3D dynamic character resume generation method
By using the BERT deep learning model to encode text resumes into character feature vectors and generate 3D dynamic character resumes, the problem of converting text resumes into 3D dynamic character resumes in the existing technology is solved, and personalized display and efficient job hunting are achieved.
Patent Information
- Application Number
- CN202510142723.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-10
- Publication Date
- 2025-05-30
AI Technical Summary
The existing technology is difficult to convert text resumes quickly and efficiently into 3D dynamic character resumes, resulting in limitations in job seekers' display of themselves and attracting recruitment companies.
The BERT deep learning model is used to encode text data from multiple dimensions, generate overall and stage character image feature vectors, and form 3D dynamic character resumes through splicing and synchronous processing.
It realizes highly personalized 3D character image generation, which can more accurately display the characteristics of job seekers' different stages, reduces the production time of self-introduction of 3D characters, and improves job search efficiency.
Smart Images
Figure CN120070685A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of job seeking and recruitment, and particularly relates to a method for generating a 3D dynamic character resume. Background Art
[0002] In the current job market, as an important bridge for the initial communication between job seekers and employers, the forms of personal resumes are becoming increasingly diverse. Although traditional text resumes can list in detail the educational background, work experience and other information of job seekers, they have limitations in attracting attention and demonstrating personal charm. As a more vivid and intuitive expression, 3D character images are gradually becoming a new means for job seekers to stand out. However, producing high-quality 3D dynamic characters requires professional skills and time investment, which is a great challenge for most job seekers. Therefore, it is particularly urgent to develop a method that can automatically convert text resumes into 3D dynamic character resumes. Summary of the Invention
[0003] The purpose of the present invention is to provide a method for generating a 3D dynamic character resume. The present invention automatically converts a text resume into a 3D dynamic character resume, which can intuitively display the ability characteristics of job seekers, help job seekers better display themselves and help recruitment enterprises quickly understand job seekers.
[0004] The technical solution of the present invention: A method for generating a 3D dynamic character resume includes the following steps:
[0005] Step 1: Obtain the resume text of a person, preprocess the resume text to form labeled text data;
[0006] Step 2: Use the BERT deep learning model to encode the labeled text data into a general character image feature vector from multiple dimensions, and generate a general character image according to the general character image feature vector;
[0007] Step 3: Divide the experiences of the job seeker into different stages according to the resume text, and assign a unique identifier to each stage; then extract the specific information of the job seeker for each stage respectively, and use the BERT deep learning model to encode the specific information into an image feature vector for this stage, and generate a stage character image that matches the image feature vector for this stage according to the image feature vectors for different stages;
[0008] Step 4: For each stage, generate the corresponding introduction text according to the extracted specific information, and then convert the introduction text into a text feature vector;
[0009] Step 5: Use the general character image feature vector as the cover character image of the 3D dynamic character resume, and splice the stage character images and embed them into the text feature vectors corresponding to the stages to form the self-introduction character information in the 3D dynamic character resume.
[0010] In the above 3D dynamic character resume generation method, in step 1, the preprocessing of the resume text is to extract the text content of the resume text using the optical character recognition method, and then remove noise, segment words, and perform part-of-speech tagging on the text content to form annotated text data.
[0011] In the aforementioned 3D dynamic character resume generation method, in step 2, the multi-dimensions include basic information, educational background, work experience, professional skills, and honors and specialties;
[0012] The feature vector representation of the overall character image features is as follows:
[0013] V total = f BERT (text);
[0014] In the formula: V total represents the feature vector of the overall character image features, and f BERT represents the BERT deep learning model.
[0015] In the aforementioned 3D dynamic character resume generation method, the specific information includes school name, major, job responsibilities, skills, and project experience.
