A post recommendation method and device

Abstract

The application provides a post recommendation method and device, and a computer readable storage medium. The post recommendation method comprises the following steps: a plurality of questions are asked to a candidate person for a plurality of competencies of the candidate person; answer texts of the candidate person answering each question are collected, and each answer text and a corresponding question text thereof form a question and answer pair; the question text and the answer text of each question and answer pair are segmented to obtain a plurality of question sentences and a plurality of answer sentences of each question and answer pair; a pre-trained reasoning model is used to perform information interaction, semantic reasoning and result classification on the plurality of question sentences and the plurality of answer sentences of each question and answer pair to obtain a scoring result of the candidate person for each competency; and a recommended post suitable for the candidate person is determined according to the scoring result of the candidate person for each competency and at least one competency combination required by a recruitment post.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of artificial intelligence (AI), and particularly relates to a post recommendation method based on artificial intelligence, an apparatus for executing the post recommendation method, and a computer readable storage medium. BACKGROUND

[0002] With the development of artificial intelligence technology, automated post recommendation technology or system gradually appears. Some existing post recommendation technology uses simple natural language processing to perform word segmentation or keyword extraction on the resume content of a candidate, the content of a current post, and the content of a target post, and performs matching or similarity calculation between the resume and the post based on some keyword weights, thereby achieving the purpose of simple post recommendation. Some other relatively advanced post recommendation technology trains a word vector model based on external encyclopedic data to represent a personal semantic vector of a candidate (including the resume content and personal information of the candidate) and a post semantic vector, and further obtains the suitability of the candidate for the current post by calculating the semantic similarity. Some higher level post recommendation technology trains a matching model (a simple neural network model such as a convolutional neural network) based on the resume content and the post content on the basis of the word vector, to predict the suitability of the candidate for the post. Although these methods can achieve automatic post recommendation to some extent, they still have the following shortcomings:

[0003] (1) Weak theoretical support. Post recommendation based only on a resume and post requirements can only indicate that the work experience of a candidate meets the requirements of a post to some extent, and it is difficult to ensure that the recommended candidate can perform the current work. In other words, the recommended candidate still needs to go through tedious manual interviews or video interviews for further assessment, and does not bring substantial liberation to the interviewers of an enterprise.

[0004] (2) Rough technology. Neither word segmentation, keyword extraction, nor similarity calculation of semantic vectors or matching models based on basic word vector technology can deeply mine the internal correlation between the work experience of a candidate and the requirements of a post. Further, such methods can hardly ensure that the recommended candidate is suitable for the corresponding post.

[0005] (3) Lack of assessment of the candidate's competency. According to a kind of interview auxiliary technology disclosed by Shanghai Jiaoliang Intelligent Technology Co., Ltd. (Patent No. / Application No. 202011060966.6), competency refers to the comprehensive quality of an individual, which is a non-technical ability that can be sustained throughout a person's life, mainly including work responsibility, ability to communicate and coordinate with colleagues, time management ability, stress resistance and professional enthusiasm for the post engaged in. Competency is an important factor to assess whether a candidate is competent for a recruitment post. Without competency assessment of the candidate, the post recommendation technology cannot comprehensively and accurately assess whether the candidate is competent for the recruitment post.

[0006] Based on the above pain points, the present application provides a post recommendation technology, which can on the one hand use artificial intelligence technology to do question and answer interaction on the question text and answer text of the candidate's answers to various questions from the perspective of competency assessment, thereby improving the evaluation accuracy of artificial intelligence on the candidate's various competencies, and on the other hand automatically recommend suitable posts for the candidate according to the evaluation results of the candidate's various competencies. SUMMARY

[0007] The following gives a brief summary of one or more aspects to provide a basic understanding of these aspects. This summary is not an exhaustive overview of all contemplated aspects, and neither is it intended to identify key or critical elements of all aspects nor to delineate the scope of any or all aspects. Its only purpose is to present some concepts of one or more aspects in a simplified form before the more detailed description is given later.

[0008] In order to overcome the above-mentioned defects existing in the prior art, the present application provides a post recommendation method, a post recommendation device and a computer readable storage medium, which are used to use artificial intelligence technology to do question and answer interaction on the question text and answer text of the candidate's answers to various questions from the perspective of competency assessment, thereby improving the evaluation accuracy of artificial intelligence on the candidate's various competencies, and automatically recommending suitable posts for the candidate according to the evaluation results of the candidate's various competencies.

[0009] Specifically, the first aspect of the present application provides the above-mentioned job recommendation method, which comprises the following steps: asking a plurality of questions to a candidate person with respect to a plurality of competencies of the candidate person; collecting answer texts of the candidate person answering each of the questions, and forming a question-answer pair by combining each of the answer texts with a question text corresponding thereto; segmenting the question text and the answer text of each of the question-answer pairs to obtain a plurality of question sentences and a plurality of answer sentences of each of the question-answer pairs; performing information interaction, semantic reasoning and result classification on the plurality of question sentences and the plurality of answer sentences of each of the question-answer pairs by using a pre-trained reasoning model to obtain a scoring result of the candidate person with respect to each of the competencies; and determining a recommended job suitable for the candidate person according to the scoring result of the candidate person with respect to each of the competencies and a competency combination required by at least one recruitment job. By performing these steps, the job recommendation method can on the one hand utilize artificial intelligence technology to perform question-answer interaction on the question text and the answer text of the candidate person answering each question from the perspective of competency assessment, thereby improving the evaluation accuracy of the artificial intelligence on each of the competencies of the candidate person, and on the other hand automatically recommend a suitable job for the candidate person according to the evaluation result of each of the competencies of the candidate person.

[0010] Preferably, in some embodiments of the present application, the step of asking a plurality of questions to a candidate person with respect to a plurality of competencies of the candidate person comprises: determining the plurality of competencies according to competency models of a plurality of jobs; retrieving one or more questions corresponding to each of the competencies from a pre-constructed competency question bank; and playing a video or audio of each of the questions to the candidate person.

[0011] Preferably, in some embodiments of the present application, before performing the step of asking a plurality of questions to a candidate person with respect to a plurality of competencies of the candidate person, the job recommendation method further comprises the following steps: interviewing a plurality of employees of a plurality of jobs to obtain interview records of the plurality of employees; scoring the interview records of each of the employees according to pre-defined competency dimensions to obtain scores of each of the employees in each of the competencies; dividing the plurality of employees into excellent employees and ordinary employees of each of the jobs according to work performance; performing difference test on the scores of the excellent employees and the ordinary employees in each of the competencies to determine a plurality of competencies that the excellent employees of each of the jobs should possess; constructing a competency model of each of the jobs according to the plurality of competencies that the excellent employees of each of the jobs should possess; and preparing at least one question for each of the competencies and constructing the competency question bank based on the plurality of competencies that the excellent employees of the plurality of jobs should possess.

[0012] Optionally, in some embodiments of the present application, the step of collecting the answer texts of the candidate's answers to each of the questions comprises: collecting a video or an audio of the candidate's answers to each of the questions; and performing speech-to-text transcription on audio data in the video or the audio to obtain the answer texts of the candidate's answers to each of the questions.

[0013] Optionally, in some embodiments of the present application, the step of segmenting the question text and the answer text of each of the question-answer pairs comprises: segmenting the question text and the answer text of each of the question-answer pairs in units of sentences; or segmenting the question text and the answer text of each of the question-answer pairs in units of RDF triples.

