Search

CN-122020191-A - Candidate recommendation model construction method and device and terminal equipment

CN122020191ACN 122020191 ACN122020191 ACN 122020191ACN-122020191-A

Abstract

The application discloses a candidate recommendation model construction method, a candidate recommendation model construction device and terminal equipment, belonging to the field of artificial intelligence, wherein the method comprises the following steps: according to the resume feature vector of each candidate and the post feature vector of the target post, calculating to obtain the matching confidence coefficient of each resume text, and taking the resume text with high matching confidence coefficient as an alternative resume text; the method comprises the steps of carrying out rule verification on each candidate resume text, calculating to obtain rule matching degree, screening candidate resume texts with matching confidence degree and rule matching degree exceeding corresponding thresholds as candidate resume texts, and training a reinforcement learning model according to intention data of each candidate, historical business transaction data and post feature vectors of a target post, and resume feature vectors and rule verification results corresponding to each candidate resume text to obtain a candidate recommendation model. By implementing the application, the problems of long time consumption and incomplete talent coverage caused by the fact that the traditional hunter service relies on manual processing of unstructured texts can be solved.

Inventors

  • QI CHAOYU

Assignees

  • 广州叮咚科技集团有限公司

Dates

Publication Date
20260512
Application Date
20260104

Claims (10)

