a demand matching model is as follows:

wherein y is _i Can represent candidate u _i Whether a successful tag is actually applied for or not,

Can represent model prediction candidates u _i Probability of success of application.

In one possible implementation, as shown in fig. 8, training the demand matching model and the plurality of capability prediction models based on the demand matching result to update parameters of the demand matching model and the plurality of capability prediction models includes:

s801, obtaining an overall loss function based on the loss function of the demand matching model and the loss functions of the multiple capacity prediction models;

s802, under the condition that the integral loss function is determined to need to be updated, updating the demand matching model by using the loss function of the demand matching model, and respectively updating the corresponding capacity prediction models by using the loss functions of the plurality of capacity prediction models.

For example, an example of an overall loss function is as follows:

loss＝α*loss ^E +β*loss ^R +γ*loss ^A +η*loss ^D

where α, β, γ, and η are weights corresponding to the loss functions of the capacity prediction model and the demand matching model, and may be empirical values. In practical application, specific values of alpha, beta, gamma and eta can be adjusted according to convergence speed and the like of each model.

In the embodiment of the disclosure, in the case that the value of the overall loss function is not smaller than the set threshold value, the parameters of the demand matching model and the capacity prediction models may be updated. When the value of the overall loss function is smaller than the set threshold value, the overall loss function converges, and updating of the overall model may be stopped. The multiple models included in the joint capacity model can be jointly trained through the integral loss function and the loss function of each model, and training efficiency is improved.

In one possible implementation, updating the demand matching model with the loss function of the demand matching model includes:

at least one of the attention weight, the overall skill characterization information, and the post skill requirement information in the requirement matching model is updated based on a loss function of the requirement matching model. Updating parameters of the demand matching model can enable the demand matching model to obtain more suitable matching results of candidates and posts.

In embodiments of the present disclosure, based on the above examples, parameters in the demand matching model that need to be updated may include one or more of the following: attention weights, e.g. a _E 、a _R 、a _A Global skill characterization information such as h ^S And post skill demand information such as gamma _j Is embedded in matrix W of (2) ^J . If the loss function of the demand matching model is not less than the set threshold, at least one of the attention weight, the overall skill characterization information, and the post skill demand information in the demand matching model may be updated. In the case where the value of the overall loss function is smaller than the set threshold, the updating of the demand matching model may be stopped.

In one possible implementation, the corresponding capacity prediction model is updated by using the loss functions of the plurality of capacity prediction models, and the method comprises at least two steps of:

Updating at least one of the trial capability information, the question difficulty information, the question distinguishing degree information and the overall skill characterization information of the trial capability prediction model based on a loss function of the trial capability prediction model;

updating at least one of full-connection layer parameters, first topic characterization information, and overall skill characterization information of the interview capability prediction model based on a loss function of the interview capability prediction model;

and updating at least one of full-connection layer parameters, second topic characterization information and overall skill characterization information of the resume capability prediction model based on a loss function of the resume capability prediction model.

For example, if the joint capacity model includes a interview capacity prediction model and an interview capacity prediction model, the parameters of the interview capacity prediction model are updated according to the loss function of the interview capacity prediction model, and the parameters of the interview capacity prediction model are updated according to the loss function of the interview capacity prediction model. For another example, if the joint capacity model includes a try capacity prediction model and a resume capacity prediction model, the parameters of the try capacity prediction model are updated according to the loss function of the try capacity prediction model, and the parameters of the resume capacity prediction model are updated according to the loss function of the resume capacity prediction model.

In the embodiment of the disclosure, in updating of the plurality of models, the overall skill representation information needs to be updated, so the overall skill representation information can be updated in a certain order. For example, the overall skill characterization information is updated in the order of the resume ability prediction model, the interview ability prediction model, and the interview ability prediction model. For another example, the overall skill characterization information is updated in the order of the interview ability prediction model, the resume ability prediction model, and the interview ability prediction model. The whole skill characterization information updated by the previous model can be used as the updating basis of the next model.

By updating the parameters of each capacity prediction model, more comprehensive and more dimensional capacity predictions can be satisfied.

FIG. 9 is a flow diagram of a capability requirement matching method according to an embodiment of the present disclosure. The matching method may include:

s901, processing various data to be processed of a candidate by adopting various capability prediction models to obtain various capability information of the candidate;

s902, processing the plurality of capability information and the requirement information by adopting a requirement matching model to obtain a requirement matching result of the candidate, wherein the requirement information comprises capability information which needs to be met by a target scene.

The capacity prediction model and the demand matching model in this embodiment may be a capacity prediction model and a demand matching model in a joint capacity model that is trained by using any one of the training methods in the foregoing embodiments. The plurality of data to be processed may include, for example, one or more of real try-on data, real interview data, and real resume data of the candidate. For example, the real try-on data, the real interview data and the real resume data of a certain candidate are input into a joint capability model, and various capability prediction models of the joint capability model respectively process the real try-on data, the real interview data and the real resume data of the candidate to obtain try-on capability information, interview capability information and resume capability information of the candidate.

In the embodiment of the disclosure, the combined capacity model obtained after training can meet more comprehensive and multidimensional capacity prediction and demand matching, and is suitable for richer application scenes.

In one possible implementation, as shown in fig. 10, the processing of the multiple types of data to be processed of the candidate by using multiple types of capability prediction models to obtain multiple types of capability information of the candidate includes at least two steps as follows:

S1001, processing the to-be-processed pen test data by adopting a pen test capability prediction model to obtain pen test capability information;

s1002, processing interview data to be processed by adopting an interview capability prediction model to obtain interview capability information;

s1003, processing resume data to be processed by adopting a resume capability prediction model to obtain resume capability information.

In the embodiment of the disclosure, the to-be-processed pen test data may be real pen test data of the candidate, such as a pen test answer record of a school test paper, an on-job test paper, and the like. The interview data to be processed may be real interview data of a candidate such as an interview officer's assessment text of various skills of a certain candidate in the interview process, text related to the skills of a certain candidate extracted from the interview audio-visual data, etc. The resume data to be processed may be real resume data of a candidate such as resume electronic text of a certain candidate.