[0014] Optionally, in some embodiments of the present application, the inference model comprises a pre-trained encoding module and a pre-trained relation graph network module. The step of performing information interaction, semantic reasoning and result classification on the multiple question sentences and the multiple answer sentences of each of the question-answer pairs comprises: performing vector representation on each of the question sentences and each of the answer sentences in each of the question-answer pairs by using the encoding module to generate an encoding representation vector of each of the question sentences and an encoding representation vector of each of the answer sentences in each of the question-answer pairs, respectively; performing iterative reasoning on the encoding representation vector of each of the question sentences and the encoding representation vector of each of the answer sentences in each of the question-answer pairs by using the relation graph network module to realize information interaction between each of the question sentences and each of the answer sentences in each of the question-answer pairs, and to generate a logical relation representation vector of each of the question sentences and a logical relation representation vector of each of the answer sentences in each of the question-answer pairs, respectively; performing semantic reasoning on the logical relation representation vector of each of the question sentences and the logical relation representation vector of each of the answer sentences in each of the question-answer pairs to generate a semantic representation vector of each of the question-answer pairs; and performing result classification according to the semantic representation vector of each of the question-answer pairs to obtain a score result of the candidate for each of the competencies, respectively.

[0015] Preferably, in some embodiments of the present application, the step of iteratively reasoning on the encoded representation vectors of the question sentences and the encoded representation vectors of the answer sentences in each of the question-answer pairs comprises: weighting and integrating the encoded representation vector of each of the question sentences in each of the question-answer pairs according to the encoded representation vector of each of the question sentences, the encoded representation vectors of the rest of the question sentences, and the encoded representation vectors of the answer sentences, to realize information interaction between each of the question sentences and the rest of the question sentences and the answer sentences in each of the question-answer pairs, and to generate a logical relationship representation vector of each of the question sentences, respectively; and weighting and integrating the encoded representation vector of each of the answer sentences in each of the question-answer pairs according to the encoded representation vector of each of the answer sentences, the encoded representation vectors of the rest of the answer sentences, and the encoded representation vectors of the question sentences, to realize information interaction between each of the answer sentences and the rest of the answer sentences and the question sentences in each of the question-answer pairs, and to generate a logical relationship representation vector of each of the answer sentences, respectively.

[0016] Preferably, in some embodiments of the present application, the step of iteratively reasoning on the encoded representation vectors of the question sentences and the encoded representation vectors of the answer sentences in each of the question-answer pairs further comprises: determining the neighbor nodes of each question sentence node and each answer sentence node in each of the question-answer pairs according to the order of the question text and the answer text in each of the question-answer pairs, wherein the neighbor nodes of the question sentence node include all question sentence nodes in the question-answer pair that are spaced less than a preset window value from the question sentence node, and all answer sentence nodes in the question-answer pair, and the neighbor nodes of the answer sentence node include all answer sentence nodes in the question-answer pair that are spaced less than the preset window value from the answer sentence, and all question sentence nodes in the question-answer pair; weighting and integrating the encoded representation vector of each of the question sentence nodes according to the encoded representation vectors of each of the question sentence nodes and all of its neighbor nodes, to realize information interaction between each of the question sentence nodes and all of its neighbor nodes in each of the question-answer pairs, and to generate a logical relationship representation vector of each of the question sentence nodes, respectively; and weighting and integrating the encoded representation vector of each of the answer sentence nodes according to the encoded representation vectors of each of the answer sentence nodes and all of its neighbor nodes, to realize information interaction between each of the answer sentence nodes and all of its neighbor nodes in each of the question-answer pairs, and to generate a logical relationship representation vector of each of the answer sentence nodes, respectively.

[0017] Optionally, in some embodiments of the present application, the inference model further comprises a pre-trained semantic inference module. The step of performing semantic inference on the logical relation representation vector of each question and the logical relation representation vector of each answer in each question-answer pair to generate a semantic representation vector of each question-answer pair comprises: performing semantic inference on the logical relation representation vector of each question and the logical relation representation vector of each answer in each question-answer pair by using the semantic inference module to realize semantic interaction between each question and each answer in each question-answer pair, and generating a semantic representation vector of each question-answer pair respectively.

[0018] Preferably, in some embodiments of the present application, the inference model further comprises a plurality of pre-trained classification network modules. The step of performing result classification according to the semantic representation vector of each question-answer pair to obtain a score result of the candidate for each competency comprises: inputting the semantic representation vector of each question-answer pair into a corresponding classification network module to obtain a score result of the candidate for each competency respectively.

[0019] Optionally, in some embodiments of the present application, the step of determining a recommended position suitable for the candidate according to the score result of the candidate for each competency and a competency combination required by at least one recruitment position comprises: screening the score result of the candidate for each competency according to a preset score threshold to determine at least one qualified competency of the candidate; comparing the at least one qualified competency with the competency combination required by at least one recruitment position to determine that the at least one qualified competency can comprehensively cover at least one recruitment position of the competency combination; and determining the at least one recruitment position that the at least one qualified competency can comprehensively cover as a recommended position of the candidate.

[0020] Preferably, in some embodiments of the present application, after performing the step of determining a recommended position suitable for the candidate according to the score result of the candidate for each competency and a competency combination required by at least one recruitment position, the position recommendation method further comprises the following steps: screening a recommended position list of a plurality of candidates according to a recruitment position provided by a recruitment enterprise to determine a recommended candidate of the recruitment position, wherein the recommended position list records at least one recommended position suitable for a corresponding candidate; and recommending the recommended candidate to the recruitment enterprise.

[0021] Preferably, in some embodiments of the present application, after the step of recommending the recommended candidate to the recruitment enterprise, the job recommendation method further comprises the step of providing the recruitment enterprise with the answer text of the recommended candidate answering each of the questions and the corresponding question text thereof, and / or the video and / or audio of the recommended candidate answering each of the questions, and / or the score result of the recommended candidate for each of the competencies.

[0022] Optionally, in some embodiments of the present application, before the step of putting multiple questions to the candidate for multiple competencies, the job recommendation method further comprises the steps of: putting multiple corresponding questions to multiple candidate samples respectively according to the multiple competencies; collecting answer text samples of each of the candidate samples answering each of the questions respectively, and forming question and answer pair samples respectively by combining each of the answer text samples with the corresponding question text thereof; segmenting the question text and the answer text sample of each of the question and answer pair samples to obtain multiple question sentence samples and multiple answer sentence samples of each of the question and answer pair samples; performing competency scoring on the answer text samples of each of the candidate samples answering each of the questions respectively for each of the competencies of each of the candidate samples; and training the reasoning model to perform the information interaction, the semantic reasoning and / or the result classification by using the multiple question sentence samples and the multiple answer sentence samples of each of the question and answer pair samples, and the competency scores of the corresponding candidate samples answering each of the questions.

[0023] According to the second aspect of the present application, the above-mentioned job recommendation device comprises a memory and a processor. The processor is connected to the memory and is configured to implement the above-mentioned job recommendation method provided by the first aspect of the present application. By implementing the job recommendation method, the job recommendation device can on the one hand use artificial intelligence technology to perform question and answer interaction on the question text and the answer text of the candidate answering each question from the perspective of competency assessment, thereby improving the evaluation accuracy of the artificial intelligence on the competencies of the candidate, and on the other hand automatically recommend suitable positions for the candidate according to the evaluation results of the competencies of the candidate.

[0024] According to the third aspect of the present application, the above-mentioned computer readable storage medium has computer instructions stored thereon. The computer instructions are executed by a processor to implement the above-mentioned job recommendation method provided by the first aspect of the present application. By implementing the job recommendation method, the computer readable storage medium can on the one hand use artificial intelligence technology to perform question and answer interaction on the question text and the answer text of the candidate answering each question from the perspective of competency assessment, thereby improving the evaluation accuracy of the artificial intelligence on the competencies of the candidate, and on the other hand automatically recommend suitable positions for the candidate according to the evaluation results of the competencies of the candidate. BRIEF DESCRIPTION OF DRAWINGS

[0025] The above features and advantages of the present application will be better understood by reading the following detailed description of the embodiments of the present application in conjunction with the drawings, in which:

[0026] Figure 1 A flowchart of constructing a competency model and a competency question bank is shown according to some embodiments of the present application.

[0027] Figure 2 A flowchart of training an inference model is shown according to some embodiments of the present application.

[0028] Figure 3 A schematic diagram of a relationship graph network is shown according to some embodiments of the present application.