[0016] In the aforementioned 3D dynamic character resume generation method, the process of encoding the specific information into the image feature vector at this stage using the BERT deep learning model is as follows:
[0017] The self-attention mechanism is used to calculate the relationship between each word and all other words and update its representation:
[0018]
[0019] In the formula: Q, K, and V are the matrices of query, key, and value respectively, d k is the dimension of the key vector, and T is the transpose operation;
[0020] Then, the multi-head attention mechanism is adopted, and this process is repeated multiple times, and the results are concatenated:
[0021] MultiHead(Q, K, V) = Concat(head 1 , ···, head h )W O ;
[0022] In the formula: head i = Attention(QW i Q , KW i K , VW i V),W i Q ,W i K ,W i V and W O are the training weight matrices respectively;
[0023] Then, a feed - forward neural network is used to perform a further non - linear transformation on the output of the self - attention layer:
[0024] FFN(x)=max(0,xW 1 +b 1 )W 2 +b 2 ;
[0025] In the formula, W 1 and W 2 are the trainable weight matrices respectively, b 1 and b 2 are the trainable bias terms; FFN(x) represents the output stage image feature vector.
[0026] For the aforementioned 3D dynamic resume generation method, after obtaining the stage image feature vector, the attention mechanism is used to optimize and adjust the stage image feature vector, and the process is as follows:
[0027] Calculate the attention scores of the stage image feature vector V i and other stage feature vectors through dot product:
[0028]
[0029] In the formula: V i and V j represent the i - th stage and j - th stage image feature vectors respectively; V ik and V jk represent the k - th elements of the image feature vectors V i and V j respectively, and d is the dimension of the feature vector;
[0030] To avoid the values being too large or too small, the attention scores are scaled, and the scaling factor is the reciprocal of the square root of the dimension d of the feature vector:
[0031]
[0032] The attention scores of all stages are normalized through the softmax function and converted into attention weights to ensure that the sum of the weights of all stages is 1:
[0033]
[0034] where m is any stage index other than i;
[0035] According to the attention weights, the feature vectors of each stage are weighted and summed to obtain the adjusted feature vectors;
[0036]
[0037] In the formula: is the stage image feature vector of the i-th stage after optimization and adjustment.
[0038] In the aforementioned 3D dynamic character resume generation method, in step 3, multi-dimensional clustering analysis is also performed on the overall character image feature vector through the clustering algorithm and the stage image feature vector to identify the main characteristics and career tendencies of the job seeker. The specific calculation is as follows:
[0039]
[0040] where SSE is the sum of squared errors within the cluster, C i represents the i-th cluster of the overall character image feature vector, and μ i represents the centroid of the i-th cluster of the overall character image feature vector, is the stage image feature vector of the i-th stage after optimization and adjustment.
[0041] In the aforementioned 3D dynamic character resume generation method, in step 4, the specific process of generating the corresponding introduction text for the specific information is as follows:
[0042] Decompose the specific information text extracted from the resume into words and perform part-of-speech tagging on them. Then name the entities of the words and determine the key information points in the resume; then use the term frequency-inverse document frequency method to calculate the scores of the named entities, and determine the keywords by sorting the entity scores:
[0043] tfidf(t, d, D) = tf(t, d) × idf(D);
[0044] In the formula: f(t, d) is the term frequency of word t in document d; idf(D) is the inverse document frequency of word t in the document set D; tfidf(t, d, D) is the TF-IDF score of word t in document d;
[0045] Finally, according to the extracted key information points and keywords, use the sequence-to-sequence Seq2Seq model to generate the character introduction text:
[0046]
[0047] In the formula: P = (y 1 , y 2 , ···, yT |x 1 ,x 2 ,···,x S ) is the probability of output sequence y given the input sequence x; P = (y t |y 1 ,···,y t-1 ,x 1 ,···,x S ) is the probability of the t-th word in the output sequence given the previous outputs and inputs; o represents the length of the generated introduction text;
[0048] Then, the generated introduction text is converted into a text feature vector through the word embedding method.
[0049] In the 3D dynamic character resume generation method described above, in step 5, the phased text feature vectors are synthesized into introduction voices through the text-to-speech method, and then the generated introduction voices are synchronized with the phased character images to ensure the matching degree and fluency.
