Method and device for recommending a job based on external knowledge in the form of chat records

Abstract

The application discloses a person-post recommendation method and device based on chat record form external knowledge, which is based on a resume document representation, a post description document representation and a chat record representation obtained through coding, predicts a person-post similarity, and thus obtains a person-post recommendation result. Through the application, the gradient signal of the gain regularization term of the person-post similarity prediction task is fed back to the encoder and the matching network, and the gradient of the loss function of the resume recognition task is fed back to the chat record encoder in the encoder, so that the accuracy of the person-post recommendation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to natural language processing technology and recommendation system technology, in particular to a job recommendation method and device based on chat record form external knowledge. BACKGROUND

[0002] The job recommendation task refers to that a machine predicts the probability of matching between a user and a job based on a resume document and a job description document, so as to match a suitable user with a job. However, in addition to the semantic similarity information of the resume document and the job description document, the subjective personalized preference information of the job seeker user is also a key factor to determine whether the user and the job are matched.

[0003] Generally, there are chat records of job seeker users consulting / joining in some jobs on the job recommendation application platform. These chat record information not only has a large quantity, good real-time performance and easy acquisition, but also often contains the job seeking preference information of the user.

[0004] However, the existing job recommendation methods ignore the above-mentioned characteristics of the chat record, resulting in limited accuracy of job matching. SUMMARY

[0005] To solve the above problems, the present application provides a job recommendation method and device based on chat record form external knowledge, which uses chat record information as external knowledge to assist the resume document and the job description document, so as to make more accurate job similarity prediction.

[0006] To achieve the above purpose, the technical content of the present application includes:

[0007] A job recommendation method based on chat record form external knowledge, which predicts job similarity based on the encoded resume document representation, job description document representation and chat record representation, so as to obtain a job recommendation result.

[0008] Further, the resume document representation is obtained by:

[0009] 1) adding a token 'SEP' between each sentence in the resume document, and adding a token 'CLS' at the beginning of the resume document;

[0010] 2) inputting the processed resume document into a pre-trained language model to obtain the resume document representation.

[0011] Further, the chat record representation is obtained by the following steps:

[0012] 1) constructing an improved recurrent neural network model by setting the update gate in the recurrent neural network model to two;

[0013] 2) Collect the chat logs from each round. i ={u i J i ,a i Input the improved recurrent neural network model;

[0014] 3) Use two update gates to control the chat content respectively. i Chatting with J i And combined with resetting the gating to monitor chat behavior a i By embedding the data, we obtain the hidden state representation of each round of chat history;

[0015] 4) Based on the resume document representation, the hidden state representations of each round of chat records are weighted by an attention mechanism to obtain the chat record representation.

[0016] Furthermore, the hidden state representation is obtained through the following steps:

[0017] 1) Based on the hidden state representation h at time i-1 i-1 Chat log at time i i To obtain the updated gate p at time i. i Update the gating q i and reset the gate control v;

[0018] 2) Based on the hidden state representation h at time i-1 i-1 and the update gate p at time i i Update the gating q i Chat content u i Chat partner J i To obtain the parameters at time i.

[0019] 3) Based on the hidden state representation h at time i-1 i-1 Reset gate v at time i i ,parameter Obtain the hidden state representation h at time i. i .

[0020] Furthermore, the similarity between people and positions is predicted through the following steps:

[0021] 1) Using two symmetric feedforward fully connected networks, calculate the similarity s between the resume document representation and the job description document representation. rj And the similarity s between the chat history representation and the job description document representation. dj ;

[0022] 2) Calculate the similarity s using another feedforward fully connected network. rj Similarity s dj The weighting coefficient λ between them;

[0023] 3) according to the similarity s rj , the similarity s dj and the weight coefficient λ, a person-post similarity is obtained.