At least multiple capability information of the candidate can be obtained through prediction by at least two of a written test capability prediction model, an interview capability prediction model and a resume capability prediction model, and then more comprehensive and accurate demand matching is realized based on the multiple capability information of the candidate.

In a possible implementation manner, as shown in fig. 11, the processing of the to-be-processed pen test data by using the pen test capability prediction model to obtain pen test capability information includes:

s1101, extracting candidate features from the to-be-processed pen test data;

and S1102, obtaining the trial capability information of the candidate based on the candidate characteristics.

In the embodiment of the disclosure, the candidate features in the to-be-processed pen test data, namely the real pen test data, can comprise macro characterization of the candidate about the pen test, and the specific capability of the candidate in the pen test can be accurately obtained through the candidate features in the pen test data. For example, the candidate feature extracted from the to-be-processed try data, i.e. the real, may be candidate u ₁ Is a one-hot vector

The candidate's interview capability information may be a macroscopic representation of the candidate with respect to the interview

May be a d-dimensional vector. />

Wherein W is ^U May be an embedding matrix, W ^U May have an initial value. W can be aligned in the subsequent training process ^U Learning and updating are performed, and further, the written test capability information is updated. After training, W is used for predicting the trained W in the model by using the pen test capability ^U The one-hot encoding of the real try data with a candidate may predict the try ability information of the candidate.

In addition, the trained trial capacity prediction model can be dynamically updated based on real trial data, and specific processes can be seen in the process of updating with the training trial data. The construction manner of the calculation loss function of the trial capacity prediction model can also be referred to as the related description in the training method of the above embodiment.

In one possible implementation, as shown in fig. 12, the interview capability prediction model is used to process interview data to be processed to obtain interview capability information, including:

s1201, extracting initial interview characterization information from the interview data to be processed;

s1202, obtaining a first mean value and a first variance based on the interview characterization information;

s1203, obtaining first characterization information conforming to Gaussian distribution based on the first mean value and the first variance, and sampling from the first characterization information to obtain first distribution potential variables;

s1204, obtaining first theme distribution information based on the first distribution potential variable;

s1205, obtaining a first topic collection based on the first topic representation information and the overall skill representation information;

s1206, obtaining predicted interview characterization information based on the first topic distribution information and the first topic set;

S1207, obtaining interview capability information based on the first theme distribution information and the first theme characterization information.

Referring to the example of fig. 5, candidate u is extracted from interview data to be processed, i.e., real interview data ₁ Initial interview characterization information of (a)

As input vector to the interview capability prediction model. Vector->

List the first topicCharacterization information, e.g. topic characterization matrix t ^A And overall skill characterization information such as skill characterization matrix h ^S Cross-multiplying and passing the softmax layer to obtain a first topic set beta ^A . Finally, the first topic distribution information is->

With the first topic set beta ^A Multiplication results in a predicted value, i.e. predicted interview characterization information +.>

Furthermore, interview ability information +.>

Initial interview characterization information, predicted interview characterization information, topic distribution, topic collection and the like related to interviews can be obtained through the real interview data of the candidate, and more accurate interview capability information of the candidate can be obtained.

In addition, the trained trial capacity prediction model may be dynamically updated based on real interview data, and specific procedures may be referred to as procedures updated with the training interview data. The construction of the computational loss function of the interview capacity prediction model can also be seen from the relevant description in the training method of the above embodiment.

In one possible implementation, as shown in fig. 13, processing resume data to be processed by using a resume capability prediction model to obtain resume capability information includes:

s1301, extracting initial resume characterization information from the resume data to be processed;

s1302, obtaining a second mean and a second variance based on the resume characterization information;

s1303, obtaining second characterization information conforming to Gaussian distribution based on the second mean and the second variance, and sampling from the second characterization information to obtain second distribution potential variables;

s1304, obtaining second topic distribution information based on the second distribution potential variable;

s1305, obtaining a second topic collection based on the second topic characterization information and the overall skill characterization information;

s1306, obtaining predicted resume characterization information based on the second theme distribution information and the second theme set;

And S1307, obtaining resume capability information based on the second theme distribution information and the second theme characterization information.

Referring to the example of fig. 6, candidate u is extracted from resume data to be processed, i.e., real resume data ₁ Interview characterization information of (a)

As input vector to the resume capability prediction model. Vector->

Second topic representation information such as topic representation matrix t ^R And overall skill characterization information such as skill characterization matrix h ^S Cross-multiplying and passing the softmax layer to obtain a second topic set beta ^R . Finally, the second topic distribution information->

Furthermore, interview capability information

Initial resume characterization information, predicted resume characterization information, topic distribution, topic collection and the like related to the resume are obtained through the real resume data of the candidate, so that more accurate resume capability information of the candidate can be obtained.

In addition, the trained resume capacity prediction model can be dynamically updated based on the real resume data, and specific processes can be seen in the process of updating by using the training resume data. The construction manner of the computational loss function of the resume capacity prediction model can also be referred to the related description in the training method of the above embodiment.

In one possible implementation, as shown in fig. 14, the processing the multiple capability information and the requirement information by using the requirement matching model to obtain the requirement matching result of the candidate includes:

s1401, obtaining overall capability information based on the plurality of capability information and the attention weight;

s1402, obtaining specific skill information based on the overall capability information and skill characterization information;

s1403, obtaining a predicted requirement matching result based on the specific skill information and the post skill requirement information.

For example, based on candidate u ₁ The written test capability information of (1) is

The resume ability information is +.>

Interview ability information is

The whole macroscopic capability of the candidate, namely the whole capability information, is obtained by adopting an attention mechanism and is h ₁ ∈R ^1×d 。

Wherein a is _E 、a _R 、a _A For attention weights, these weights may be trained weights.