  1. 1. A candidate recommendation model construction method, comprising: acquiring historical business transaction data of a target post, post feature vectors and target post rules, and resume text, resume feature vectors and intention data of each candidate; Calculating to obtain the matching confidence coefficient of each resume text according to each resume feature vector and the post feature vector, and taking the resume text with the matching confidence coefficient meeting a first screening condition as an alternative resume text, wherein the first screening condition is that the matching confidence coefficient exceeds a preset confidence coefficient threshold value; For each alternative resume text, carrying out rule check on the alternative resume text according to the target post rule to generate a rule check result, and calculating to obtain rule matching degree based on the rule check result; screening out candidate resume texts with the matching confidence degree meeting the first screening condition and the rule matching degree meeting the second screening condition as candidate resume texts, wherein the second screening condition is that the rule matching degree exceeds a preset matching degree threshold value; Training a preset reinforcement learning model according to the intention data, the historical business transaction data, the post feature vector, the resume feature vector corresponding to each candidate resume text and the rule verification result, and taking the reinforcement learning model after training as a candidate recommendation model.
  2. 2. The method for constructing a candidate recommendation model according to claim 1, wherein training a preset reinforcement learning model according to the intention data, the historical business transaction data, the post feature vector, and the resume feature vector and rule check result corresponding to each candidate resume text, and taking the trained reinforcement learning model as the candidate recommendation model comprises: For each candidate resume text, the post feature vector, the resume feature vector corresponding to the candidate resume text and the post history recommendation success rate are combined to form a fusion feature vector; Repeatedly executing training operation on a preset reinforcement learning model according to the fusion feature vector, the historical business transaction data, the rule check result and the intention data until reaching the preset iteration times, and obtaining a reinforcement learning model after training; taking the reinforced learning model after training as the candidate recommendation model; wherein the training operation comprises: Updating the real-time iteration times; inputting the fusion feature vector into a current value network to obtain the state value output by the current value network, wherein the current value network is an initial value network after the reinforcement learning model is initialized when training operation is performed for the first time; for the candidate corresponding to each candidate resume text, calculating to obtain the total rewards of the current candidate according to the historical business transaction data, the rule check result and the intention data of the current candidate; Inputting the fusion feature vector into a current strategy network to obtain the action probability output by the current strategy network, wherein the current strategy network is an initial strategy network after the reinforcement learning model is initialized when training operation is performed for the first time; Calculating to obtain a current objective function value according to the action probability output by the strategy network during the previous training operation, the action probability output by the current strategy network, the total rewards and the state value; According to the current objective function value, adjusting the current strategy network to obtain an updated strategy network; Judging whether the current real-time iteration number reaches the preset iteration number, If yes, terminating the training operation, and storing the updated strategy network to obtain a reinforced learning model after training; if not, taking the updated strategy network as the strategy network in the next training operation, and carrying out the next training operation.
  3. 3. The method of claim 2, wherein the historical business transaction data comprises post history recommendation success rate and post history average return rate; The step of calculating the total rewards of the current candidate according to the historical business transaction data, the rule check result and the intention data of the current candidate, comprising the following steps: Calculating to obtain a service rewarding value according to the post history recommending success rate, the post history average refund rate and the estimated acceptance rate of the current candidate; calculating to obtain a rule rewarding value according to the rule checking result; and calculating to obtain the total rewards of the current candidate according to the service rewards and the rule rewards.
  4. 4. The candidate recommendation model construction method as claimed in claim 3, wherein said calculating a current objective function value based on a probability of action output by a policy network at a previous training operation, a probability of action output by a current policy network, said total rewards and said status value comprises: calculating to obtain a current probability ratio according to the action probability output by the strategy network in the previous training operation and the action probability output by the current strategy network; Calculating to obtain a dominance function value according to the total rewards and the state value; and calculating to obtain a current objective function value according to the current probability ratio and the dominance function value.
  5. 5. The candidate recommendation model construction method as defined in claim 1, wherein the post feature vector and resume feature vector generation process comprises: acquiring a post description text of a target post; vectorizing each post description text through a feature coding model to generate post feature vectors corresponding to each post description text; Vectorizing each resume text through a feature coding model to generate resume feature vectors corresponding to the resume texts; the construction process of the feature coding model comprises the following steps: The method comprises the steps of obtaining a person post matching annotation data set, wherein the person post matching annotation data set comprises historical resume texts, historical business adaptation capability annotation weights and historical cross-industry adaptation capability annotation weights of a plurality of candidates, historical post description texts, historical business demand capability annotation weights and historical cross-industry demand capability annotation weights of a plurality of posts and a plurality of person post matching annotation data; carrying out structuring treatment on each history post description text to obtain a history basic post feature vector corresponding to each history post description text; Combining historical business demand capability labeling weights and historical cross-industry demand capability labeling weights of all posts, respectively carrying out feature dimension expansion on each historical basic post feature vector to generate a plurality of historical post feature vectors; Carrying out structuring processing on each history resume text to obtain a history basic resume feature vector corresponding to each history resume text; combining the historical business adaptation capability labeling weight and the historical cross-industry adaptation capability labeling weight corresponding to each candidate, and respectively carrying out feature dimension expansion on each historical basic resume feature vector to generate a plurality of historical resume feature vectors; Training a preset neural network model according to the historical post feature vectors of the candidates, the historical resume feature vectors of the posts and the post matching labeling data, and taking the trained neural network model as the feature coding model.
  6. 6. The method for constructing a candidate recommendation model according to claim 1, wherein the calculating a confidence of matching each resume text according to each resume feature vector and the post feature vector comprises: and for each resume text, calculating cosine similarity between the resume feature vector corresponding to the current resume text and the post feature vector, and determining the matching confidence of the current resume text according to the cosine similarity.
  7. 7. The candidate recommendation model construction method as defined in claim 1, wherein each target post rule corresponds to a priority, and different priorities correspond to different weight coefficients; And performing rule verification on the alternative resume text according to the target post rule to generate a rule verification result, and calculating rule matching degree based on the rule verification result, wherein the rule matching degree comprises the following steps: identifying the matching degree between the alternative resume text and each target post rule one by one, and generating a rule check result corresponding to each target post rule, wherein the rule check result is a binary discrimination value; And calculating the rule matching degree of the alternative resume text according to the weight coefficient of each target post rule and the rule checking result.
  8. 8. The candidate recommendation model construction method of claim 1 wherein said resume feature vectors comprise basic resume feature vectors, resume business feature vectors and resume implicit feature vectors; the candidate recommendation model construction method further comprises the following steps: calculating a difference between the rule matching degree and the matching confidence degree; screening rule matching degree and matching confidence degree of which the difference value exceeds a preset threshold value, and taking the rule matching degree and the matching confidence degree as the rule matching degree to be optimized and the matching confidence degree to be optimized; positioning the corresponding resume feature vector according to the rule matching degree to be optimized and the matching confidence degree to be optimized, and taking the resume feature vector as the resume feature vector to be optimized; and adjusting the resume business feature vector and the resume recessive feature vector in the resume feature vector to be optimized to obtain an adjusted resume feature vector.
  9. 9. A candidate recommendation model construction apparatus, comprising: the data acquisition module is used for acquiring historical business transaction data of a target post, post feature vectors and target post rules, and resume text, resume feature vectors and intention data of each candidate; The first screening module is used for calculating and obtaining the matching confidence coefficient of each resume text according to each resume feature vector and the post feature vector, and taking the resume text with the matching confidence coefficient meeting a first screening condition as an alternative resume text, wherein the first screening condition is that the matching confidence coefficient exceeds a preset confidence coefficient threshold value; the matching degree calculation module is used for carrying out rule check on each alternative resume text according to the target post rule, generating a rule check result and calculating to obtain rule matching degree based on the rule check result; The second screening module is used for screening candidate resume texts with the matching confidence degree meeting the first screening condition and the rule matching degree meeting the second screening condition as candidate resume texts, wherein the second screening condition is that the rule matching degree exceeds a preset matching degree threshold value; the model training module is used for training a preset reinforcement learning model according to the intention data, the historical business transaction data, the post feature vector, the resume feature vector corresponding to each candidate resume text and the rule checking result, and taking the reinforcement learning model after training as a candidate recommendation model.
  10. 10. A terminal device, comprising: one or more processors; A memory coupled to the processor for storing one or more programs; The one or more programs, when executed by the one or more processors, cause the one or more processors to implement the candidate recommendation model construction method of any one of claims 1-8.