[0024] Further, a person-post recommendation model is constructed to obtain the resume document representation, the post description document representation, the chat record representation and the predicted person-post similarity, wherein the person-post recommendation model is trained by the following steps:

[0025] 1) a plurality of sets of training data are obtained, wherein each set of training data comprises a plurality of resume document samples, a post description document sample and a chat record sample, and a resume document sample in the plurality of resume document samples matches the post description document sample and the chat record sample;

[0026] 2) the matched resume document sample and the post description document sample are set as a positive sample pair; and the unmatched resume document sample and the post description document sample are set as a negative sample pair;

[0027] 3) one mini-batch is sampled from each of the positive sample pair and the negative sample pair, and a cross-entropy loss L c is calculated by maximizing the matching score of the positive sample pair and minimizing the matching score of the negative sample pair;

[0028] 4) a gain regularization term loss L reg is calculated by using the higher person-post similarity of the positive sample pair when using the chat record sample and the lower person-post similarity of the negative sample pair when not using the chat record sample;

[0029] 5) according to the cross-entropy loss L c and the gain regularization term loss L reg , a person-post similarity prediction loss L is obtained;

[0030] 6) for a chat record encoder obtaining the chat record representation, a resume recognition task loss L r is calculated by using the similarity score between the matched resume document sample and the chat record sample and the similarity score between the unmatched resume document sample and the chat record sample;

[0031] 7) the person-post similarity prediction loss L and the resume recognition task loss L r are used for multi-task learning by using a step-by-step training method, and parameters in the person-post recommendation model are obtained.

[0032] Further, the similarity score between the matched resume document sample and the chat record sample is calculated by the following steps:

[0033] 1) encode the resume document sample representation using word vector summation

[0034] 2) according to the resume document sample representation and the chat record sample representation d, calculate a similarity score.

[0035] Further, a gain regularization term L is calculated reg or the resume identification task loss L r When used, the loss function includes: hinge loss.

[0036] A storage medium, the storage medium has a computer program stored therein, wherein the computer program is set to execute the above method when running.

[0037] An electronic device comprising a memory and a processor, wherein the memory stores a program for executing the above method.

[0038] Compared with the prior art, the application has the advantages and positive effects as follows:

[0039] 1) The present application is aimed at the person-post recommendation task, based on the user resume document and the post description document, the user's job search preference is mined from the user's chat record information, so that the person-post similarity is more accurately predicted.

[0040] 2) The present application introduces an identifier module to ensure that the information mined from the chat record by the model can be associated with the user's resume.

[0041] 3) The present application introduces a gain regularization term in the loss function to ensure that the information mined from the chat record by the model can have a gain for the person-post recommendation task. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 An example diagram for the job seeker's chat record.

[0043] Figure 2 A model architecture diagram for assisting the resume document and the post description document in the person-post recommendation task based on the user's chat record.

[0044] Figure 3 The specific structure of the chat record encoder of the encoding module in the model.

[0045] Figure 4 The pseudo code of the training method for the person-post recommendation model proposed in the present application.

[0046] Figure 5The following is a case study of how the method provided in the present application utilizes chat record information to make more accurate job-person similarity predictions. DETAILED DESCRIPTION

[0047] The present application will be described in detail below with reference to the accompanying drawings and examples. It should be noted that the described examples are only intended to facilitate understanding of the present application and do not limit the present application in any way.

[0048] The job recommendation method of the present application includes a job recommendation model based on user chat records and a set of training algorithms for the proposed model.

[0049] In a first aspect, the present application provides a job recommendation model based on user chat records, which includes:

[0050] The encoding module can be specifically divided into three parts: resume document encoder, job description document encoder, and chat record encoder. The resume document encoder first performs word embedding calculation on each word in the resume document, and then calculates the resume document vector representation using each word vector in the document. The job description document encoder adopts a structure symmetrical to the resume document encoder to calculate the job description document vector representation. The chat record encoder first encodes each round of chat record. Each round of conversation includes three parts of information: chat content, chat object, and chat behavior. The chat content is a sentence spoken by the job seeker user or the job recruiter, the chat object is the job document consulted or communicated by the job seeker user in the round of chat, and the chat behavior represents whether the round of conversation comes from the job seeker user or the job recruiter. For the three parts of information in each round of conversation, the chat record encoder calculates the chat content sentence vector representation, the chat object document vector representation, and the chat behavior embedding representation, and takes the three vector representations as the representation of the round of chat. After obtaining the representation of each round of chat, the chat record encoder integrates the representations of each round of chat using an improved recurrent neural network, and calculates the chat record vector representation using an attention mechanism.