[0050] Compared with the prior art, the present invention has the following beneficial effects:
[0051] 1. The present invention combines deep learning and feature vector technology to realize the generation of highly personalized 3D character images, meeting the diverse needs of job seekers.
[0052] 2. Through phased processing and information extraction technology, the present invention can more accurately capture and display the characteristics of different stages of job seekers, facilitating the display of the advantages of job seekers and also enabling recruitment companies to understand and screen.
[0053] 3. The present invention greatly shortens the production time of the self-introduction 3D character, reduces the production threshold, and improves the job hunting efficiency.
[0054] In summary, the present invention improves the application scenarios of the recruitment platform and provides a more comprehensive and convenient way for job seekers and recruiters to understand each other. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 is a schematic flow chart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0056] The present invention will be further described below with reference to the drawings and embodiments, but it shall not be used as a basis for limiting the present invention.
[0057] Embodiment: A 3D dynamic character resume generation method includes the following steps:
[0058] Step 1: Obtain the resume text of a person, preprocess the resume text to form labeled text data. In this step, the resume file is a word, PDF, or picture file uploaded by the user to the corresponding recruitment platform, and then the OCR method is used to extract the text content. Among them, the OCR method is an optical character recognition method. Its main process is to process the input file, including grayscale conversion, noise reduction, binarization, skew correction, etc., to improve the file quality, then separate the characters in the text image, and then extract features according to the structure, strokes, etc. of the characters. Finally, the extracted features are compared with the predefined font library to find the most matching characters.
[0059] After the text is extracted, preprocess the extracted text content, including removing noise, word segmentation, part-of-speech tagging, etc., to form labeled text data.
[0060] Step 2: Use the BERT deep learning model to encode the labeled text data into a feature vector of the overall personal image characteristics from multiple dimensions, and generate an overall personal image according to the feature vector of the overall personal image characteristics.
[0061] In this step, use the pre-trained BERT deep learning model to encode the preprocessed labeled text data into a feature vector of the overall personal image characteristics from multiple dimensions such as the applicant's basic information, educational background, work experience, professional skills, honors and specialties.
[0062] Among them, the BERT deep learning model is a pre-trained language representation model based on the Transformer architecture, which is used for various tasks in natural language processing (NLP). The training of the BERT deep learning model is divided into two stages: pre-training and fine-tuning. In the pre-training stage, the BERT deep learning model performs unsupervised training on a large-scale corpus, mainly using two tasks: Masked Language Model (MLM) and Next Sentence Prediction (NSP). In the fine-tuning stage, the BERT deep learning model performs supervised training on specific tasks, such as classification, question answering, etc.
[0063] The feature vector of the overall personal image characteristics described in this embodiment is represented as follows:
[0064] V total =f BERT (text);
[0065] In the formula: V total represents the feature vector of the overall personal image characteristics, and f BERT represents the BERT deep learning model.
[0066] Step 3: According to the resume text, divide the job seeker's experiences into different stages (such as educational experience, work experience, etc.), and assign a unique identifier to each stage (the identifier can be generated using the hash coding method, which is used to associate specific information, stage image feature vectors, and corresponding introduction texts, etc. for different stages); then extract the specific information of the job seeker for each stage (such as school name, major, job responsibilities, skills, project experience, etc.), and use the BERT deep learning model to encode the specific information into the stage image feature vector, and generate a stage character image that matches the stage image feature vector according to the different stage image feature vectors;
[0067] In this step, the process of encoding the specific information into the stage image feature vector using the BERT deep learning model is as follows:
[0068] Use the self-attention mechanism to calculate the relationship between each word and all other words, and update its representation:
[0069]
[0070] where: Q, K, and V are the matrices of query, key, and value respectively, d k is the dimension of the key vector, and T is the transpose operation;
[0071] Then use the multi-head attention mechanism, repeat this process multiple times, and concatenate the results:
[0072] MultiHead(Q, K, V) = Concat(head 1 , ···, head h )W O ;
[0073] where: head i = Attention(QW i Q , KW i K , VW i V ), W i Q , W i K , W i V and W O are the training weight matrices respectively;
[0074] Then use the feed-forward neural network to perform further non-linear transformation on the output of the self-attention layer:
[0075] FFN(x) = max(0, xW 1 + b 1 )W2 +b 2 ;
[0076] Wherein, W 1 and W 2 are respectively trainable weight matrices, and b 1 and b 2 are trainable bias terms; FFN(x) represents the output stage image feature vector.