In obtaining macroscopic capability h of candidate ₁ After that, h ₁ And skill characterization matrix h ^S Cross-multiplying to obtain candidates such as job seekers' mastery of a particular skill ₁ ∈R ^1×K 。

α ₁ ＝sigmoid(h ₁ ×(h ^S ) ^T )

Post p ₁ Skills requirement gamma of (2) ₁ ∈R ^1×K The formula of (a) is exemplified as follows:

is post p ₁ One-hot vector, W ^J Is a trained matrix.

Finally, candidate u ₁ Mastery case vector alpha at a particular skill ₁ And post p ₁ Skill demand vector gamma of (2) ₁ Para-multiplication, probability of success of predicted application:

wherein alpha is _1,s Representing candidate u ₁ The mastery condition vector of skill s, gamma _1,s Representation post p ₁ Demand vector gamma of skill s of (2) ₁ 。

FIG. 15 is a schematic structural diagram of a training apparatus for a federated capability model that includes a plurality of capability prediction models and a demand matching model, in accordance with an embodiment of the present disclosure, which may include:

the capability prediction module 1501 is configured to process sample data by using multiple capability prediction models to obtain multiple capability information;

A requirement matching module 1502, configured to process the multiple capability information and requirement information by using a requirement matching model to obtain a requirement matching result, where the requirement information includes capability information that needs to be satisfied by a target scene;

the training module 1503 is configured to train the demand matching model and the multiple capability prediction models based on the demand matching result to update parameters of the demand matching model and the multiple capability prediction models.

In one possible implementation, as shown in fig. 16, the capability prediction module 1501 includes at least two of:

the test capability prediction submodule 15011 is used for processing sample test data by adopting a test capability prediction model to obtain test capability information;

an interview capability prediction submodule 15012, configured to process sample interview data by using an interview capability prediction model to obtain interview capability information;

the resume capability prediction submodule 15013 is configured to process the sample resume data by using a resume capability prediction model to obtain resume capability information.

In one possible implementation, the try ability prediction sub-module 15011 is further configured to extract candidate features from the sample try data; candidate trial capability information is obtained based on the candidate features.

In one possible implementation, the loss function of the test capability prediction model is constructed based on the probability of correct answer predicted by the test capability prediction model and the actual answer result; the probability of correct answer is predicted by the test capacity prediction model based on the interaction information of the candidate obtained by the sample test data and the questions.

In one possible implementation, the try ability prediction submodule 15011 is further configured to obtain interaction information of the candidate with the topic, where the step includes:

extracting candidate features and topic features from the sample pen test data;

obtaining the try-on capability information based on the candidate feature;

obtaining written test skill characterization information based on the written test capability information and the overall skill characterization information;

acquiring topic difficulty information and topic distinguishing degree information based on the topic features;

and obtaining interaction information of the candidate and the topic based on the pen test skill characterization information, the topic difficulty information, the topic distinguishing degree information and the topic association skill information.

In one possible implementation, the interview capability prediction submodule 15012 is configured to:

extracting initial interview characterization information from the sample interview data;

Based on the interview characterization information, a first mean and a first variance are obtained;

obtaining first characterization information conforming to Gaussian distribution based on the first mean value and the first variance, and sampling from the first characterization information to obtain first distribution potential variables;

obtaining first topic distribution information based on the first distribution potential variables;

obtaining a first topic collection based on the first topic representation information and the overall skill representation information;

obtaining predicted interview characterization information based on the first topic distribution information and the first topic set;

and obtaining interview capability information based on the first topic distribution information and the first topic characterization information.

In one possible implementation, the loss function of the interview capability prediction model is constructed based on the first topic distribution information, the first distribution latent variable, the predicted interview characterization information, and the first topic set.

In one possible implementation, the resume capability prediction submodule 15013 is further configured to:

extracting initial resume characterization information from the sample resume data;

obtaining a second mean and a second variance based on the resume characterization information;

obtaining second characterization information conforming to Gaussian distribution based on the second mean and the second variance, and sampling from the second characterization information to obtain second distribution potential variables;

Obtaining second topic distribution information based on the second distribution potential variable;

obtaining a second topic collection based on the second topic representation information and the overall skill representation information;

obtaining predicted resume characterization information based on the second topic distribution information and the second topic set;

and obtaining resume capability information based on the second theme distribution information and the second theme characterization information.

In one possible implementation, the loss function of the resume capability prediction model is constructed based on the second topic distribution information, the second distribution latent variable, the predicted resume characterization information, and the second topic set.

In one possible implementation, as shown in fig. 16, the requirement matching module 1502 includes:

an overall capability sub-module 15021 for deriving overall capability information based on the plurality of capability information and the attention weight;

a specific skills sub-module 15022 for obtaining specific skills information based on the overall ability information and the skill characterization information;

a demand matching sub-module 15023 for obtaining a predicted demand matching result based on the specific skill information and the post skill demand information.

In one possible implementation, the loss function of the demand matching model is constructed based on the predicted demand matching result and the actual demand matching result.

In one possible implementation, the training module 1503 includes:

the overall loss submodule 15031 is used for obtaining an overall loss function based on the loss function of the demand matching model and the loss functions of the multiple capacity prediction models;

an updating submodule 15032, configured to update the demand matching model with a loss function of the demand matching model and update the corresponding capacity prediction models with loss functions of the plurality of capacity prediction models, respectively, when it is determined that updating is required based on the overall loss function.

In one possible implementation, the updating sub-module 15032 is configured to update the demand matching model with a loss function of the demand matching model, including: at least one of the attention weight, the overall skill information, and the post skill requirement information in the requirement matching model is updated based on a loss function of the requirement matching model.