[0029] Figure 4 A flowchart of a post recommendation method is shown according to some embodiments of the present application. DETAILED DESCRIPTION

[0030] The present application is described in detail by specific embodiments, and those skilled in the art can easily understand other advantages and effects of the present application from the disclosure. Although the description of the present application will be introduced in conjunction with the preferred embodiments, it does not mean that the features of the present application are limited to the embodiments. On the contrary, the purpose of introducing the embodiments in conjunction with the present application is to cover other options or modifications that can be extended based on the claims of the present application. In order to provide a deep understanding of the present application, many specific details will be included in the following description. The present application can also be implemented without using these details. In addition, in order to avoid confusion or obscure the focus of the present application, some specific details will be omitted in the description.

[0031] As described above, in order to overcome the defects of weak theoretical support, rough technology and lack of assessment of candidate's competence in the prior art, the present application provides a post recommendation method, a post recommendation device, and a computer readable storage medium, which are used to use artificial intelligence technology to do question and answer interaction on the question text and answer text of each question answered by the candidate from the perspective of competence assessment, so as to improve the evaluation accuracy of artificial intelligence on the competence of the candidate, and automatically recommend suitable posts for the candidate according to the evaluation results of the competence of the candidate.

[0032] In some non-limiting embodiments, the first aspect of the present application provides that the above-mentioned job recommendation method can be implemented by the above-mentioned job recommendation device provided by the second aspect of the present application, using a pre-constructed and trained competency model and an inference model. Specifically, the above-mentioned job recommendation device provided by the second aspect of the present application can include a memory and a processor. The memory includes but is not limited to the above-mentioned computer readable storage medium provided by the third aspect of the present application, on which computer instructions are stored. The processor is connected to the memory and is configured to execute the computer instructions stored on the memory to implement the above-mentioned job recommendation method provided by the first aspect of the present application.

[0033] The working principle of the above-mentioned job recommendation device will be introduced below in combination with some methods of constructing a competency model, some methods of training an inference model, and some methods of using the competency model and the inference model for job recommendation. Those skilled in the art can understand that these methods of constructing a competency model, training an inference model, and job recommendation are only some non-limiting embodiments provided by the present application, which are intended to clearly demonstrate the main idea of the present application and provide some specific cases for the public to implement, rather than to limit the protection scope of the present application.

[0034] Please refer to Figure 1 , Figure 1 A flowchart for constructing a competency model and a competency question bank according to some embodiments of the present application is shown.

[0035] As Figure 1 shown, in some embodiments of the present application, in the construction phase of the competency model and the competency question bank, the constructor can first determine a plurality of positions and interview a plurality of employees of each position to obtain interview records of the plurality of employees. Then, the constructor can score the interview records of each employee according to a plurality of pre-defined competency dimensions such as professional knowledge, skill level, responsibility, stress resistance, time management ability, communication ability, adaptability, and enthusiasm for the position, to obtain the score of each employee in each competency for each position.

[0036] In some embodiments, the work of scoring the interview records of each employee according to a plurality of competencies required by the position can be implemented by a pre-trained competency evaluation model. The competency evaluation model is an artificial intelligence model that preliminarily determines the score of each employee in each competency based on a neural network model (such as CNN, RNN or its model variants) and a classification network. Its basic principle can refer to the prior art published by Shanghai Jinuyu Intelligent Technology Co., Ltd. (Patent No. / Application No. 202011060966.6).

[0037] Alternatively, in other embodiments, the task of scoring each employee's interview records for multiple competencies required for the position can also be obtained by human resources experts manually annotating the interview content for each competency dimension.

[0038] After obtaining the scores for each employee in each position for each competency, the builders of the competency model and competency question bank can organize the interview content of each employee to obtain the competency question set Q[q1, q2, ..., q] for each employee. k And the corresponding competency scores S[s1, s2, ..., s] k Afterwards, the builder can categorize multiple employees in each position into excellent and average employees based on their actual work performance in their respective roles, and combine this with competency assessment models or competency levels L[l] provided by human resources experts. q1C1 , l q2C1 , ..., l qkC1 , l q1C2 , l q2C2 , ..., l qkC2 ] Calculate the total scores of excellent employee C1 and average employee C2 across multiple competency dimensions:

[0039] total_score_C1=s1*l q1C1 +s2*l q2C1 +…+s k *l qkC1

[0040] total_score_C2=s1*l q1C2 +s2*l q2C2 +…+s k *l qkC2

[0041] In the formula, l qiCj Let C represent the j-th employee. j In the i-th competency problem q i The competency level.

[0042] Furthermore, the builder can combine the aforementioned competency scores S and competency levels L to calculate the mean total score, mean competency level, and standard deviation for each competency dimension for the excellent employee group C1 and the ordinary employee group C2. Then, the builder can determine the k competencies needed to differentiate between excellent and ordinary employees for a given position by performing difference tests on the mean total score, mean competency level, and standard deviation for each competency dimension between the excellent employee group C1 and the ordinary employee group C2.

[0043] Specifically, if the total score average of the excellent employee group C1 is significantly higher than the total score average of the ordinary employee group C2 in a certain competency dimension, the constructor of the competency model and the competency question bank can determine that this competency can significantly distinguish the excellent employees from the ordinary employees of the post. According to this scheme, the constructor can determine whether each competency can significantly distinguish the excellent employees from the ordinary employees of the post one by one, thereby determining the multiple competencies that the excellent employees of a post should possess, and further determining the multiple competencies that the excellent employees of each post should possess.

[0044] After determining the multiple competencies that the excellent employees of each post should possess, the constructor can construct a competency model for each post based on the multiple competencies that the excellent employees of each post should possess. Meanwhile, the constructor can also prepare at least one question for each competency and construct a competency question bank based on the multiple competencies that the excellent employees of each post should possess, so that the competency model can retrieve questions for each competency to comprehensively assess the competencies of the candidate.

[0045] In some preferred embodiments, the competency combination that distinguishes the excellent employees from the ordinary employees of a post is not unique. For example, for an excellent office clerk, he needs to have a strong sense of responsibility, good communication and coordination skills, and time management skills to ensure timely completion of work. On the other hand, another office clerk who has a strong sense of responsibility, hardworking spirit, and good communication and coordination skills, but lacks good time management skills, can also ensure timely completion of work and meet the standards of excellent employees of the post. At this time, the constructor of the competency model and the competency question bank can also configure multiple different competency combinations for the same post to construct the above-mentioned competency model and competency question bank, so as to comprehensively and individually attract excellent talents who meet the requirements of the recruitment post.

[0046] Those skilled in the art can understand that the constructor described above is a non-limiting description, including but not limited to technical personnel who perform the above-mentioned construction method, as well as processors and other related devices that perform the above-mentioned construction method.

[0047] After completing the construction of competency models and competency question banks for a large number (e.g., 1000 or more, 2000 or more, 5000 or more, 10000 or more) of posts, those skilled in the art can retrieve one or more questions corresponding to each competency from the pre-constructed competency question bank according to all competencies involved in the competency models of these posts, and play a video or audio of each question to the candidate sample for subsequent training of the inference model and use of the inference model for post recommendation.

[0048] Please refer to Figure 2 , Figure 2A flowchart of training an inference model is shown according to some embodiments of the present application.

[0049] As shown in Figure 2 , in some embodiments of the present application, in the training phase of the inference model, the trainer can first use the competency model constructed in the above embodiments to obtain questions for assessing each competency from the competency question bank constructed in the above embodiments according to a plurality of competencies required by a large number of positions, to construct a question set Q [q1, q2, …, q k ]. Then, the trainer can recruit a large number of candidate samples to answer these questions, and collect answer text samples of each candidate sample answering these questions, to construct an answer set A [a1, a2, …, a k ]. After that, the trainer can hire experts in the field of human resources to score the ability of each candidate sample in each competency dimension T [T1, T2, …, T k ] based on the above question set Q and the above answer set A, so as to comprehensively obtain the data of a large number of candidate samples in each competency as the basis for subsequent training of the inference model. In some embodiments, the score T of each candidate sample in each competency dimension can be in the form of a multi-level score system such as a binary system, a ternary system, a quinary system, etc., and the higher the score, the better the ability of the candidate sample in the corresponding competency dimension.