[0051] The matching module is used to predict the matching degree between the job seeker user and the job by using the resume document representation, the job description document representation, and the chat record representation calculated by the encoding module. Specifically, it includes three feedforward fully connected networks for predicting the similarity between the resume document and the job description document, the similarity between the user chat record and the job description, and the weight coefficient of the above two similarities. The final prediction result of the job-person similarity by the model is the weighted average of the two similarities by the weight coefficient.

[0052] In a second aspect, for the above-mentioned person-post recommendation model, the present application also proposes a set of model training algorithms. In order to enable the model to effectively encode from the user's chat records and to utilize the chat record information to assist the resume document and the post description document to make more effective person-post similarity prediction, in the training process, in addition to training the model to learn to predict the person-post similarity, the present application also introduces an auxiliary task to enhance the information encoded from the chat records, which is both associated with the user's resume and complementary to the user's resume, so as to bring gain to the prediction of the person-post similarity.

[0053] In order to train the model to correctly predict the person-post matching degree, for the person-post similarity prediction result output by the matching module, the present application uses the cross-entropy loss function to maximize the similarity score between the matched person-post sample pairs and minimize the similarity score between the unmatched person-post sample pairs.

[0054] In order to ensure that the information encoded from the chat records by the model is associated with the user's resume, the present application introduces an identifier auxiliary module and a resume identification auxiliary task to the model. Specifically, the present application trains the model to identify which candidate resume belongs to the job seeker user based on the encoded chat record vector representation. The resume identification auxiliary task uses pair-wise hinge loss as the loss function for training, and the gradient signal is back-propagated to the chat record encoder, so as to constrain the chat record encoder to encode information associated with the user's resume.

[0055] In order to ensure that the information encoded from the chat records by the model is complementary to the user's resume and helps the prediction of the person-post similarity, the present application introduces an additional gain regularization term based on the cross-entropy loss function of the person-post similarity. Specifically, on the basis of maximizing the similarity score between the matched person-post sample pairs and minimizing the similarity score between the unmatched person-post sample pairs, the present application also constrains the difference between the person-post similarity prediction result in the matching module and the similarity prediction result of the resume document and the post description document, that is, on the matched person-post sample pairs, the person-post similarity prediction result after introducing the chat record is higher than the similarity prediction result of the resume document and the post description document, and on the unmatched person-post sample pairs, the person-post similarity prediction result after introducing the chat record is lower than the similarity prediction result of the resume document and the post description document. In actual application, this regularization term uses pair-wise hinge loss as the loss function.

[0056] The model proposed in the present application includes two calculation processes in the training process: one is to sample a positive and negative person-post sample pair for each mini-batch, which is used to calculate the person-post similarity cross-entropy loss and the gain regularization term loss, and the sum of the two is used as the loss function to update the parameters of the encoding module and the matching module in the model; the second is to sample a mini-batch from all job seeker users, which is used to calculate the pair-wise hinge loss loss in the resume recognition task, and update the chat record encoder parameters in the model. During the training process, the present application calculates the two calculation processes in an alternating iterative manner. In the inference process, only the first calculation process is retained, that is, for each person-post sample pair, the person-post similarity prediction result calculated by the matching module is taken as the output.

[0057] Exemplary, Figure 1 The historical chat record of the user in the person-post recommendation platform is shown in the figure. As Figure 1 shown, from the chat record information, it can be seen that the expected salary of the job seeker user is 15k and the user is unwilling to travel frequently, and these preference information does not exist in the user resume document. Therefore, the historical chat record can be used as a supplement to the resume information, especially the preference information. For a given person-post sample pair, when predicting whether the similarity of both parties will reach a match, we will consider the content of the chat record between the job seeker user and other some posts of the recruiter in the past, and mine the preference information of the user from the chat record, and use these preference information to combine the resume document and the post description document information to predict the person-post similarity.