In a possible implementation manner, the updating sub-module 15032 is further configured to update the corresponding capability prediction models with the loss functions of the multiple capability prediction models, respectively, including at least two of the following:

FIG. 17 is a schematic diagram of a capability requirement matching device according to an embodiment of the disclosure, which may include:

the capability prediction module 1701 is configured to process multiple types of data to be processed of a candidate by using multiple capability prediction models, so as to obtain multiple types of capability information of the candidate;

the requirement matching module 1702 is configured to process the multiple capability information and the requirement information by using a requirement matching model to obtain a requirement matching result of the candidate, where the requirement information includes capability information that needs to be satisfied by the target scene.

In one possible implementation, as shown in fig. 18, the capability prediction module 1701 includes at least two of the following:

the test capability prediction submodule 17011 is used for processing test data to be processed by adopting a test capability prediction model to obtain test capability information;

An interview capability prediction submodule 17012, configured to process interview data to be processed by using an interview capability prediction model, so as to obtain interview capability information;

the resume capability prediction submodule 17013 is configured to process resume data to be processed by using a resume capability prediction model to obtain resume capability information.

In one possible implementation, the try ability prediction sub-module 17011 is further configured to extract candidate features from the to-be-processed try data; candidate trial capability information is obtained based on the candidate features.

In one possible implementation, the interview capability prediction submodule 17012 is further configured to:

extracting initial interview characterization information from the interview data to be processed;

In one possible implementation, the resume capability prediction submodule 17013 is configured to:

extracting initial resume characterization information from the resume data to be processed;

In one possible implementation, the requirement matching module 1702 includes:

an overall capability sub-module 17021 for deriving overall capability information based on the plurality of capability information and the attention weight;

a specific skill sub-module 17022 for obtaining specific skill information based on the overall capability information and skill characterization information;

A demand matching sub-module 17023 for obtaining a predicted demand matching result based on the specific skill information and post skill demand information.

For descriptions of specific functions and examples of each module and sub-module of the apparatus in the embodiments of the present disclosure, reference may be made to the related descriptions of corresponding steps in the foregoing method embodiments, which are not repeated herein.

In the related art, expert decisions include: based on past experience, experts (personnel departments and the like) comprehensively decide factors such as the test results of candidates, mastering skills in resume, project experience, interview evaluation reports and the like. However, expert decisions have strong dependence on artificial experience, and have the problems of low decision efficiency, lack of interpretability, strong subjectivity of comprehensive decisions and the like. Moreover, due to the previous experience difference of different experts, different decision preference can be caused, and the expansibility is insufficient.

In the related art, algorithms based on matching simply match the resume information of candidates and the skill requirements of posts using natural language processing techniques, and make decisions based on the matching situation. The scheme can assist an expert in making decisions, but only matches the resume and post information of the candidate, does not comprehensively consider multi-source evaluation data (trial and error evaluation) of the candidate, and lacks a certain interpretability.

The training method of the joint capacity model can be used in the joint capacity diagnosis process in talent recruitment, and the capacity evaluation performance in talent recruitment is improved through joint modeling of the multi-source heterogeneous evaluation result.

Application scenarios include, but are not limited to: and integrating the multisource evaluation results of the candidates to obtain comprehensive evaluation of the candidates, selecting the most suitable candidates according to the skill requirements of different posts, realizing the best matching of posts and talents, improving the talent recruitment efficiency and reducing talent loss caused by competition among posts.

In the joint capacity diagnostic process in talent recruitment, the capacity of the candidate may be modeled based on the candidate's resume data, interview data, and interview data, respectively. Such as resume-based capabilities, pen-test-based capabilities, interview-based capabilities, and the like. And the attention mechanism is adopted to intelligently integrate the multi-source heterogeneous evaluation results of the candidates to obtain comprehensive capability evaluation of the candidates. The expert decision process in the historical recruitment record is restored by predicting how well the comprehensive capabilities of the candidate match the skill requirements of the post (or what is called the post's capability requirements). And, the training learning of the relevant parameters of each model can be realized in iteration by combining the predicted result and the real result, such as the actual job-in result. Referring to fig. 19, the joint capacity diagnostic process in talent recruitment may specifically include the following links.

The specific steps of each link are described in detail below with reference to the main mathematical symbols and descriptions in table 1 in the embodiments of the present disclosure.

Table 1: mathematical symbols and description

/>

Multi-source candidate capability modeling:

1. candidate capability modeling based on pen test data

Assume that N candidates exist in the system, such as job seekers, M-channel problems (which may also be referred to as topics, practice problems, etc.), K skills.

Wherein the candidate number u _i ∈R ^1×N Problem number e _j ∈R ^1×M Skill characterization matrix h ^S ∈R ^K×d ,W ^U 、W ^diff 、W ^disc Respectively a learnable embedding matrix. Candidate u _i Problem e _j The one-hot vectors of (a) are respectively

And->

Macroscopic characterization of candidate with respect to try-on (vector)

R ^1×d The superscript d of (c) indicates the macroscopic capability dimension.

Characterization of candidates on a particular skill (vector) with respect to pen trials

R ^1×K The superscript K of (1) represents the dimension of the skill.

Difficulty parameter (vector) of problem

Degree of distinction parameter of problem

Problem associated skill vector

Representing the skill of the problem investigation. For example, if problem e _j Let's examine skills s->

And the s-th element position of (2) is 1, otherwise is 0.

Project reaction theoretical formula based on educational psychology:

wherein the symbol->

Representing the para-multiplication.

Referring to FIG. 20, for a neural network FCL of interactions between candidates (e.g., students, job seekers, etc.) and problems in a trial capacity prediction model, x is taken as an input to the FCL and output as

I.e. the probability that the prediction is correct.