[0050] After obtaining the question set Q, the answer set A, and the single score T of each candidate sample in each competency dimension of a plurality of candidate samples, the trainer can construct an inference model, and train the information interaction module, the semantic inference module, and / or the result classification module of the inference model based on these sample data, so that the trained inference model has the function of accurately evaluating the competencies of candidates.

[0051] Specifically, the trainer of the inference model can first group each answer text sample a i and its corresponding question text q i to form a question-answer pair sample <q i , a i > for each candidate sample. After that, for each question-answer pair sample <q i , a i >, the trainer cuts the question text q i into a plurality of question samples [q s1 , q s2 , …, q sc ] in units of sentences, and cuts the answer text sample a i into a plurality of answer samples [a s1 , a s2 , …, a sp], where c represents the question text q i The number of questions in the text, p represents the number of response text samples a. i The number of answers in the text.

[0052] Afterwards, the trainer can use encoding modules composed of sequence models such as CNN, RNN, LSTM, GRU, and BERT to process each question sample [q]. s1 q s2 , ..., q sc Encode the data to obtain the encoded representation vector [r] for each question sample. qs1 r qs2 ,...,r qsc ], and for each answer sample [a s1 a s2 , ..., a sp Encode the responses to obtain the encoded representation vector [r] for each response sample. as1 r as2 ,...,r asp For specific schemes for encoding and representing text data using sequence models, please refer to the prior art with patent number / application number 202011060966.6, which will not be elaborated here.

[0053] After that, the trainer can build a relational graph network module based on graph models (including but not limited to GATConv, RelGraphConv, GCN, and TAGConv), and use this relational graph network module to train question-answer pair samples. i a i Sample questions [q] s1 q s2 , ..., q sc ] and sample answers [a s1 a s2 , ..., a sp Information exchange is conducted to characterize the deep or shallow dependency information and semantic relationships between questions, between questions and answers, and between answers, thereby improving the accuracy of artificial intelligence in assessing candidates' various competencies.

[0054] Taking the RelGraphConv relational graph network as an example, the trainer can train samples based on question-and-answer pairs. i a i The encoded representation vectors of each question and the remaining questions in the > structure (e.g., r qs1 With r qs2 ~r qsc ), the encoded representation vectors of each question and each answer (e.g., r qs1 With r as1 ​​~r asp ), and the encoded representation vectors of each answer and the remaining answers (e.g., r as1 With r as2 ~r qsp Construct a relational graph network and rely on the information transmission between nodes in the relational graph network to realize information interaction between questions, between questions and answers, and between answers.

[0055] Please refer to Figure 3 , Figure 3 A schematic diagram of a relational graph network provided according to some embodiments of the present invention is shown.

[0056] exist Figure 3 In the illustrated embodiment, question-answer pairs are used as samples. i a i The graph network can include two questions q1-q2 and three answers a1-a3. Correspondingly, the graph network can include five nodes: q1-q2 and a1-a3. When using the graph network for iterative reasoning, the representation vector h of each node... i They will all be based on their own h i and its associated neighboring nodes h j Perform weighted integration, that is Among them, w i w n w j All are weights to be learned, and nc represents node h. i The number of neighboring nodes. In this way, the relational graph network can rely on the information transmission between nodes to realize the information interaction between questions, between questions and answers, and between answers, and represent the semantic associations between questions, between questions and answers, and between answers.

[0057] Compared to existing technologies that use traditional sequence models, this invention uses a graph model to analyze question-answer pair samples. i a i Graph network modeling can effectively shorten the distance between questions and answers, thus avoiding the gradient vanishing problem that occurs in sequence modeling due to the large number of sentences, and ensuring that samples of the same question-answer pair are consistent. i a i The deep interaction of causal relationships between the questions and answers in the text.

[0058] Based on the above description, relational graph networks can analyze question-answer pair samples. i a i The encoded representation vector of each question sample in the > qs1 ,r qs2 ,…,r qsc ​​​​], and the encoded representation vector [r] of each answer sample. as1 ,r as2 ,…,r asp Perform iterative reasoning to obtain question-answer pair samples. i a i Sample questions q si Logical relation representation vector [l qs1 ,l qs2 ,…,l qsc ] and sample answers a si Logical relation representation vector [l as1 ,l as2 ,…,l asp These logical relationships represent vectors [l] qs1 ,l qs2 ,…,l qsc ] and [l as1 ,l as2 ,…,l asp [Able to characterize question-answer pair samples] i a i The semantic relationships between the questions, between the questions and the answers, and between the answers.

[0059] Those skilled in the art will understand that Figure 3 The diagram network shown, which includes only 5 nodes, is merely an embodiment of a simplified structure provided by this invention. It is intended to clearly demonstrate the main concept of this invention and to provide a specific solution that is easy for the public to implement, rather than to limit the scope of protection of this invention.

[0060] Alternatively, in other embodiments, for complex relational graph networks that include a large number of question nodes and a large number of answer nodes, the trainer can train the network based on samples of each question-answer pair. i a i >Chinese question text q i and answer text sample a i The word order and preset window value are used to determine the word order of each question-answer pair. i a i >Each question node q si and each answer node a si The neighboring nodes. For example, for an embodiment with a preset window value of 2, the trainer can use question-answer pair samples. i a i >Middle and question node q si All question nodes q with an interval less than 2 s(i-1) and q s(i+1) and question-answer pairs samples i a i All answer nodes a s1 ​​​​​​~a sp All were identified as question nodes q si The neighboring nodes. Similarly, the trainer can also use question-answer pair samples. i a i >Middle and answer node a si All answer nodes a with an interval less than 2 s(i-1) and a s(i+1) and question-answer pairs samples i a i All question nodes q in > s1 ~q sc All are determined to be answer node a si The neighboring nodes.

[0061] Afterwards, the trainer can adjust the samples based on the question-and-answer format. i a i > Each question node q si and all its neighboring nodes, as well as question-answer pair samples i a i > Each answer node a si Construct question-answer pairs samples from the node and all its neighboring nodes. i a i >Chinese question text q i and answer text sample a i A relational graph network. This relational graph network is capable of analyzing question-answer pair samples. i a i The network iteratively infers from each node to achieve information exchange and semantic association between neighboring nodes. In this way, the relational graph network can analyze samples based on question-and-answer relationships. i a i Each question node q in > si The encoded representation vectors of the node q and all its neighboring nodes are used to represent each question node q. si The encoded representation vectors are weighted and integrated to achieve question-answering pairs of samples. i a i Each question node q in > si It exchanges information with all its neighboring nodes and generates each question node q. si Logical relation representation vector [l qs1 ,l qs2 ,…,l qsc Similarly, this relational graph network can also analyze samples based on question-and-answer pairs. i a i Each answer node a si The encoded representation vectors of the node and all its neighboring nodes are used to represent each answer node a. si ​​​​​​​​​The encoded representation vectors are weighted and integrated to achieve question-answer pairing of samples. i a i Each answer node a si It exchanges information with all its neighboring nodes and generates a response node a for each node. si Logical relation representation vector [l as1 ,l as2 ,…,l asp ].

[0062] By configuring the aforementioned window value setting interface, trainers can set appropriate window values ​​based on the accuracy requirements of the inference model's assessment of candidates' various competencies, as well as the data processing capabilities of the training equipment and job recommendation device's processors, to balance both assessment accuracy and processing speed. The specific value of the window does not affect the basic requirements of those skilled in the art for constructing relational graph networks, and will not be elaborated upon here.