[0058] In the person-post recommendation task scenario studied in the present application, each post corresponds to a post description document J composed of multiple sentences related to post responsibilities and post requirements. Each job seeker resume document R is composed of multiple sentences from personal statements, project experience, work experience, etc. In addition, each user also corresponds to a historical chat record D = {d1, d2,..., d n}。 Each round d i in the chat record can be represented as a triple d i = {u i , J i , a i}, where is a sentence of chat content, J i represents the consultation or communication of the job seeker to which post, i.e. the chat object, a i is the behavior information, indicating whether the person speaking in this round of conversation is a job seeker user or a recruiter, a i = 1 represents the job seeker, and a i = 0 represents the recruiter.

[0059] For each person-job sample pair, there is a class label y e {0, 1} corresponding to whether the pair is matched or not. The goal of our model is to predict the probability of a person-job pair being matched based on the user resume document R, the job description document J, and the user’s historical chat record information D. Note that the user’s chat record is with other jobs that the user has previously interacted with, and does not contain the job that the model is trying to predict whether the user is matched with.

[0060] Figure 2 The person-job recommendation model based on user chat records proposed in this application is used to assist the matching of resume documents and job description documents. The model consists of three modules, namely a representation learning module, a matching module, and an identifier module. The representation learning module includes three parts, two document encoders to encode the resume document and the job description document respectively, and a dialogue encoder to encode the chat record. In order to ensure that the representation of the chat record calculated by the chat record encoder has sufficient relevance to the user resume document, we introduce an identifier module. The training task of the identifier is to identify which one of a set of candidate resumes belongs to the job seeker user according to the chat record representation. During training, the identifier will backpropagate the gradient signal to the dialogue encoder to supervise the chat record encoder to encode information associated with the resume. The matching module is used to predict the degree of person-job matching. Specifically, the model predicts the similarity between the job description document and the resume document, the similarity between the job description document and the chat record representation, and a weight coefficient. By weighting the above two similarities, the model’s predicted person-job similarity can be calculated. In addition, in order to ensure that the encoded representation has sufficient gain for the person-job recommendation task, we propose a regularization term to constrain the model to have a higher match score for the resume document-chat record-job document than for the resume document-job document on positive samples. Symmetrically, the situation on negative samples can be obtained. The specific details of the representation learning module, the matching module, and the identifier module of the model proposed in this application will be introduced below.

[0061] Representation learning module:

[0062] The representation learning module comprises two parts: a document encoder and a chat log encoder. The document encoder is used to encode resume documents and job description documents. Specifically, this application uses the BERT model as the encoder. Taking the resume document as an example, this application adds the token 'SEP' between sentences in the document and the token 'CLS' at the beginning of the document. The processed document is then input into the resume document encoder. After processing through the embedding layer and multiple Transformer layers of the BERT model, the final hidden state of the 'CLS' token is used as the resume document representation. Using the same method, this application can calculate the job description document representation using another set of BERT network parameters.

[0063] Correspondingly, the role of the chat log encoder is to calculate the vector representation of the chat log. For example... Figure 1 As shown, each round of chat d in chat history D i Includes a chat message u i A chat partner (job title) J i And a chat action icon 'a' i To capture what job seekers say / hear in each round of conversation and which job position the user is chatting for, we use two BERT-based encoders to encode the chat content sentences and the chat object (job position) document, respectively. Consistent with the previous approach, we also utilize the last hidden state representation of the 'CLS' token as the representation of the previous chat content sentences. i And the chat partner's job description document J i To integrate the chat content and chat partner information from each round of chat and thus compute the chat record representation, we propose a Dialogue Tracking Gated Recurrent Units (DT-GRU) network.