Modeling the capacity parameters and parametric characterization of the problem of the candidate based on the candidate's trial writing data, and then following the psychology of the project inverseAfter the interaction formula of the candidate and the problem is constructed according to the theory idea, the neural network with stronger fitting capability can be adopted to automatically learn the interaction between the candidate and the problem, and the probability of the candidate to correctly answer the problem is output. And comparing the probability of correctly answering the problem with the true answering result to calculate a loss function, and updating the parameters of the trial capacity prediction model. Candidate ability modeling loss function loss based on pentry ^E ：

as candidate u _i Answer questions e _j Is (are) true>

Predicting candidate u for model _i Answer questions e _j Is the case in (a). Ability to learn each candidate based on minimizing the loss function based on the trial recording +.>

2. Candidate ability modeling based on resume data/interview data:

referring to FIG. 21, the input to the resume capability prediction model is resume r _i Representation (vector)

Including skills extracted from the profile text. For example if skills s are present in resume text +.>

And the s-th element position of (2) is 1, otherwise is 0.

Will be

Respectively passing through two fully connected neural networks to obtain a mean value mu ^R Sum of variances sigma ^R Then, from the line of Gaussian distribution N (mu) ^R ,(σ ^R ) ² ) Is sampled to obtain a distribution latent variable such as Gaussian random vector +.>

Then obtaining the topic distribution (vector) by softmax>

Characterizing (matrix) t ^R And skill characterization (matrix) h ^S Cross-multiplying and cross-softmax to obtain a topic set (matrix) beta ^R Finally, will->

And beta ^R Multiplying to obtain a predicted value.

Theme distribution (vector)

The probability corresponding to each topic represents the mastery under each topic (skill set).

Theme characterization (matrix)

Suppose that K exists in the system for skills in resume ^R Topics, where each topic represents a set of closer skill words, such as convolution, machine learning, and the like. Each topic has a token vector, i.e., a topic specific vector. />

A specific vector representing the topic k.

Distribution of topics in skill words

Representing the proficiency of each topic k in the corresponding skill,

h ^S representing a skill characterization matrix.

Finally, the topic of resume is distributed (vector)

Characterization of topic (matrix) t ^R Multiplying can result in macroscopic capabilities (vectors) of candidates such as job seekers on resume:

interview modeling based on principles similar to resume modeling, interview-assessed topic distributions (vectors)

Characterization of topic (matrix) t ^A Multiplication results in the macroscopic ability (vector) of the candidate on the interview evaluation:

after modeling the candidate's capabilities based on the resume/interview data, the candidate's capabilities may be learned from the candidate's resume/interview based on their resume/interview data. Since the topic model is known to interpret hidden decision logic in an efficient representation with high interpretability, potential capability characterizations embodied in the resume may be mined here based on the topic model. The resume-based candidate capability modeling loss function may be:

is resume r _i Subject distribution of->

Is a distributed latent variable,/->

Is resume r _i Characterization of beta ^R Is the topic set of the resume.

The interview-based candidate capability modeling loss function may be:

is interview a _i Subject distribution of->

Is a distributed latent variable,/->

Representing a priori distribution +.>

And posterior distribution->

Is a distance of (3).

The following is an example of a specific formula reasoning process for the resume's loss function, and the interviews are similar and will not be repeated:

(1) Loss function:

(2) The generation process comprises the following steps: a generative model may be defined for each resume using the above model,

subscript s in the above formula represents a count value of skill; k represents the number of knowledge points, corresponding to the skill dimension K described above.

Wherein, it is assumed that

Is from a Gaussian random vector +.>

The obtained product is obtained by the method,

(3) The reasoning process comprises the following steps: inference of topic distribution probability of resume using variational inference algorithm

The basic idea is to optimize the parameters +.>

Thus using KL distribution to approximate realityPosterior distribution. The variation distribution may be defined herein as a gaussian distribution,

f ₁ ,f ₂ is a parameter of the fully connected neural network layer.

Resume R _i Is embedded in (a)

f _r Is a parameter of the neural network, see fig. 21.

Finally, can be to

Approximating the posterior distribution of (a) and deducing each candidate u by minimizing the evidence of the lower limit _i Is->

The loss function formula of (c) is as follows,

(II) Multi-Source candidate Capacity integration: based on the multi-source capability of the candidate obtained in the step (one), adopting an attention mechanism to learn the emphasis on different capabilities of the candidate respectively, and integrating the multi-source candidate capability into the comprehensive capability of the candidate by using a linear superposition function.

(III) modeling post skill requirements: and extracting the requirement condition of the post on the specific skill based on the information data of the post.

(four) candidate-post matching:

the modeling process of candidate, e.g., job seeker, matching with posts is exemplified as follows:

respectively at pensAbility to model candidates in try, resume, interview data

The attention mechanism is then used to derive the overall macroscopic capability (vector) h of the candidate _i ∈R ^1×d I.e.

Wherein a is _E 、a _R 、a _A Is the attention weight.

After obtaining the macroscopic ability of the candidate, the technique characterization matrix h is used for ^S Cross-multiplying to obtain a candidate's mastery case vector alpha at a particular skill _i ∈R ^1×K ，

α _i ＝sigmoid(h _i ×(h ^S ) ^T )

Skill demand vector gamma for post _p ∈R ^1×K : representation post p _p The need for a particular skill (vector), the formula,

is post p _p One-hot vector, W ^J Is a learnable training matrix.

Finally, the staff is mastered in the specific skill alpha _i Skill demand vector gamma for post _p Multiplying the positions, predicting the job entering probability,

wherein alpha is _i,s A mastery condition vector, gamma, representing skills s (s-th skill) _p,s Representing the positionA demand vector for skills s.

And (3) matching the comprehensive capacity of the candidate obtained in the step (II) with the post skill requirement obtained in the step (III), judging whether the capacity of the candidate is suitable for the post, and predicting whether the candidate successfully applies for the post based on a matching result. Candidate-post matching modeled loss function:

Wherein y is _i Is candidate u _i Whether a successful tag is actually applied for or not,

is a model prediction candidate u _i Probability of success of application.