[0063] Those skilled in the art will understand that the above-described text segmentation scheme based on sentences is merely a non-limiting embodiment of the present invention, intended to clearly demonstrate the main concept of the invention and provide a specific solution easy for the public to implement, rather than being intended to limit the scope of protection of the present invention. Optionally, in other embodiments, the trainer can also segment question-answer pair samples based on the subject-verb-object (RDF) triple structure. i a i >The question text q i and answer text sample a i To obtain question-answer pairs in the same way i a i Multiple question samples [q] s1 ,q s2 ,…,q sc ] and multiple sample answers [a s1 ,a s2 ,…,a sp Afterwards, the trainer can train each question sample q as described above. si and sample answers a si Encoding is performed, and a relational graph network is constructed to facilitate information exchange between neighboring nodes, thereby generating sample question-answer pairs. i a i Sample questions q si Logical relation representation vector [l qs1 ,l qs2 ,…,l qsc ] and sample answers a si Logical relation representation vector [l as1 ,l as2 ,…,l​​​​asp The specific process of encoding representation and information interaction in this scheme is similar to that in the above embodiments, and will not be repeated here.

[0064] After constructing the relational graph network module of the inference model, the trainer can continue to build a semantic inference module based on graph models such as Graph Attention Network (GAT), and use this semantic inference module to process question-answer pair samples. i a i Sample questions q si Logical relation representation vector [l qs1 ,l qs2 ,…,l qsc ], and sample answers a si Logical relation representation vector [l as1 ,l as2 ,…,l asp Perform semantic reasoning to further enhance the individual question-answer pair with the sample. i a i The semantic interactions between the questions, between the questions and the answers, and between the answers themselves further enhance the accuracy of artificial intelligence in assessing the candidates' various competencies.

[0065] Specifically, targeting Figure 3 The relational graph network module shown can also have the same two question nodes q1-q2 and three answer nodes a1-a3 in its backend semantic reasoning module. The difference between the semantic reasoning module and the relational graph network module lies in the fact that in the semantic reasoning module, each question node q1-q2... si The representation vector h i It is no longer the encoded representation vector r mentioned above. qsi Instead, it is the logical relation representation vector l obtained through inference and iteration of the relation graph network module. qsi Correspondingly, each answer node a in the semantic reasoning module si The representation vector h i It is no longer the encoded representation vector r mentioned above. asi Instead, it is the logical relation representation vector l obtained through inference and iteration of the relation graph network module. asi .

[0066] When performing semantic reasoning using the semantic reasoning module, the representation vector h of each node in the graph attention network... i They will all be based on their own h i and its associated neighboring nodes h j Perform weighted integration, that is Among them, w i w n ​​, w j are to be learned weights, and nc represents the number of neighbor nodes of the node h i . In this way, the graph attention network can rely on information transmission between nodes to perform semantic interaction between questions, between questions and answers, and between answers, so as to realize semantic reasoning between questions, between questions and answers, and between answers.

[0067] Compared with the prior art using a traditional sequence model, the present application uses a graph attention network model to model semantic reasoning of question and answer pair samples <q i , a i >, which can effectively shorten the distance between questions and answers, thereby avoiding the gradient vanishing problem caused by too many sentences in the sequence model modeling process, and further ensuring the deep interaction between the semantics of the questions and answers in the same question and answer pair sample <q i , a i >.

[0068] Based on the above description, the semantic reasoning module can further perform semantic reasoning on the logical relationship representation vectors [l i , l i ,..., l qs1 ] of the question samples and the logical relationship representation vectors [l qs2 , l qsc ,..., l as1 ] of the answer samples in the question and answer pair sample <q as2 , a asp > to obtain the semantic representation vectors [h1, h2,..., h i ] of the question and answer pair samples <q i , a c+p > respectively. The semantic representation vectors [h1, h2,..., h c+p ] can represent the semantic association between questions, between questions and answers, and between answers in the question and answer pair samples <q i , a i >.

[0069] In some embodiments, for the case that the number of answer sentences in the answer text samples of different candidate samples is different, the trainer can further perform max-pooling or mean-pooling operation on the obtained multiple semantic representation vectors [h1, h2,..., h c+p ] to obtain a semantic representation vector H[h1, h2,..., h c+p ] of uniform dimension.

[0070] After constructing the semantic reasoning module of the reasoning model, the trainer can further build multiple classification network modules on the backend of the semantic reasoning module. These classification network modules can respectively obtain question-answer pair samples for the corresponding competency dimension. i a i > semantic representation vector h i And based on the obtained semantic representation vector h i Determine the corresponding classification result. This classification result indicates to the candidate the rating result T for a corresponding competency. i The basic principle of using a classification network to determine the corresponding classification result based on the semantic representation vector can be found in the prior art with patent number / application number 202011060966.6, and will not be elaborated here.

[0071] Based on the above construction process, the trainer can build a reasoning model that includes an encoding module, a relational graph network module, a semantic reasoning module, and multiple classification network modules. In this scenario with multiple classification network modules, the entire reasoning model can be viewed as multiple end-to-end multi-task models based on single-question scoring. Question-answering pair samples i a i Interaction information can be transmitted from the graph nodes of the relational graph network module to the backend semantic reasoning module and various classification network modules for the optimization of the entire reasoning model.

[0072] After completing the construction process of the reasoning model, the trainer can input the question set Q, the answer set A, and the competency score results T of each candidate sample in each competency dimension into the constructed reasoning model in sequence to simultaneously train the learning parameters of the above-mentioned encoding module, relational graph network module, semantic reasoning module and / or multiple classification network modules, so that the trained reasoning model has the above-mentioned functions of encoding representation, information interaction, semantic reasoning and competency score result classification.

[0073] Those skilled in the art will understand that the training scheme for each module in the above-described synchronous training inference model is merely a non-limiting implementation provided by the present invention, intended to clearly demonstrate the main concept of the invention and provide a specific solution that is easy for the public to implement, rather than being used to limit the scope of protection of the present invention. Optionally, in some other embodiments, the trainer may also progressively annotate the intermediate parameters output by each module and train each module of the inference model separately based on each intermediate parameter to achieve the same training effect.

[0074] Those skilled in the art will also understand that the above-described trainer is only a non-limiting description and includes, but is not limited to, a technician performing the above-described training method, as well as a processor and other related devices performing the above-described training method.

[0075] ​​After training the aforementioned reasoning model, the user can use the job recommendation device provided in the second aspect of the present invention to implement the job recommendation method provided in the first aspect of the present invention by utilizing the pre-built competency model and competency question bank, as well as the trained reasoning model.

[0076] Please refer to Figure 4 , Figure 4 A flowchart illustrating a job recommendation method provided according to some embodiments of the present invention is shown.

[0077] like Figure 4 As shown, in some embodiments of the present invention, when recommending a candidate for a job, the job recommendation device can first query a large number of job competency models to determine the multiple competencies involved in these competency models. Then, the job recommendation device can retrieve one or more questions corresponding to each competency from a pre-built competency question bank and play a video or audio recording of each question to present these multiple questions to the candidate. By selecting questions based on the multiple competencies involved in a large number of job competency models, the present invention can comprehensively assess the candidate's competencies in various aspects, thereby accurately and comprehensively recommending jobs of various fields, industries, and types to the candidate.

[0078] Candidates can answer each question sequentially based on the video or audio played by the job recommendation device. Simultaneously, the job recommendation device can collect the video or audio of the candidate's answers to each question. Then, the job recommendation device can first extract the audio data of the candidate's answers from the collected video, and then use Automatic Speech Recognition (ASR) technology to transcribe the extracted audio data into speech text, obtaining the candidate's answer text A[a1,a2,…,a…]. k ].

[0079] Those skilled in the art will understand that the above-described solutions for presenting multiple questions to candidates via video or audio are merely non-limiting embodiments provided by this invention, intended to clearly demonstrate the main concepts of the invention and provide specific solutions convenient for public implementation, rather than limiting the scope of protection of this invention. Optionally, in other embodiments, the job recommendation device may also present these multiple questions to candidates by displaying the question text on a human-computer interaction interface such as a display screen or touch screen to achieve the same effect. Furthermore, the job recommendation device may also directly collect the candidate's answers to each question via a human-computer interaction interface such as a physical / virtual keyboard, thereby omitting the aforementioned ASR speech-to-text transcription step.

