System and method for automatically resolving query requests submitted by job seekers or recruiters

Abstract

The application discloses a system and method for automatically analyzing query requests submitted by job seekers or employers, belonging to the technical field of data processing. By receiving the query request submitted by the user, the query statement and the user portrait information are spliced to form an input sequence, and a deep learning model is used to perform deep semantic analysis on the input sequence to generate an embedding vector. The system uses a prefix array to quickly calculate the word vector interval sum, and combines a word segmentation tool to confirm the boundary of the word to generate the embedding representation of the complete word. Through a multi-intent classification module, the embedding vector at the non-MASK position is classified by using a Softmax layer, multiple intents in the query are identified, and corresponding category labels are output. The application can realize deep semantic understanding, multi-intent separation and personalized recommendation, and provide efficient and accurate service support for the recruitment platform and job seekers.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of data processing, and specifically relates to a system and method for automatically analyzing query requests submitted by job seekers or employers. BACKGROUND

[0002] In the modern recruitment field, with the rapid growth of job information and the number of job seekers, how to efficiently and accurately analyze the query requests submitted by job seekers or employers has become the key to improving the efficiency of job matching.

[0003] The problems faced in the modern recruitment field include:

[0004] - Query intent is ambiguous: Users often search queries containing brief expressions, industry jargon or non-standard language, making it difficult to understand the intent. For example, users may search for "front-end development", but do not specify whether it refers to web front-end or mobile application front-end.

[0005] - Multiple intent mixing: A single query may contain multiple recruitment needs or job-seeking intentions, and traditional analysis methods are difficult to effectively distinguish. For example, users may search for "data analyst" and "data scientist" in the same query, hoping to find both types of positions.

[0006] - Insufficient personalized matching: The matching needs of job seekers' preferences, skills and enterprise culture are complex, and existing technologies are difficult to fully meet. Different job seekers have different expectations for work environment, job requirements or company culture, and traditional methods are difficult to accurately capture these personalized needs.

[0007] - Information overload processing: In the face of massive job information and resumes, how to quickly extract key information and reduce information noise is another challenge. Users' search queries contain a large amount of information, and how to find the most relevant positions or resumes from them is the core task of recruitment platforms.

[0008] The main technologies of the existing technology include:

[0009] Keyword matching: Based on keyword extraction for preliminary screening, but it is easy to miss or misjudge due to insufficient semantic understanding. For example, "Java development engineer" is found through keyword matching, but it does not consider that the user may need a developer with specific industry experience. Rule engine: According to the preset rules to filter information, but it is difficult to flexibly respond to changing query requirements and individual preferences. For example, the rule engine may set a rule that "job seekers must have 3 years of work experience", but it does not consider that some positions may require 5 years or more experience. NLP preliminary exploration: Apply natural language processing technology for semantic understanding, but the depth and breadth still need to be improved. For example, NLP can understand the keywords in the user's query, but the depth of understanding of the user's intent and the ability of context analysis are limited.

[0010] However, the existing technology still has many problems:

[0011] For example, the semantic understanding is shallow: the deep semantic understanding of complex query statements is insufficient, and it is difficult to accurately grasp the user's real intention. For example, the user queries "need to understand SEO copywriting editor", the traditional NLP may only understand "SEO" and "copywriting editor", and does not understand that the user needs is a person who is good at both SEO skills and content creation.

[0012] Poor dynamic adaptability: slow technology update iteration, difficult to quickly adapt to new demands and new terms in the recruitment market. For example, new technical terms such as "blockchain development" may not be recognized in the old analysis system.

[0013] Lack of personalized analysis: lack of dynamic analysis strategy for personalized needs of users, resulting in low matching accuracy. For example, when matching the personalized needs of different users, the traditional algorithm may not be able to distinguish the specific requirements of the user for the position.

[0014] Weak multi-intention processing capability: when processing queries containing multiple intentions, it can only provide single or generalized results. For example, a user may search for "front-end development" and "back-end development" in the same query, and the existing system may only provide job information for either. SUMMARY

[0015] Problems to be solved

[0016] Based on the above reasons, the present application aims to solve the following technical problems:

[0017] 1. Deep semantic analysis: through deep learning and advanced NLP technology, deeply understand the intention and context behind the query. For example, using BERT model for deep semantic analysis can better understand the implicit intention in the user query.

[0018] 2. Multi-intention intelligent separation: develop algorithm model to effectively identify and separate multiple intentions in the query, and realize accurate response. For example, design a multi-classification model to decompose the user's complex query into multiple single intentions.