[0077] To highlight the characteristics of each stage, after obtaining the stage image feature vector, an attention mechanism is used to optimize and adjust the stage image feature vector, and the process is as follows:
[0078] Calculate the attention scores of the stage image feature vector V i with other stage feature vectors through dot product:
[0079]
[0080] Where: V i and V j represent the stage image feature vectors of the i-th stage and the j-th stage respectively; V ik and V jk represent the k-th elements of the image feature vectors V i and V j respectively, and d is the dimension of the feature vector;
[0081] To avoid the numerical value being too large or too small, the attention scores are scaled, and the scaling factor used is the reciprocal of the square root of the dimension d of the feature vector:
[0082]
[0083] The attention scores of all stages are normalized through the softmax function to be converted into attention weights, ensuring that the sum of the weights of all stages is 1:
[0084]
[0085] where m is any stage index other than i;
[0086] According to the attention weights, the feature vectors of each stage are weighted and summed to obtain the adjusted feature vector;
[0087]
[0088] Where: is the stage image feature vector of the i-th stage after optimization and adjustment.
[0089] The optimized and adjusted stage image feature vector incorporates information from other stages and highlights the characteristics of the current stage.
[0090] In addition, in this step, a multi-dimensional clustering analysis is also performed on the overall personal image characteristic feature vector through the clustering algorithm and the stage image feature vector to identify the main characteristics and career inclinations of the job seekers. The specific calculation is as follows:
[0091]
[0092] Among them, SSE is the sum of squared errors within the cluster, and C i represents the i-th cluster of the overall personal image characteristic feature vector, and μ i represents the centroid of the i-th cluster of the overall personal image characteristic feature vector. is the stage image feature vector of the i-th stage after optimization and adjustment.
[0093] This way can make the overall personal image show corresponding styles, such as the images of sales elites, technical personnel, mature and steady, eager to learn, etc., which is convenient for the recruitment company to form a first impression through the overall personal image.
[0094] Step 4: For each stage, generate corresponding introduction texts according to the extracted specific information, and then convert the introduction texts into text feature vectors;
[0095] In this step, the specific process of generating corresponding introduction texts from specific information is to decompose the specific information text extracted from the resume into words and perform part-of-speech tagging on them (such as nouns, verbs, adjectives); then name the words as entities (such as names, schools, companies, majors, etc.) and use a method based on configuration rules to determine the key information points in the resume (such as work experience, educational background, skills, etc.); then use the term frequency-inverse document frequency method to calculate the scores of the named entities, and determine the keywords by sorting the entity scores:
[0096] tfidf(t, d, D) = tf(t, d) × idf(D);
[0097] In the formula: f(t, d) is the term frequency of word t in document d; idf(D) is the inverse document frequency of word t in the document set D; tfidf(t, d, D) is the TF-IDF score of word t in document d;
[0098] Finally, according to the extracted key information points and keywords, use the sequence-to-sequence Seq2Seq model to generate the personal introduction text:
[0099]
[0100] In the formula: P = (y 1 , y2 , ···, y T |x 1 , x 2 , ···, x S ) is the probability of output sequence y given the input sequence x; P = (y t |y 1 , ···, y t-1 , x 1 , ···, x S ) is the probability of the t-th word in the output sequence given the previous outputs and inputs; O represents the length of the generated introduction text;
[0101] Then, the generated introduction text is converted into a text feature vector through a word embedding method.
[0102] For example: Hello, I am XXX, 23 years old, from Wuhu, Anhui. I am about to pursue a master's degree in Electronic Information at Anhui University.
[0103] I studied Internet of Things Engineering at the undergraduate level, mastered professional basic knowledge such as C language and data structures, and won the second prize in the C++ group of the Blue Bridge Cup.