[0064] like Figure 3 As shown, DT-GRU represents the hidden state of the entire dialogue up to time i-1 as h. i-1 And use the chat content to express u i J represents the chat partner i Update the hidden state. Furthermore, to capture whether the job seeker is speaking or listening in each round of the chat, DT-GRU introduces two behavioral embeddings. i These correspond to speaking and listening, respectively. The behavior embeddings will be trained together with other parameters in the network.

[0065]

[0066] Equation (1) describes the hidden state update process under the gating mechanism of DT-GRU, where pi with q i are two update gates, controlling the information source from the chat content and the chat object respectively, v i is a reset gate. All the W in the formula represent weight matrices, ⊙ denotes the product of two vectors by dimension, and σ is a sigmoid function. By using the gating mechanism in the DT-GRU, the model can selectively extract preference information from the chat content and the chat object information on the premise of realizing that the job seeker is speaking or listening in each round.

[0067] After calculating the hidden layer state of the DT-GRU at each time, the present application uses the attention mechanism to calculate the chat record representation:

[0068]

[0069] The matching module:

[0070] For a person-job sample pair, the model can obtain the resume document vector R, the job description document vector J, and the chat record vector d through the representation learning module. The role of the matching module is to use the three vector representations to predict the person-job matching similarity. Specifically, the matching module first uses two symmetric feedforward fully connected networks to predict the job document-resume document similarity s rj , the job document-chat record similarity s dj , and uses another feedforward fully connected network and the weight coefficient λ between the two similarities. The prediction result s of the person-job similarity is the weighted average of the two similarities:

[0071]

[0072] The recognizer module:

[0073] In order to ensure that the chat record encoder can obtain information associated with the user's resume from the chat record, the present application designs an auxiliary recognizer module in addition to the representation learning module and the matching module. The goal of the recognizer module is to select the resume of the job seeker user from multiple candidate resume representations according to the chat record representation d. The gradient signal of the recognizer will be fed back to the chat record encoder, so that the chat record encoder can encode as much information associated with the user's resume as possible from the chat record.

[0074] Since the resume representation is used in the attention mechanism of the model when calculating the chat record representation, if the candidate resume representation input to the recognizer is calculated by the resume encoder, the recognizer only needs to remember which resume is used to calculate the chat record representation. Due to this information leakage problem, the present application uses the word vector addition method to obtain the resume representation Instead of using the resume encoder to compute the representation.

[0075] The recognizer uses a fully connected network to compute the similarity between the chat record representation and each candidate resume representation:

[0076]

[0077] where v and W r are the model parameters in the recognizer. During training, the gradient signal will also be backpropagated to the chat record encoder to update the network parameters of the chat record encoder when updating the model parameters of the recognizer. Since the parameters of the recognizer itself only have a matrix W r and a vector v, the model can constrain the chat record encoder to produce a representation that is more conducive to the identification task of the model, i.e., a representation that is more relevant to the resume of the job seeker. In this way, under the joint action of the chat record encoder and the recognizer, it can be ensured that the chat record representation computed by the model has sufficient relevance to the user's resume.

[0078] So far, the present application has introduced the specific details of each module of the proposed person-post recommendation model. The training algorithm designed for the proposed model will be introduced below.

[0079] As Figure 4 shown, for a given person-post sample pair, the model proposed by the present application includes two computing processes. One is to use each encoder in the representation learning module to encode the resume document, the job description document, and the chat record, and then use the matching module to predict the person-post similarity. The second is to input the chat record representation and multiple candidate resume representations computed by the model into the recognizer to let the recognizer compute the similarity between the chat record representation and the resume representation, and select the one that belongs to the user's own resume representation.

[0080] For the first computing process, at each step of the training process, the present application samples one mini-batch for the matched positive person-post sample pair and the unmatched person-post sample pair. In the loss function, in addition to maximizing the matching score of the positive sample and minimizing the matching score of the negative sample, we also add a regularization term L reg to ensure that the introduction of chat record information can bring gains to the model.