[0019] 3. Personalized matching optimization: combine user portrait and behavior analysis to optimize query analysis strategy and improve personalized matching effect. For example, according to the user's historical behavior and preferences, dynamically adjust the weight of the matching algorithm.

[0020] Deep semantic parsing: through deep learning and advanced natural language processing (NLP) technology, deeply understand the intention and context behind the query request submitted by the user. Traditional methods often only stay on the surface when understanding the true intention of the user, and cannot accurately grasp the user's needs. Specifically, the invention aims to solve the following problems: use a pre-trained model (such as BERT) to represent the query as a vector, capturing the semantic information in the query; calculate the interval sum of the token through the prefix array (preSum), obtain the word vector representation, and further improve the accuracy of semantic parsing.

[0021] Multi-intention intelligent separation: develop an algorithm model to effectively identify and separate multiple intentions in the query, and achieve accurate response. The user's search query may contain multiple recruitment needs or job-seeking intentions, and traditional parsing methods cannot effectively distinguish these multiple intentions. Specifically, the invention solves this problem by using the following methods: using a word segmentation tool to identify the boundaries of words, and performing multi-classification on the vector of each non-MASK position through a softmax layer to identify and separate multiple intentions in the query.

[0022] Personalized matching optimization: combine user portrait and behavior analysis to optimize query parsing strategy and improve personalized matching effect. The matching needs of job seekers' preferences, skill strengths and enterprise culture are complex, and existing technologies cannot fully meet them. Specifically, the invention does this by: concatenating the user's query statement and user portrait information (associated with the industry in which the user is located according to the user's userid) through special characters to obtain more rich user portrait information. In the query parsing process, dynamically adjust the weight of the matching algorithm, and perform personalized matching optimization according to the user's historical behavior and preferences.

[0023] Technical solutions

[0024] To solve the above problems, the invention uses the following technical solutions.

[0025] A system for automatically parsing query requests submitted by job seekers or recruiters,

[0026] A data input module for receiving query requests submitted by job seekers or recruiters, and concatenating query statements and user portrait information through special characters [seq] to form an input sequence;

[0027] A semantic parsing module based on a Transformer Encoder model for deep semantic parsing of the input sequence, decomposing the query statement into several basic units of tokens and generating corresponding embedding vectors;

[0028] A prefix array processing module for calculating the interval sum of the word vector through the prefix array preSum, including:

[0029] Confirming the boundaries of words according to a word segmentation tool;

[0030] The word vector of any interval [i, j] is quickly calculated by the formula preSum[j]-preSum[i-1], and the average of the vectors in the interval is calculated to generate the embedding representation of the complete word;

[0031] The multi-intent classification module inputs the generated embedding vector of each non-MASK position into a Softmax layer for multi-classification, identifies multiple intents in the query request, and outputs the corresponding category label;

[0032] The personalized matching module dynamically adjusts the weight of the classification result in combination with user portrait information and historical behavior data, optimizes the matching strategy, and generates personalized recommendation results that meet user needs.

[0033] Preferably, the semantic analysis module realizes deep semantic analysis through the following steps:

[0034] The user's query statement and user portrait information are spliced by a special character [seq] to form an input sequence;

[0035] The input sequence is vectorized using a Transformer Encoder to generate an embedding vector for each basic unit token;

[0036] The word vector sum of any interval is calculated by the prefix array preSum, specifically including:

[0037] After confirming the boundaries of the word, the word vector of the interval [i, j] is quickly calculated using the formula preSum[j]-preSum[i-1];

[0038] The average of the vectors in the interval is calculated to generate the embedding representation of the complete word;

[0039] In combination with context information, the self-attention mechanism is used to capture the dependency between words in the query statement, enhancing the context awareness ability of the semantic representation;

[0040] Position encoding is introduced to incorporate word order information into word embedding, enabling the model to capture word order structure features.

[0041] Preferably, the multi-intent classification module realizes multi-intent identification through the following steps:

[0042] Each embedding vector of a non-MAS position generated by the semantic analysis module is input into a Softmax layer for multi-classification;

[0043] The Softmax layer outputs the corresponding category label, where the category label includes but is not limited to "position", "industry", "skill", and "language", for identifying multiple intents in the query request;

[0044] The identified multiple intents are separated, and independent resolution results are generated for each intent, realizing hierarchical recognition of query intent;

[0045] Combined with user portrait information, the weight of the classification result is dynamically adjusted to ensure that the classification result is highly relevant to user needs.