[0104] I participated in the back-end development of user reviews and the development of the iFlytek digital human service system, and am proficient in development frameworks such as Java and Spring Boot.
[0105] Then, the generated introduction text is converted into a text feature vector through a word embedding method (such as Word2Vec, GloVe, etc.).
[0106] Step 5: Use the overall character image feature vector as the cover character image of the 3D dynamic resume, and splice the stage character images and embed the corresponding stage text feature vectors to form the self-introduction character information in the 3D dynamic resume.
[0107] In this step, the splicing of the stage character images is to obtain the corresponding stage character images from the model library according to the stage image feature vectors, and then use video editing techniques (such as cutting, synthesis, special effects, etc.) to splice the stage character images into a coherent image video, and at the same time, corresponding animation effects can be matched.
[0108] Synthesize the introduction speech from the stage text feature vectors through text-to-speech (TTS) technology. The TTS system supports simulating different speech styles, speech rates, and intonations to meet the needs of different scenarios and audiences. Then, synchronize the generated introduction speech with the stage character images to ensure the matching degree and fluency of the speech and video content. Finally, synthesize a personalized 3D dynamic resume.
[0109] In this embodiment, both the overall character image and the stage character image can be obtained from the corresponding basic model library. The basic model library can be obtained in the following ways:
[0110] Data collection: Collect high-quality pictures or videos of character images and typical scenes (such as schools, enterprises, factories, laboratories, etc.). These data will be used to train the AI model.
[0111] Preprocessing: Clean the data to remove noise and irrelevant information.
[0112] Data annotation: Add labels to different parts of the character (such as the head, body, limbs).
[0113] Feature extraction: Use a deep learning model, such as a convolutional neural network (CNN), to extract key features from the image, such as facial features, body proportions, postures, etc.
[0114] Model training: Adopt generative adversarial networks (GANs) and variational autoencoders (VAEs) to train one or more AI models using the extracted features; use GPT technology to integrate character features with scene features to construct a comprehensive large-scale composite scene.
[0115] 3D model generation:
[0116] Adopt mesh generation and texture mapping to use the trained AI model to convert 2D images into 3D models.
[0117] Optimization and refinement:
[0118] Optimize the generated 3D model to improve its geometric accuracy and details, such as removing geometric defects and smoothing the surface, etc.
[0119] In summary, the present invention automatically converts a text resume into a 3D dynamic character resume, which can intuitively display the ability characteristics of job seekers, help job seekers better display themselves, and help recruitment enterprises quickly understand job seekers.
Claims
1. A 3D dynamic character resume generation method, characterized by: The steps include: Step 1: Obtain the resume text of the person, pre-process the resume text, and form annotated text data; Step 2: Use the BERT deep learning model to encode the annotated text data into a feature vector of the overall character image characteristics from multiple dimensions, and generate the overall character image based on the feature vector of the overall character image characteristics; Step 3: Based on the resume text, the job seeker's experience is divided into different stages, and a unique identifier is assigned to each stage; then, specific information of the job seeker is extracted for each stage, and the BERT deep learning model is used to encode the specific information into the image feature vector of the stage, and a stage character image matching the image feature vector of the stage is generated based on the image feature vector of the different stages; Step 4: For each stage, generate corresponding introduction text according to the extracted specific information, and then convert the introduction text into a text feature vector; Step 5: Use the overall character image feature vector as the cover character image of the 3D dynamic character resume, and embed the stage character image into the text feature vector of the corresponding stage after splicing to form the self-introduction character information in the 3D dynamic character resume.
2. The 3D dynamic character resume generation method according to claim 1, characterized in that: In step 1, the resume text is preprocessed by using an optical character recognition method to extract the text content of the resume text, and then the text content is de-noised, segmented, and tagged with parts of speech to form tagged text data.
3. The 3D dynamic character resume generation method according to claim 1, characterized in that: In step 2, the multiple dimensions include basic information, educational background, work experience, professional skills and honors; The feature vector of the overall character image characteristics is expressed as follows: V total =f BERT (text); Where: V total The feature vector representing the overall character image characteristics, f BERT Represents the BERT deep learning model.