[0080] After obtaining the candidate's answer text A for each question, the job recommendation device can process each answer text a.i Each of them corresponds to the question text q i Forming a question-and-answer pair i a i Afterwards, the job recommendation device can use the same segmentation method as the training process of the inference model, segmenting each question-answer pair one by one, either by sentence or by RDF triples. i a i >The question text q i Divide into multiple questions [q] s1 ,q s2 ,…,q sc ] and each question and answer pair i a i >The answer text a i Divide into multiple answer sentences [a s1 ,a s2 ,…,a sp ], where c represents the question text q i The number of questions in the text, p represents the number of answers in the text a. i The number of answers in the text.

[0081] After segmenting to obtain each question-and-answer pair i a i > Multiple questions [q s1 ,q s2 ,…,q sc ] and multiple answers [a s1 ,a s2 ,…,a sp Afterwards, the job recommendation device can use the trained encoding module to match each question-answer pair one by one. i a i > Multiple questions [q s1 ,q s2 ,…,q sc Encode the question and answer pair to obtain the answer pair. i a i The encoded representation vector of each question in the text [r] qs1 r qs2 ,...,r qsc Similarly, the job recommendation device can also use this coding module to match each question-and-answer pair one by one. i a i Multiple answers [a s1 a s2 , ..., a sp Encode the question and answer pair to obtain the answer pair. i a i The encoded representation vector of each answer in > as1 r as2 ​​​​​​​​,...,r asp ].

[0082] In obtaining each question and answer pair i a i The encoded representation vector of each question in the text [r] qs1 r qs2 ,...,r qsc ] and the encoded representation vector of each answer [r as1 r as2 ,...,r asp After that, the job recommendation device can match each question and answer pair one by one. i a i The encoded representation vector of each question [r] qs1 r qs2 ,...,r qsc ] and the encoded representation vector of each answer [r as1 r as2 ,...,r asp Input the trained relational graph network module, and use each node of the relational graph network module to process each question-answer pair. i a i The question and answer statements are iteratively reasoned to achieve a complete question-answer pair. i a i The information exchange between the questions and answers in the text, and the generation of each question-answer pair. i a i >The logical relationship representation vector of each question in the text [l qs1 , l qs2 , ..., l qsc ] and the logical relationship representation vector of each answer [l as1 , l as2 , ..., l asp ].

[0083] Specifically, targeting Figure 3 The diagram shows a relational graph network with fewer nodes. The relational graph network module can rely on information transfer between nodes in the network to achieve information interaction between questions, between questions and answers, and between answers, thus representing semantic relationships between questions, between questions and answers, and between answers. When using the relational graph network for iterative reasoning, the relational graph network module can be based on the learning weights w determined in previous training steps. i w n w j For each node's representation vector h i Perform weighted integration, that is Here, the representation vector h of each question node i For question and answer <q​​​​​i a i The encoded representation vector of each question in the text [r] qs1 r qs2 ,...,r qsc The representation vector h of each answer node. i For question and answer i a i The encoded representation vector of each answer in > as1 r as2 ,...,r asp The relational graph network module can analyze each question-answer pair. i a i The encoded representation vector of each question [r] qs1 r qs2 ,...,r qsc ], and the encoded representation vector [r] of each answer. as1 r as2 ,...,r asp Iterate through each question-and-answer pair to obtain them one by one. i a i The logical relationship representation vector of each question in the text [l] qs1 , l qs2 , ..., l qsc ] and the logical relationship representation vector of each answer [l as1 , l as2 , ..., l asp These logical relationships represent vectors [l] qs1 , l qs2 , ..., l qsc ] and [l as1 , l as2 , ..., l asp [Able to characterize each question-answer pair] i a i The semantic relationships between the questions, between the questions and the answers, and between the answers.

[0084] Compared to existing techniques using traditional sequence models, the graph model employed in this invention can effectively shorten the distance between each question and each answer, thereby avoiding the gradient vanishing problem that occurs during sequence modeling due to the large number of sentences, and thus ensuring that the same question-answer pair... i a i The deep interaction of causal relationships between the questions and answers in the text.

[0085] Optionally, in other embodiments of the present invention, for complex relational graph networks including a large number of question nodes and a large number of answer nodes, the relational graph network module can also classify each question-answer pair according to the relationship graph network. i a​​​​​​i >Chinese question text q i and answer text a i The word order and preset window values ​​are used to determine the word order of each question-answer pair. i a i >Each question node q si and each answer node a si The neighboring nodes. For example, in an embodiment with a preset window value of 2, the relationship graph network module can identify question-answer pairs. i a i >Middle and question node q si All question nodes q with an interval less than 2 s(i-1) and q s(i+1) and Q&A pairs i a i All answer nodes a s1 ~a sp All were identified as question nodes q si The neighboring nodes. Similarly, the relational graph network module can also connect question-and-answer pairs. i a i >Middle and answer node a si All answer nodes a with an interval less than 2 s(i-1) and a s(i+1) and Q&A pairs i a i All question nodes q in > s1 ~q sc All are determined to be answer node a si The neighboring nodes. Then, the relationship graph network module can, based on the question-and-answer pairs... i a i Each question node q in > si The encoded representation vectors of the node q and all its neighboring nodes are used to represent each question node q. si The encoded representation vectors are weighted and integrated to achieve question-answer pairs. i a i Each question node q in > si It exchanges information with all its neighboring nodes and generates each question node q. si Logical relation representation vector [l qs1 ,l qs2 ,…,l qsc Similarly, this relational graph network can also be based on question-and-answer pairs. i a i Each answer node a in > si The encoded representation vectors of the node and all its neighboring nodes are used to represent each answer node a. si The encoded representation vectors are weighted and integrated to achieve question-answer pairs. <q​​​​​​​​i , a i > each answer sentence node a si interacts with all its neighbor nodes and generates a logical relation representation vector [l si , l as1 , …, l as2 ] of each answer sentence node a asp respectively.

[0086] By configuring the setting interface of the above window value, the user can set appropriate window values according to the demand of the inference model for the evaluation accuracy of the candidate's various competences, and the data processing ability of the processor of the training device and the post recommendation device, so as to balance the performance of evaluation accuracy and processing speed. The specific value of the window value does not affect the basic requirement of the skilled person in the art to construct a relation graph network, which is not described here.

[0087] After obtaining the logical relation representation vectors [l i , l i , …, l si ] of each question q qs1 in each question-answer pair <q qs2 , a qsc > and the logical relation representation vectors [l si , l as1 , …, l as2 ] of each answer a asp , the post recommendation device can input the logical relation representation vectors [l i , l i , …, l qs1 ] and [l qs2 , l qsc , …, l as1 ] of each question-answer pair <q as2 , a asp > into the trained semantic inference module one by one to initialize each node of the semantic inference module.

[0088] As mentioned above, the representation vector h si of each question node q i in the semantic inference module is the logical relation representation vector l qsi obtained after inference iteration of the relation graph network module, and the representation vector h si of each answer node a i is the logical relation representation vector l asi obtained after inference iteration of the relation graph network module. When using the semantic inference module for semantic inference, the representation vector h i of each node of the trained graph attention network (GAT) is based on its own h iand neighbor nodes h associated therewith j performing weighted integration, i.e. wherein w i , w n , and w j are learning weights determined through the foregoing training process, and nc represents the number of neighbor nodes of the node h i . In this way, the graph attention network can rely on information transmission between nodes to perform semantic interaction between questions, between questions and answers, and between answers, so as to realize semantic reasoning between questions, between questions and answers, and between answers, and generate semantic representation vectors [h1, h2,..., h i ] of each question-answer pair <q i , a c+p > respectively. The semantic representation vectors [h1, h2,..., h c+p ] can represent semantic associations between questions, between questions and answers, and between answers in each question-answer pair <q i , a i >.