[0081]

[0082] where s + and s - represent the similarity predicted by the model using the resume document, the job description document, and the chat record information on the positive and negative person-post sample pairs, respectively, and L c is the cross-entropy loss. and respectively represent the similarity scores predicted by the resume document and the job description document in the positive and negative person-job sample pairs, L reg The core idea is to constrain the model to obtain higher similarity scores on positive samples and lower similarity scores on negative samples after introducing the chat record, so as to constrain the model to obtain the gain of the chat record information obtained by the model encoding on the person-job recommendation task. The loss function L in the model training process is the cross-entropy loss L c and the gain regular term L reg , where Θ is the weight coefficient.

[0083] For the second computing process, at each step in the training, the present application samples a mini-batch in all person-job sample pairs. For the resume encoder, the matching module, and the job encoder in the model, the parameters of these modules are fixed, and the parameters of the chat record encoder and the identifier are updated, so as to constrain the chat record encoder to obtain as much information as possible from the chat record related to the user resume. Specifically, for a chat record representation calculated by the chat record encoder in the model, the present application randomly samples a resume as a negative sample in addition to the resume of the job seeker user. For the two resumes, the model calculates the similarity scores and respectively using formula (4), and uses the hinge loss as the loss function for training.

[0084]

[0085] In this way, the model in the present application can obtain the part of the chat record related to the user resume and having additional gain on the person-job recommendation task, and use the encoded chat record representation to assist in predicting the person-job similarity. In the training process, for each loop, we perform the first process for n c steps and the second process for n r steps. When testing, for each person-job sample pair, we only perform the first computing process, that is, to encode the resume document, the job description document, and the chat record, and use the matching module to predict the person-job similarity.

[0086] To further illustrate the effectiveness of the model and the training method proposed in the above, the present application displays the calculation results of the model on actual cases. As Figure 5 shown, the case includes a resume of a job seeker user, a chat record, a job matched with the user, and a job not matched with the user. The s rj , s dj, λ, s have the same meaning as formula (3), s' represents the prediction result of the model proposed in the application after removing the chat record model. The job seeker in the case has video editing and shooting skills and some practical experience. In the chat record, the job seeker communicated with the recruiters of two video production related job positions. It can be seen from the chat record that the job seeker is a graduate and he pays more attention to the company's welfare and treatment in insurance and housing. The two candidate positions in the case are both related to video editing and require a bachelor's degree and video editing skills. The difference between the two candidate positions is that the first position is open to graduates and explicitly states that it will provide insurance and housing benefits, while the second position requires work experience.

[0087] However, if the model does not introduce external knowledge from the chat record, since the user's skills meet the requirements of the two candidate positions, the user's matching degree with the two positions is quite high in this case. As can be seen from the prediction result s' of the model after removing the chat record information in this case, the model scores for the two candidate positions are close and both exceed the threshold value 0.5. After introducing the external knowledge from the chat record, the model can capture the reason why the second candidate position does not match the job seeker. On the one hand, the chat record mentions that the job seeker is a graduate, which cannot meet the requirement of work experience for the candidate position; on the other hand, the chat record mentions that the job seeker values the welfare and treatment of the position, while the position does not mention this.

[0088] Although the model predicts the position document-resume document similarity s rj is still close on the two candidate position documents and also exceeds the threshold value 0.5, the model predicts a lower chat record-position document similarity score s dj on the second candidate position, and when calculating the weight λ, a lower weight is assigned to s rj , so as to correctly predict that the second candidate position should not be matched.

[0089] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art should understand that any modification, equivalent replacement or improvement within the spirit and principle of the present application should be included in the protection scope of the present application, and the protection scope is defined by the claims.