4. The 3D dynamic character resume generation method according to claim 1, characterized in that: The specific information includes school name, major, job responsibilities, skills and project experience.
5. The 3D dynamic character resume generation method according to claim 1, characterized in that: The process of using the BERT deep learning model to encode specific information into the image feature vector of this stage is: The self-attention mechanism is used to calculate the relationship between each word and all other words and update its representation: Where: Q, K, V are the matrices of query, key, and value respectively, d k is the dimension of the key vector, T is the transpose operation; Then use the multi-head attention mechanism to repeat this process multiple times and concatenate the results: MultiHead(Q,K,V)=Concat(head1,···,head h )W O ; Where: head i =Attention(QW i Q ,KW i K ,VW i V ),W i Q ,W i K ,W i V and W O They are the training weight matrices respectively; The output of the self-attention layer is further transformed nonlinearly using a feedforward neural network: FFN(x)=max(0,xW1+b1)W2+b2; Where W1 and W2 are trainable weight matrices, b1 and b2 are trainable bias terms; FFN(x) represents the output stage image feature vector.
6. The 3D dynamic character resume generation method according to claim 5, characterized in that: After obtaining the stage image feature vector, the attention mechanism is used to optimize and adjust the stage image feature vector. The process is as follows: Calculate the stage image feature vector V by dot product i Attention scores with feature vectors from other stages: Where: V i and V j Respectively represent the image feature vectors of the i-th stage and the j-th stage; V ik and V jk Represent the image feature vector V i and V j The kth element of , d is the dimension of the feature vector; In order to avoid values that are too large or too small, the attention score is scaled, and the scaling factor is the inverse of the square root of the feature vector dimension d: The attention scores of all stages are normalized by the softmax function and converted into attention weights to ensure that the sum of the weights of all stages is 1: where m is any phase index except i; According to the attention weight, the feature vectors of each stage are weighted summed to obtain the adjusted feature vector; Where: is the stage image feature vector of the i-th stage after optimization and adjustment.
7. The 3D dynamic character resume generation method according to claim 6, characterized in that: In step 3, a multi-dimensional cluster analysis is also performed on the overall character image feature vector through clustering algorithm and stage image feature vector to identify the main characteristics and career orientation of job seekers. The specific calculation is as follows: Among them, SSE is the sum of squares of intra-cluster errors, C i Represents the i-th cluster of the feature vector of the overall character image, μ i Represents the centroid of the ith cluster of the feature vector of the overall character image characteristics, is the stage image feature vector of the i-th stage after optimization and adjustment.
8. The 3D dynamic character resume generation method according to claim 1, characterized in that: In step 4, the specific process of generating corresponding introduction text from specific information is as follows: Decompose the specific information text extracted from the resume into words and tag them with parts of speech; then name the words as entities and determine the key information points in the resume; then use the word frequency inverse document frequency method to calculate the score of the named entity, and determine the keywords by sorting the entity scores: tfidf(t,d,D)=tf(t,d)×idf(D); Where: f(t,d) is the frequency of word t in document d; idf(D) is the inverse document frequency of word t in document set D; tfidf(t,d,D) is the TF-IDF score of word t in document d; Finally, based on the extracted key information points and keywords, the Seq2Seq model is used to generate the character introduction text: Where: P = (y1, y2, ···, y T |x1,x2,···,x S ) is the probability of outputting sequence y given input sequence x; P = (y t |y1,···,y t-1 ,x1,···,x S ) is the probability of outputting the tth word in the sequence given the previous output and input; O represents the length of the generated introduction text; The generated introduction text is then converted into a text feature vector through the word embedding method.
9. The 3D dynamic character resume generation method according to claim 8, characterized in that: In step 5, the feature vectors of the staged text characteristics are used to synthesize the introduction speech through the text-to-speech method, and then the generated introduction speech is synchronized with the stage character image to ensure matching and fluency.