[0089] Compared with the prior art using a traditional sequence model, the present application uses a graph attention network model to model semantic reasoning of question-answer pairs <q i , a i >, which can effectively shorten the distance between questions and answers, thereby avoiding the gradient vanishing problem caused by excessive number of sentences in the sequence model modeling process, and further ensuring deep interaction between semantics of questions and answers in the same question-answer pair <q i , a i >.

[0090] In some embodiments, the post recommendation device can perform max-pooling or mean-pooling operations on the obtained semantic representation vectors [h1, h2,..., h c+p ] to obtain a semantic representation vector H[h1, h2,..., h c+p ] of uniform dimension. In this way, even if the number of answer sentences in the answer texts of different candidates is different, the classification network module of the backend can accurately determine the score results T i of the candidate for each competency according to the semantic representation vector H of uniform dimension.

[0091] After generating the semantic representation vectors H[h1, h2,..., h i ] of each question-answer pair <q i , a c+p > using the semantic reasoning module, the post recommendation device cani Each input is a classification network module corresponding to a competency dimension, which then classifies the input semantic representation vector h. i Determine the classification result for each corresponding competency. This classification result indicates to the candidate the rating result T for a competency. i .

[0092] like Figure 4 As shown, the candidate's scores for each competency are determined as T[T1, T2, ..., T]. k After that, the job recommendation device can use these rating results T[T1, T2, ..., T] k [and at least one combination of competencies required for the advertised position C] i [c1, c2, ..., c n [This will help determine the suitable recommended position for the candidate.]

[0093] Specifically, in some binary scoring implementations, the various scoring results T i The results can include values ​​of 0 and 1, where 0 indicates the candidate does not possess the competency, and 1 indicates the candidate possesses the competency. The job recommendation device can filter candidates based on their ratings for each competency according to a preset rating threshold 1 to determine at least one qualified competency for each candidate. For example, if a candidate x's rating result T... x = [1,1,0,1,0], the job recommendation device can determine the competence of candidate x, including C. x [c1,c2,c4].

[0094] Subsequently, the job recommendation device can use a simple sequence matching method to sequentially determine at least one competency C of candidate x. x [c1,c2,c4] and the competency combination C required for each job position. i A comparison is conducted to determine the at least one competency C that meets the standard. x Capable of comprehensively covering competency combination C i At least one job opening. For example, if a job opening i requires a competency combination of C... i [c1,c2], the job recommendation device can determine at least one of the candidate x's competencies C. x [c1,c2,c4] can comprehensively cover the competency combination C required for the job position i. i [c1,c2], thus determining the job posting i as the recommended job posting for candidate x.

[0095] Furthermore, if another job posting j requires a competency combination of C... j[c2,c4] the post recommendation device can determine that the candidate x has at least one up-to-standard competence C x [c1,c2,c4] can also comprehensively cover the required competence combination C of the recruitment post j j [c2,c4], so as to determine the recruitment post j as the recommended post of the candidate x, and add the recruitment post i and the recruitment post j to the recommended post list of the candidate x.

[0096] On the contrary, if the candidate x has at least one up-to-standard competence C x [c1,c2,c4] cannot comprehensively cover the required competence combination C of any recruitment post i The post recommendation device can determine that there is no recommended post suitable for the candidate x at present, and output a result of no recommendation.

[0097] In this way, the post recommendation device provided by the present application can on the one hand use artificial intelligence technology to perform question and answer interaction on the question text and answer text of the candidate's answers to various questions from the perspective of competence assessment, thereby improving the evaluation accuracy of artificial intelligence on the various competences of the candidate, and on the other hand automatically recommend suitable posts for the candidate according to the evaluation results of the various competences of the candidate, thereby substantially liberating the interviewers of the enterprise.

[0098] Those skilled in the art can understand that the above-mentioned binary scoring method is only a non-limiting embodiment of the present application, which is intended to clearly demonstrate the main idea of the present application and provide a specific scheme for the public to implement, rather than to limit the protection scope of the present application.

[0099] Optionally, in some other ternary scoring method embodiments, each scoring result T i may include the value results of 0, 1 and 2, wherein 0 indicates that the candidate does not have the competence, 1 indicates that the candidate basically has the competence, and 2 indicates that the competence of the candidate is relatively outstanding. The post recommendation device can set a corresponding scoring threshold according to the specific needs of the recruitment enterprise, and then filter the scoring results of the candidate for each competence according to the set scoring threshold, to determine at least one up-to-standard competence of the candidate. For example, if the recruitment enterprise is in urgent need of employees, the post recommendation device can set the scoring threshold to 1, and widely filter the candidates who basically have the required competences, in order to improve the number of up-to-standard candidates. For another example, if the recruitment enterprise only needs outstanding talents, the post recommendation device can set the scoring threshold to 2, and finely filter the candidates whose required competences are relatively outstanding, in order to further improve the quality of the up-to-standard candidates.

[0100] Further, in some embodiments of the present application, in order to meet the recruitment needs of the recruitment enterprise, the post recommendation device can also obtain the demand information of the recruitment post provided by the recruitment enterprise from the recruitment platform. Then, the post recommendation device can screen the recommended post list of a large number of candidates who have undergone the above-mentioned competence assessment according to the recruitment post provided by the recruitment enterprise, determine the candidates recorded in the recommended post list as the recommended candidates suitable for the recruitment post, and recommend the recommended candidates to the recruitment enterprise. By adopting this asynchronous recruitment mode of pre-assessing the competence of the candidates and screening suitable recommended candidates according to the recruitment needs of the recruitment enterprise, the present application can further get rid of the time limit of the interview behavior for both the recruitment parties, so as to comprehensively recommend suitable posts for the candidates and comprehensively recommend suitable talents for the recruitment enterprise.

[0101] Further, after the recommended candidates are recommended to the recruitment enterprise, the post recommendation device can also provide the recruitment enterprise with the answer text of the recommended candidate to each question and the corresponding question text, and / or the video and / or audio of the recommended candidate answering each question, and / or the scoring results of the recommended candidate for each competence, so as to provide theoretical support and data support for the recommended results, and provide guidance for the subsequent talent employment of the recruitment enterprise according to the evaluation results of each competence of the recommended candidate.

[0102] Based on the above description, the post recommendation device provided by the present application can use the pre-constructed and trained competence model and inference model to interact with the question text and answer text to improve the evaluation accuracy of artificial intelligence for each competence of the candidate, and automatically recommend suitable posts for the candidate according to the evaluation results of each competence of the candidate.

[0103] Although the above-described methods are illustrated and described as a series of acts, it will be appreciated and understood that the methods are not limited by the order of acts, as some acts can, in accordance with one or more embodiments, occur simultaneously or in different order than shown and described herein, or can be omitted entirely, depending on the circumstances. Accordingly, not all steps of the methods are required, and the methods can include additional steps not shown or described herein.

[0104] The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be apparent to those skilled in the art, and the generic principles defined herein can be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein, but is to be accorded the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A job recommendation method characterized by comprising: The method comprises the following steps: asking the candidate a plurality of questions on a plurality of competencies of the candidate, wherein the plurality of competencies are determined according to competency models of a plurality of job positions; collecting answer texts of the candidate answering each of the questions, and forming each of the answer texts and a corresponding question text into a question-answer pair; segmenting the question text and the answer text of each of the question-answer pairs to obtain a plurality of question sentences and a plurality of answer sentences of each of the question-answer pairs; using an encoding module in a pre-trained reasoning model to perform vector representation on each of the question sentences and each of the answer sentences in each of the question-answer pairs to generate an encoding representation vector of each of the question sentences and an encoding representation vector of each of the answer sentences in each of the question-answer pairs; using a relation graph network module in the pre-trained reasoning model to perform iterative reasoning on the encoding representation vector of each of the question sentences and the encoding representation vector of each of the answer sentences in each of the question-answer pairs to realize information interaction between each of the question sentences and each of the answer sentences in each of the question-answer pairs, and generate a logical relation representation vector of each of the question sentences and a logical relation representation vector of each of the answer sentences in each of the question-answer pairs; performing semantic reasoning on the logical relation representation vector of each of the question sentences and the logical relation representation vector of each of the answer sentences in each of the question-answer pairs to generate a semantic representation vector of each of the question-answer pairs; and classifying results according to the semantic representation vector of each of the question-answer pairs to obtain a score result of the candidate on each of the competencies; and determining a recommended position suitable for the candidate according to the score result of the candidate on each of the competencies and a required competency combination of the plurality of job positions.