(V) loss function calculation: and (3) comparing the prediction result obtained in the step (IV) with the actual recruitment situation, finding out all the situations of prediction errors, and calculating a loss function. For example, an example of an overall loss function:

loss＝α*loss ^E +β*loss ^R +γ*loss ^A +η*loss ^D

and (six) updating parameters: and (3) updating parameters of the whole model by using the loss function in the step (V). Joint training: and if the calculation result of the loss function (V) shows that the parameters need to be updated, respectively adjusting the parameters of each model according to the loss function corresponding to each model. If the calculation result of the loss function of (fifth) shows that convergence is needed, all models stop tuning.

After diagnosing the comprehensive ability of the candidate, the solution of the embodiment of the present disclosure may have the following specific application.

(1) Candidate capability characterization:

the knowledge of the candidate's skill in a particular skill can be obtained through a mapping operation, knowing the candidate's weak/skilled skill points.

After obtaining the new skill representation, the potential representation of the comprehensive ability of the candidate can be mapped to the skill to obtain the mastery condition of the candidate on the new skill.

And obtaining the overall mastering condition of the candidate on the related skill set through the linear superposition function.

(2) Talent retrieval service:

the framework can sort the candidates according to the mastering condition of a certain skill combination, and then recommend the optimal K candidates for a specific post according to the requirements of the post on the skill, so that talent loss caused by competition in the same post can be effectively avoided, and the post recruitment is ensured to be the most suitable candidate.

The scheme of the embodiment of the disclosure is applied to talent recruitment scenes and has at least one of the following characteristics:

and (5) comprehensively. The framework based on the scheme of the embodiment of the disclosure can provide a joint capacity diagnosis method in talent recruitment, and compared with a traditional matching method, the method has the advantages that candidate data are considered more comprehensively, and capacity evaluation performance in talent recruitment is improved by modeling candidate capacity from different dimensions.

High efficiency. Based on the framework of the scheme of the embodiment of the disclosure, decision making can be directly carried out, expert decision making can be assisted, and compared with a traditional manual decision making method, recruitment efficiency can be greatly improved.

Interpretability. Based on the framework of the scheme of the embodiment of the disclosure, the mastering condition of the candidate on the specific skill and the requirement condition of the post on the specific skill can be described, a black box does not exist, and the method has good interpretability.

Objectivity. Based on the framework of the scheme of the embodiment of the disclosure, the comprehensive capability of the candidate is modeled based on the multi-source data of the candidate, and compared with the traditional manual expert manual integration multi-source evaluation result, the subjectivity factor is reduced.

And (5) expandability. For different scenes, different candidate data, the framework based on the scheme of the embodiment of the disclosure can be applicable, and the method has strong expandability.

Balance. The framework based on the scheme of the embodiment of the disclosure can sort candidates aiming at specific skill combinations so as to ensure that talents recruited by the posts are the most suitable, and can effectively avoid talent loss caused by competition of the same post.

In the technical scheme of the disclosure, the acquisition, storage, application and the like of the related user personal information all conform to the regulations of related laws and regulations, and the public sequence is not violated.

According to embodiments of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

Fig. 22 shows a schematic block diagram of an example electronic device 2200 that may be used to implement embodiments of the present disclosure. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular telephones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be exemplary only, and are not meant to limit implementations of the disclosure described and/or claimed herein.

As shown in fig. 22, the device 2200 includes a computing unit 2201 that can perform various suitable actions and processes according to a computer program stored in a Read Only Memory (ROM) 2202 or a computer program loaded from a storage unit 2208 into a Random Access Memory (RAM) 2203. In the RAM2203, various programs and data required for the operation of the device 2200 may also be stored. The computing unit 2201, the ROM 2202, and the RAM2203 are connected to each other via a bus 2204. An input/output (I/O) interface 2205 is also connected to bus 2204.

Various components in device 2200 are connected to I/O interface 2205, including: an input unit 2206 such as a keyboard, a mouse, or the like; an output unit 2207 such as various types of displays, speakers, and the like; a storage unit 2208 such as a magnetic disk, an optical disk, or the like; and a communication unit 2209 such as a network card, modem, wireless communication transceiver, or the like. The communication unit 2209 allows the device 2200 to exchange information/data with other devices via a computer network, such as the internet, and/or various telecommunications networks.

The computing unit 2201 may be a variety of general and/or special purpose processing components having processing and computing capabilities. Some examples of computing unit 2201 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various dedicated Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 2201 performs the respective methods and processes described above, such as a training method of the joint capacity model or a capacity demand matching method. For example, in some embodiments, the training method or the capability requirement matching method of the federated capability model may be implemented as a computer software program tangibly embodied on a machine-readable medium, such as storage unit 2208. In some embodiments, some or all of the computer programs may be loaded and/or installed onto device 2200 via ROM 2202 and/or communications unit 2209. When the computer program is loaded into RAM2203 and executed by computing unit 2201, one or more steps of the training method or capability requirement matching method of the joint capability model described above may be performed. Alternatively, in other embodiments, the computing unit 2201 may be configured to perform the training method or the capability requirement matching method of the joint capability model in any other suitable way (e.g., by means of firmware).

Various implementations of the systems and techniques described here above may be implemented in digital electronic circuitry, integrated circuit systems, field Programmable Gate Arrays (FPGAs), application Specific Integrated Circuits (ASICs), application Specific Standard Products (ASSPs), systems On Chip (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs, the one or more computer programs may be executed and/or interpreted on a programmable system including at least one programmable processor, which may be a special purpose or general-purpose programmable processor, that may receive data and instructions from, and transmit data and instructions to, a storage system, at least one input device, and at least one output device.

Program code for carrying out methods of the present disclosure may be written in any combination of one or more programming languages. These program code may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus such that the program code, when executed by the processor or controller, causes the functions/operations specified in the flowchart and/or block diagram to be implemented. The program code may execute entirely on the machine, partly on the machine, as a stand-alone software package, partly on the machine and partly on a remote machine or entirely on the remote machine or server.