Description

Candidate recommendation model construction method and device and terminal equipment Technical Field The invention relates to the field of artificial intelligence, in particular to a candidate recommendation model construction method, device and terminal equipment. Background With the increasing demand of enterprises for medium-high-end talents, hunter services have become an important component of the human resource market. According to industry data, the traditional hunter service finishes the recommendation of 50-100 resume required to be screened on average for middle-high-end talents once, the whole recruitment period is as long as 30-60 days, the core reasons of low efficiency are that the resume, post description and other core data are mainly unstructured texts, the unstructured texts cannot be processed in a large scale manually, and the efficiency of extracting key information from the unstructured texts manually is too low, so that the process of screening candidates is time-consuming and talent coverage is incomplete, and the requirements of the middle-high-end talents in the quick recruitment of enterprises are difficult to meet. Disclosure of Invention The invention provides a candidate recommendation model construction method, a candidate recommendation model construction device and terminal equipment, and the method can solve the problems of long time consumption and incomplete talent coverage caused by the fact that a traditional hunter service relies on manual processing of unstructured texts in the prior art. In order to solve the above technical problems, an embodiment of the present invention provides a candidate recommendation model construction method, including: acquiring historical business transaction data of a target post, post feature vectors and target post rules, and resume text, resume feature vectors and intention data of each candidate; Calculating to obtain the matching confidence coefficient of each resume text according to each resume feature vector and the post feature vector, and taking the resume text with the matching confidence coefficient meeting a first screening condition as an alternative resume text, wherein the first screening condition is that the matching confidence coefficient exceeds a preset confidence coefficient threshold value; For each alternative resume text, carrying out rule check on the alternative resume text according to the target post rule to generate a rule check result, and calculating to obtain rule matching degree based on the rule check result; screening out candidate resume texts with the matching confidence degree meeting the first screening condition and the rule matching degree meeting the second screening condition as candidate resume texts, wherein the second screening condition is that the rule matching degree exceeds a preset matching degree threshold value; Training a preset reinforcement learning model according to the intention data, the historical business transaction data, the post feature vector, the resume feature vector corresponding to each candidate resume text and the rule verification result, and taking the reinforcement learning model after training as a candidate recommendation model. Further, training a preset reinforcement learning model according to the intention data, the historical business transaction data, the post feature vector, the resume feature vector corresponding to each candidate resume text and the rule check result, and taking the reinforcement learning model after training as a candidate recommendation model, wherein the training comprises the following steps: For each candidate resume text, the post feature vector, the resume feature vector corresponding to the candidate resume text and the post history recommendation success rate are combined to form a fusion feature vector; Repeatedly executing training operation on a preset reinforcement learning model according to the fusion feature vector, the historical business transaction data, the rule check result and the intention data until reaching the preset iteration times, and obtaining a reinforcement learning model after training; taking the reinforced learning model after training as the candidate recommendation model; wherein the training operation comprises: Updating the real-time iteration times; inputting the fusion feature vector into a current value network to obtain the state value output by the current value network, wherein the current value network is an initial value network after the reinforcement learning model is initialized when training operation is performed for the first time; for the candidate corresponding to each candidate resume text, calculating to obtain the total rewards of the current candidate according to the historical business transaction data, the rule check result and the intention data of the current candidate; Inputting the fusion feature vector into a current strategy network to obtain the action probability outp