2. The job recommendation method according to claim 1, wherein The step of asking the candidate a plurality of questions on a plurality of competencies of the candidate comprises: calling one or more questions corresponding to each of the competencies from a pre-constructed competency question bank; and playing a video or audio of each of the questions to the candidate.

3. The job recommendation method according to claim 2, characterized by, Before the step of asking the candidate a plurality of questions on a plurality of competencies of the candidate, the position recommendation method further comprises the following steps: interviewing a plurality of employees of a plurality of positions to obtain interview records of the plurality of employees; scoring the interview records of each of the employees according to pre-defined competency dimensions to obtain a score of each of the employees on each of the competencies; dividing the plurality of employees into excellent employees and ordinary employees of each of the positions according to work performance; performing difference test on the scores of the excellent employees and the ordinary employees on each of the competencies to determine a plurality of competencies that the excellent employees of each of the positions should possess; constructing a competency model of each of the positions according to the plurality of competencies that the excellent employees of each of the positions should possess; and preparing at least one question for each of the competencies, and constructing the competency question bank based on the plurality of competencies that the excellent employees of the plurality of positions should possess.

4. The job recommendation method according to claim 2, wherein The step of collecting answer texts of the candidate answering each of the questions comprises: collecting videos or audios of the candidate answering each of the questions; and performing speech-to-text transcription on audio data in the video or audio to obtain answer texts of the candidate persons answering each of the questions.

5. The job recommendation method according to claim 1, wherein The step of segmenting the question text and the answer text of each of the question-answer pairs comprises: segmenting the question text and the answer text of each of the question-answer pairs in units of sentences; or segmenting the question text and the answer text of each of the question-answer pairs in units of RDF triples.

6. The job recommendation method according to claim 1, wherein The step of performing iterative reasoning on the encoding representation vector of each of the question sentences and the encoding representation vector of each of the answer sentences in each of the question-answer pairs comprises: weighting and integrating the encoding representation vector of each of the question sentences in each of the question-answer pairs according to the encoding representation vector of each of the question sentences, the encoding representation vector of the rest of the question sentences, and the encoding representation vector of each of the answer sentences, to realize information interaction between each of the question sentences and the rest of the question sentences and each of the answer sentences in each of the question-answer pairs, and to generate a logical relationship representation vector of each of the question sentences; and weighting and integrating the encoding representation vector of each of the answer sentences in each of the question-answer pairs according to the encoding representation vector of each of the answer sentences, the encoding representation vector of the rest of the answer sentences, and the encoding representation vector of each of the question sentences, to realize information interaction between each of the answer sentences and the rest of the answer sentences and each of the question sentences in each of the question-answer pairs, and to generate a logical relationship representation vector of each of the answer sentences.

7. The job recommendation method according to claim 6, wherein The step of performing iterative reasoning on the encoding representation vector of each of the question sentences and the encoding representation vector of each of the answer sentences in each of the question-answer pairs further comprises: determining neighbor nodes of each of the question sentence nodes and each of the answer sentence nodes in each of the question-answer pairs according to the order of the question text and the answer text in each of the question-answer pairs, wherein the neighbor nodes of the question sentence nodes comprise all question sentence nodes in the question-answer pairs that are spaced less than a preset window value from the question sentence nodes, and all answer sentence nodes in the question-answer pairs, and the neighbor nodes of the answer sentence nodes comprise all answer sentence nodes in the question-answer pairs that are spaced less than the preset window value from the answer sentence nodes, and all question sentence nodes in the question-answer pairs; weighting and integrating the encoding representation vector of each of the question sentence nodes in each of the question-answer pairs according to the encoding representation vector of each of the question sentence nodes and all of its neighbor nodes, to realize information interaction between each of the question sentence nodes and all of its neighbor nodes in each of the question-answer pairs, and to generate a logical relationship representation vector of each of the question sentence nodes; and weighting and integrating the encoding representation vector of each of the answer sentence nodes in each of the question-answer pairs according to the encoding representation vector of each of the answer sentence nodes and all of its neighbor nodes, to realize information interaction between each of the answer sentence nodes and all of its neighbor nodes in each of the question-answer pairs, and to generate a logical relationship representation vector of each of the answer sentence nodes.

8. The job recommendation method according to Claim 1, wherein The reasoning model further comprises a pre-trained semantic reasoning module, and the step of performing semantic reasoning on the logical relation representation vector of each question and the logical relation representation vector of each answer in each question-answer pair to generate a semantic representation vector of each question-answer pair comprises: The semantic reasoning module is used to perform semantic reasoning on the logical relation representation vector of each question and the logical relation representation vector of each answer in each question-answer pair to realize semantic interaction between each question and each answer in each question-answer pair, and generate a semantic representation vector of each question-answer pair.

9. The job recommendation method according to claim 8, wherein The reasoning model further comprises a plurality of pre-trained classification network modules, and the step of performing result classification according to the semantic representation vector of each question-answer pair to obtain a score result of the candidate for each competency comprises: The semantic representation vector of each question-answer pair is input into a corresponding classification network module to obtain a score result of the candidate for each competency.

10. The job recommendation method according to Claim 1, wherein The step of determining a recommended position suitable for the candidate according to the score result of the candidate for each competency and a competency combination required by at least one recruitment position comprises: According to a predetermined score threshold, the score result of the candidate for each competency is screened to determine at least one qualified competency of the candidate; The at least one qualified competency is compared with the competency combination required by at least one recruitment position to determine that the at least one qualified competency can comprehensively cover at least one recruitment position of the competency combination; and The at least one recruitment position that the at least one qualified competency can comprehensively cover is determined as the recommended position of the candidate.

11. The job recommendation method according to claim 10, wherein After the step of determining a recommended position suitable for the candidate according to the score result of the candidate for each competency and a competency combination required by at least one recruitment position, the position recommendation method further comprises the following steps: According to the recruitment position provided by the recruitment enterprise, a recommended position list of a plurality of candidates is screened to determine a recommended candidate of the recruitment position, wherein the recommended position list records at least one recommended position suitable for a corresponding candidate; and The recommended candidate is recommended to the recruitment enterprise.

12. The job recommendation method according to claim 11, wherein After the step of recommending the recommended candidate to the recruitment enterprise, the position recommendation method further comprises the following steps: The answer text of the recommended candidate to each question and the corresponding question text, and / or the video and / or audio of the recommended candidate answering each question, and / or the score result of the recommended candidate for each competency are provided to the recruitment enterprise.

13. The job recommendation method according to Claim 1, wherein Before the step of asking a plurality of questions to the candidate for a plurality of competencies, the position recommendation method further comprises the following steps: According to the plurality of competencies, a plurality of corresponding questions are asked to a plurality of candidate samples; and respectively, and each of the answer text samples and the corresponding question text of each of the candidate samples form a question and answer pair sample; segmenting the question text and the answer text sample of each of the question and answer pair samples to obtain a plurality of question sentence samples and a plurality of answer sentence samples of each of the question and answer pair samples; performing competency scoring on the answer text sample of each of the candidate samples for each of the competencies of each of the candidate samples; and using the plurality of question sentence samples and the plurality of answer sentence samples of each of the question and answer pair samples and the competency scores of the corresponding candidate sample answering each of the questions, training the inference model to perform the functions of the information interaction, the semantic reasoning and / or the result classification.

14. A job recommendation device characterized by comprising: comprise: a memory; and a processor connected to the memory and configured to implement the job recommendation method according to any one of claims 1-13.

15. A computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions are executed by the processor to implement the job recommendation method according to any one of claims 1-13.

Images