In the context of this disclosure, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. The machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and pointing device (e.g., a mouse or trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic input, speech input, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a background component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such background, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), wide Area Networks (WANs), and the internet.

The computer system may include a client and a server. The client and server are typically remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server may be a cloud server, a server of a distributed system, or a server incorporating a blockchain.

It should be appreciated that various forms of the flows shown above may be used to reorder, add, or delete steps. For example, the steps recited in the present disclosure may be performed in parallel, sequentially, or in a different order, provided that the desired results of the disclosed aspects are achieved, and are not limited herein.

The above detailed description should not be taken as limiting the scope of the present disclosure. It will be apparent to those skilled in the art that various modifications, combinations, sub-combinations and alternatives are possible, depending on design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present disclosure are intended to be included within the scope of the present disclosure.

Claims

1. A method of training a joint capacity model, the joint capacity model comprising a plurality of capacity prediction models and a demand matching model, the method comprising:

the plurality of capability prediction models are adopted to respectively process the corresponding sample data to obtain the corresponding capability information;

processing various capability information and demand information by adopting the demand matching model to obtain a demand matching result, wherein the demand information comprises capability information which needs to be met by a target scene;

Training the demand matching model and the multiple capability prediction models based on the demand matching result to update parameters of the demand matching model and the multiple capability prediction models;

the method for processing the multiple capability information and the demand information by adopting the demand matching model to obtain a demand matching result comprises the following steps:

obtaining overall capability information based on the plurality of capability information and the attention weight;

obtaining specific skill information based on the overall capability information and skill characterization information;

obtaining a predicted demand matching result based on the specific skill information and the post skill demand information;

training the demand matching model and the plurality of capacity prediction models based on the demand matching result to update parameters of the demand matching model and the plurality of capacity prediction models includes:

obtaining an overall loss function based on the loss function of the demand matching model and the loss functions of the multiple capacity prediction models;

in the case that it is determined that updating is required based on the overall loss function, updating the demand matching model with the loss function of the demand matching model, and updating the corresponding capacity prediction model with the loss functions of the plurality of capacity prediction models, respectively.

2. The method according to claim 1, wherein the processing of the corresponding sample data with the plurality of capability prediction models to obtain the corresponding capability information comprises at least two steps of:

processing sample pilot data by adopting a pilot capacity prediction model to obtain pilot capacity information;

processing sample interview data by adopting an interview capacity prediction model to obtain interview capacity information;

and processing the sample resume data by adopting a resume capability prediction model to obtain resume capability information.

3. The method of claim 2, wherein processing the sample test data using the test capability prediction model to obtain test capability information comprises:

extracting candidate features from the sample try-on data;

and obtaining the candidate pen test capability information based on the candidate characteristics.

4. A method according to claim 3, wherein the loss function of the trial capacity prediction model is constructed based on the probability that the answer predicted by the trial capacity prediction model is correct and the actual answer result; the probability of correct answering is obtained by predicting interaction information between candidates and questions obtained by the test capacity prediction model based on the sample test data.

5. The method of claim 4, wherein the step of obtaining the candidate interaction information with the topic comprises:

extracting candidate features and topic features from the sample pen test data;

obtaining trial capability information based on the candidate features;

acquiring the pilot skill characterization information based on the pilot capability information and the overall skill characterization information;

and obtaining interaction information of the candidate and the topic based on the pen test skill characterization information, the topic difficulty information, the topic distinction information and the topic association skill information.

6. The method of claim 2, processing sample interview data using an interview capability prediction model to obtain interview capability information, comprising:

interview characterization information extracted from the sample interview data;

obtaining a first mean and a first variance based on the interview characterization information;

7. The method of claim 6, wherein a loss function of the interview capability prediction model is constructed based on the first topic distribution information, the first distribution latent variable, the predicted interview characterization information, and the first topic set.

8. The method of claim 2, processing the sample resume data using a resume capability prediction model to obtain resume capability information, comprising:

resume characterization information extracted from the sample resume data;

obtaining second topic distribution information based on the second distribution potential variables;

9. The method of claim 8, wherein a loss function of the resume capability prediction model is constructed based on the second topic distribution information, the second distribution latent variable, the predicted resume characterization information, and the second topic set.

10. The method of any of claims 1 to 9, wherein the loss function of the demand matching model is constructed based on predicted demand matching results and actual demand matching results.

11. The method of any of claims 1 to 9, wherein updating the demand matching model with a loss function of the demand matching model comprises:

updating at least one of attention weight, overall skill characterization information, and post skill requirement information in the requirement matching model based on a loss function of the requirement matching model.

12. The method according to any one of claims 1 to 9, wherein updating the corresponding capacity prediction model with the loss function of the plurality of capacity prediction models, respectively, comprises at least two steps of:

updating at least one of the test capability information, the question difficulty information, the question distinguishing degree information and the overall skill characterization information of the test capability prediction model based on a loss function of the test capability prediction model;

updating at least one of full-connection layer parameters, first topic characterization information and overall skill characterization information of an interview capability prediction model based on a loss function of the interview capability prediction model;

13. A capability requirement matching method, comprising:

processing the multiple capability information and the requirement information by adopting a requirement matching model to obtain a requirement matching result of the candidate, wherein the requirement information comprises capability information which needs to be met by a target scene;

Wherein the multiple capability prediction model is a multiple capability prediction model in a combined capability model trained by the training method according to any one of claims 1 to 12;

the demand matching model is a demand matching model in a joint capacity model trained by the training method according to any one of claims 1 to 12.

14. The method of claim 13, wherein the processing of the plurality of types of data to be processed of the candidate using the plurality of types of capability prediction models to obtain the plurality of types of capability information of the candidate comprises at least two steps of:

processing the to-be-processed pilot data by adopting a pilot capacity prediction model to obtain pilot capacity information;

processing the interview data to be processed by adopting an interview capability prediction model to obtain interview capability information;

and processing resume data to be processed by adopting a resume capability prediction model to obtain resume capability information.