[0046] Preferably, the personalized matching module realizes personalized recommendation through the following steps:

[0047] Obtain the user's query request and user portrait information, including but not limited to the user's industry, position preference, and skill specialty;

[0048] According to the category label output by the multi-intent classification module, dynamically adjust the weight of each category, and preferentially match the category with higher relevance to the user portrait information;

[0049] Use user historical behavior data, including browsing records and click habits, to build a user interest model to further optimize the recommendation strategy;

[0050] Combined with the adjusted category weight and user interest model, generate personalized recommendation results that meet user needs;

[0051] Sort and filter the generated personalized recommendation results, and select the highest confidence results to output to the user end.

[0052] Preferably, the prefix array processing module realizes efficient calculation through the following steps:

[0053] Build prefix array preSum: preSum[j] represents the cumulative vector sum from the start of the sequence to position j;

[0054] Quickly calculate the cumulative sum of any interval [i, j]: use the formula preSum[j]-preSum[i-1] to calculate the cumulative sum of all token embedding vectors in the interval [i, j] in O(1) time;

[0055] Generate word embedding representation: average the cumulative vector in the interval [i, j] to get the embedding representation of the complete word;

[0056] Combine the word segmentation tool to confirm the word boundary: accurately locate the start and end positions of each word through word segmentation to ensure that the interval calculation result accurately corresponds to the semantic information of the word.

[0057] Preferably, the system further comprises a user portrait enhancement module for optimizing user portrait information in the following way:

[0058] Update the user portrait dynamically according to the user's historical behavior data;

[0059] The user portrait information is spliced with the query request through a special character [seq] to form a more comprehensive input sequence.

[0060] In the semantic analysis and multi-intent classification process, the user portrait information is used as a context reference to improve the understanding of the query intent.

[0061] The industry, position, and skill attributes in the user portrait are adjusted in weight to adapt to individual needs in different scenarios.

[0062] Preferably, the system further comprises a multi-label mapping module for label processing of the intent categories in the query request, specifically including:

[0063] Mapping the intent categories in the query request to a predefined label set, which includes but is not limited to "position", "job level", "industry", "product", "professional", "company", "school", "language", and "skill";

[0064] Assigning a unique label number to each intent category and associating it with the output of the multi-intent classification module;

[0065] Supporting the output of multi-label query analysis results to meet the expression needs of multiple intents in user queries;

[0066] Through the mapping relationship between the label number and the classification result, the processing efficiency of the subsequent personalized recommendation module is optimized.

[0067] Preferably, the system further comprises a label priority sorting module for priority sorting of the category labels output by the multi-intent classification module, specifically including:

[0068] Assigning an initial weight to the category labels according to the user portrait information;

[0069] Dynamically adjusting the priority of the category labels according to the user's historical behavior data;

[0070] Sorting the multiple category labels output by the multi-intent classification module in descending order of priority;

[0071] Using the sorted priority labels as input to the personalized recommendation module to optimize the relevance and accuracy of the recommendation results.

[0072] The method for automatically analyzing the query request submitted by the job seeker or the employer as described above includes the following steps:

[0073] Receiving the query request: receiving the query request submitted by the job seeker or the employer, and splicing the query statement with the user portrait information through a special character to form an input sequence;

[0074] Deep semantic parsing: using a deep learning model to perform deep semantic parsing on the input sequence, decomposing the query statement into multiple basic units and generating corresponding embedding vectors;

[0075] Multi-intent classification: inputting the embedding vectors generated by the semantic parsing module into the multi-intent classification module, performing multi-classification on the input, identifying multiple intents in the query request and outputting corresponding category labels;

[0076] Dynamic updating of user portrait: dynamically updating user portrait information based on user historical behavior data to reflect the user's latest needs and preferences;

[0077] Personalized recommendation generation: combining the updated user portrait information and multi-intent classification results to adjust the weights of each category, generating personalized recommendation results that meet the user's needs, and sorting and filtering the results, outputting to the user end.

[0078] Advantages

[0079] Compared with the prior art, the advantages of the present application are:

[0080] For the system involved:

[0081] (1) Deep semantic understanding ability is significantly improved

[0082] By introducing deep learning models (such as BERT) and advanced natural language processing techniques, the present application can more deeply analyze the semantic information of user query requests. By vectorizing the query statement and combining context semantic analysis, the user's true intent can be accurately captured, solving the problem of shallow semantic understanding in traditional methods.