Claims

1. A method for recommending job candidates based on external knowledge from chat logs, characterized in that, Based on the encoded resume document representation, job description document representation, and chat log representation, the similarity between the candidate and the job is predicted, thereby obtaining candidate-job recommendation results; the chat log representation is obtained through the following steps: An improved recurrent neural network model is constructed by setting two update gates in the recurrent neural network model. The chat logs from each round d i ={u i J i ,a i The input is an improved recurrent neural network model; where a BERT-based encoder encodes the chat content and the job postings discussed or communicated in the chat, and the last hidden state representation of the CLS-corresponding token is used as the sentence representation of the chat content. i And the chat partner's job description document J i The chat content is a sentence spoken by a job seeker or recruiter, and the chat behavior embedding represents a. i ∈[0,1] indicates whether the person in this round of conversation is a job seeker or a recruiter; Two update gates are used to control the chat content sentence representation u. i Job document representation with chat partner J i And combined with resetting the gating, the chat behavior is embedded to represent a i The embedding of the data yields the hidden state representations of each round of chat records; the process of obtaining the hidden state representations of each round of chat records includes: Based on the hidden state representation h at time i-1 i-1 Chat log at time i i To obtain the updated gate p at time i. i Update the gating q i and reset gate control v i ; Based on the hidden state representation h at time i-1 i-1 and the update gate p at time i i Update the gating q i Chat content sentences represent u i Chatting object job document representation J i To obtain the parameters at time i. Based on the hidden state representation h at time i-1 i-1 Reset gate v at time i i ,parameter Obtain the hidden state representation h at time i. i ; Based on the resume document representation, the hidden state representations of each round of chat history are weighted by an attention mechanism to obtain the chat history representation; Predict person-job similarity using the following steps: Using two symmetric feedforward fully connected networks, the similarity s between the resume document representation and the job description document representation is calculated. rj And the similarity s between the chat history representation and the job description document representation. dj ; Calculate the similarity s using another feedforward fully connected network. rj Similarity s dj The weighting coefficient λ between them; Based on similarity s rj Similarity s dj The similarity between people and positions is obtained by using the weighting coefficient λ.

2. The method as described in claim 1, characterized in that, The resume document is presented as follows: 1) Add the token 'SEP' between sentences in the resume document and add the token 'CLS' at the beginning of the resume document; 2) Input the processed resume document into the pre-trained language model to obtain the resume document representation.

3. The method as described in claim 1, characterized in that, A person-job recommendation model is constructed to obtain the resume document representation, job description document representation, and chat history representation, and to predict the person-job similarity. The person-job recommendation model is trained through the following steps: 1) Obtain several sets of training data, each set of training data including: several resume document samples, one job description document sample and one chat record sample, and one resume document sample among the several resume document samples is matched with the job description document sample and the chat record sample. 2) Set the matching resume document samples and job description document samples as positive sample pairs; set the non-matching resume document samples and job description document samples as negative sample pairs; 3) Sample one mini-batch from each of the positive and negative sample pairs, and calculate the cross-entropy loss L by maximizing the matching score of the positive sample pairs and minimizing the matching score of the negative sample pairs. c ; 4) Positive samples using chat log samples have higher person-job similarity compared to positive samples using chat log samples, and negative samples without chat log samples have lower person-job similarity compared to negative samples without chat log samples. Calculate the gain regularization term loss L. reg ; 5) Based on the cross-entropy loss L c With gain regularization term loss L reg The loss L for predicting the similarity between people and positions is obtained. 6) For the chat record encoder that obtains the chat record representation, calculate the resume recognition task loss L by using the similarity scores between matched resume document samples and chat record samples, and the similarity scores between unmatched resume document samples and chat record samples. r ; 7) Using a step-by-step training method, the loss L for person-job similarity prediction and the loss l for resume recognition task are adjusted. r Multi-task learning is performed to obtain parameters from the job recommendation model.

4. The method as described in claim 3, characterized in that, The similarity score between the matched resume document sample and the chat log sample is calculated using the following steps: 1) Using word vector summation, the resume document sample representation is obtained through encoding. 2) Based on the resume document sample Given a chat log sample d, calculate a similarity score.

5. The method as described in claim 3, characterized in that, Calculate the gain regularization term loss L reg Or resume recognition task loss L r When using this loss function, the following loss function is used: hinge loss.

6. A storage medium storing a computer program, wherein, The computer program is configured to execute any of the methods described in claims 1-5 at runtime.

7. An electronic device comprising a memory and a processor, the memory storing a computer program, the processor being configured to run the computer program to perform the method as claimed in any one of claims 1-5.

Images