15. The method of claim 14, wherein processing the to-be-processed pen test data using a pen test capability prediction model to obtain pen test capability information comprises:

extracting candidate features from the to-be-processed pen test data;

16. The method of claim 14, processing interview data to be processed using an interview capability prediction model to obtain interview capability information, comprising:

interview characterization information extracted from the interview data to be processed;

17. The method of claim 14, processing resume data to be processed using a resume capability prediction model to obtain resume capability information, comprising:

resume characterization information extracted from the resume data to be processed;

18. The method of any of claims 13 to 17, wherein processing the plurality of capability information and the requirement information using a requirement matching model to obtain a requirement matching result for the candidate comprises:

and obtaining a predicted demand matching result based on the specific skill information and the post skill demand information.

19. A training apparatus for a joint capacity model, the joint capacity model comprising a plurality of capacity prediction models and a demand matching model, the apparatus comprising:

the capacity prediction module is used for respectively processing the corresponding sample data by adopting the plurality of capacity prediction models to obtain corresponding capacity information;

the demand matching module is used for processing various capability information and demand information by adopting the demand matching model to obtain a demand matching result, wherein the demand information comprises capability information which needs to be met by a target scene;

the training module is used for training the demand matching model and the multiple capacity prediction models based on the demand matching result so as to update parameters of the demand matching model and the multiple capacity prediction models;

wherein, the demand matching module includes:

an overall capability sub-module for obtaining overall capability information based on the plurality of capability information and the attention weight;

a specific skill sub-module, configured to obtain specific skill information based on the overall capability information and the skill characterization information;

the demand matching sub-module is used for obtaining a predicted demand matching result based on the specific skill information and the post skill demand information;

The training module comprises:

the integral loss submodule is used for obtaining an integral loss function based on the loss function of the demand matching model and the loss functions of the multiple capacity prediction models;

and the updating sub-module is used for updating the demand matching model by using the loss function of the demand matching model and respectively updating the corresponding capacity prediction models by using the loss functions of the plurality of capacity prediction models under the condition that the whole loss function is determined to need to be updated.

20. The apparatus of claim 19, wherein the capability prediction module comprises at least two of:

the test capacity prediction sub-module is used for processing sample test data by adopting a test capacity prediction model to obtain test capacity information;

the interview capability prediction sub-module is used for processing sample interview data by adopting an interview capability prediction model to obtain interview capability information;

the resume capability prediction sub-module is used for processing the sample resume data by adopting a resume capability prediction model to obtain resume capability information.

21. The apparatus of claim 20, wherein the try-ability prediction submodule is to extract candidate features from the sample try-data; and obtaining the candidate pen test capability information based on the candidate characteristics.

22. The apparatus of claim 21, wherein the loss function of the trial capacity prediction model is constructed based on a probability that the answer predicted by the trial capacity prediction model is correct and an actual answer result; the probability of correct answering is obtained by predicting interaction information between candidates and questions obtained by the test capacity prediction model based on the sample test data.

23. The apparatus of claim 22, wherein the means for obtaining interaction information of the candidate with the topic comprises:

extracting candidate features and topic features from the sample pen test data;

obtaining trial capability information based on the candidate features;

24. The apparatus of claim 20, the interview capability prediction submodule to:

25. The apparatus of claim 24, wherein the loss function of the interview capability prediction model is constructed based on the first topic distribution information, the first distribution latent variable, the predicted interview characterization information, and the first topic set.

26. The apparatus of claim 20, the resume capability prediction submodule to:

resume characterization information extracted from the sample resume data;

27. The apparatus of claim 26, wherein a loss function of the resume capability prediction model is constructed based on the second topic distribution information, the second distribution latent variable, the predicted resume characterization information, and the second topic set.

28. The apparatus of any of claims 20 to 27, wherein the loss function of the demand matching model is constructed based on predicted demand matching results and actual demand matching results.

29. The apparatus of any of claims 20 to 27, wherein the updating sub-module to update the demand matching model with a loss function of the demand matching model comprises: and updating at least one of attention weight, overall skill information and post skill requirement information in the requirement matching model based on a loss function of the requirement matching model.

30. The apparatus of any of claims 20 to 27, wherein the updating sub-module is configured to update the corresponding capability prediction model with a loss function of the plurality of capability prediction models, respectively, comprising at least two steps of:

31. A capability requirement matching device, comprising:

the demand matching module is used for processing the plurality of capability information and the demand information by adopting a demand matching model to obtain a demand matching result of the candidate, wherein the demand information comprises capability information which needs to be met by a target scene;

Wherein the multiple capability prediction model is a multiple capability prediction model in a joint capability model trained by the training device according to any one of claims 19 to 30;

the demand matching model is a demand matching model in a joint capacity model trained by the training device according to any one of claims 19 to 30.

32. The apparatus of claim 31, wherein the capability prediction module comprises at least two of:

the test capacity prediction sub-module is used for processing test data to be processed by adopting a test capacity prediction model to obtain test capacity information;

the interview capability prediction sub-module is used for processing interview data to be processed by adopting an interview capability prediction model to obtain interview capability information;

the resume capability prediction sub-module is used for processing resume data to be processed by adopting a resume capability prediction model to obtain resume capability information.

33. The apparatus of claim 32, wherein the try-ability prediction submodule is configured to extract candidate features from the to-be-processed try-data; and obtaining the candidate pen test capability information based on the candidate characteristics.

34. The apparatus of claim 32, the interview capability prediction submodule to:

35. The apparatus of claim 32, the resume capability prediction submodule to:

36. The apparatus of any of claims 31-35, wherein the demand matching module comprises:

and the demand matching sub-module is used for obtaining a predicted demand matching result based on the specific skill information and the post skill demand information.

37. An electronic device, comprising:

at least one processor; and

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-18.

38. A non-transitory computer readable storage medium storing computer instructions for causing the computer to perform the method of any one of claims 1-18.