[0083] (2) Multi-intent processing ability is enhanced

[0084] The present application can effectively identify and separate multiple intents in the query by designing an intelligent multi-intent separation algorithm. Using a word segmentation tool and a softmax multi-classification model, complex queries are processed in detail, achieving accurate response to multi-intent queries and overcoming the shortcomings of existing technologies in multi-intent processing ability.

[0085] (3) Personalized matching effect is optimized

[0086] The present application combines user portrait and behavior analysis, dynamically associates the user's query request with their historical behavior and industry preference information, and optimizes the matching strategy through personalized weight adjustment. Compared with traditional rule-based or keyword matching methods, the present application can provide personalized recommendation results that better meet user needs, significantly improving matching accuracy.

[0087] (4) Strong dynamic adaptability

[0088] The application adopts a deep learning model to train data, which can quickly adapt to new demands, new terminologies and industry changes in the recruitment market. Compared with traditional rule engines, the application does not need to be manually updated frequently, and has higher flexibility and dynamic adaptability.

[0089] (5) Information processing efficiency is improved

[0090] The application calculates the interval sum of the token embedding vector by introducing the prefix array (preSum), and confirms the boundary of the word in combination with the word segmentation tool, so as to realize efficient semantic analysis and information extraction. The method can quickly process massive position information and resumes, reduce information noise, and improve system response speed.

[0091] (6) Wide application scenarios

[0092] Through the above technical solutions, the application not only improves the automatic analysis capability of the query request, but also greatly enhances the personalized matching effect, and provides a more efficient and accurate service experience for the recruitment platform and the job seeker.

[0093] At the same time, for the method, the matching accuracy is improved: through dynamic updating of user portrait and multi-intent classification, the demand of the job seeker and the recruitment party can be more accurately matched, so as to improve the relevance and satisfaction of the recommendation. Strengthen semantic understanding: the combination of deep semantic analysis and multi-intent recognition enables the system to effectively understand multiple intents when analyzing complex query requests, thereby improving the intelligent degree of the system. Enhance user experience: the generation process of personalized recommendation results considers the actual needs of users, which can provide more preferred services for users, thereby improving the overall user experience. Strong adaptability: the dynamic adjustment mechanism in the method enables the system to quickly adapt to changes in the recruitment market, timely update the recommendation strategy, and effectively respond to industry changes.

[0094] The application is not only suitable for matching position information and job seeking demand in the recruitment platform, but also can be extended to other fields, such as education and training, e-commerce search, and scenarios requiring multi-intent analysis and personalized recommendation, and has wide practical value. In summary, the application is significantly superior to the prior art in semantic understanding, multi-intent processing, personalization optimization, dynamic adaptability and information processing efficiency, and provides a more efficient, accurate and intelligent service experience for the recruitment platform and the job seeker. BRIEF DESCRIPTION OF DRAWINGS

[0095] Figure 1 is a system structure diagram for automatically analyzing query requests submitted by job seekers or recruitment parties;

[0096] Figure 2 is a system flow architecture diagram of the application.

[0097] Figure 3 Figure for the technical method of the application using "prefix and array (preSum)" to calculate the interval sum of word vectors in query statements;

[0098] Figure 4 Figure for the method flow of the application for automatically analyzing query requests submitted by job seekers or employers. DETAILED DESCRIPTION

[0099] The application will be further described below in conjunction with specific embodiments.

[0100] EMBODIMENT

[0101] The application provides a system for automatically analyzing query requests submitted by job seekers or employers.

[0102] As shown in Figure 1 , the application provides a system for automatically analyzing query requests submitted by job seekers or employers, which includes the following:

[0103] A data input module is configured to receive query requests submitted by job seekers or employers, and concatenate the query statement and user portrait information through a special character [seq] to form an input sequence; a semantic analysis module is configured to perform deep semantic analysis on the input sequence based on a TransformerEncoder model, decompose the query statement into a plurality of basic units token, and generate corresponding embedding vectors; a prefix array processing module is configured to calculate the interval sum of word vectors through a prefix array preSum, including: confirming the boundaries of words according to a word segmentation tool; quickly calculating the word vector of any interval [i, j] through the formula preSum[j]-preSum[i-1], and averaging the vectors in the interval to generate the embedding representation of the complete word; a multi-intent classification module is configured to input the generated embedding vectors of each non-MASK position into a Softmax layer for multi-classification, for identifying multiple intents in the query request, and outputting corresponding category labels; and a personalized matching module is configured to dynamically adjust the weight of the classification result in combination with user portrait information and historical behavior data, optimize the matching strategy, and generate personalized recommendation results that meet the user's needs.

[0104] The query analysis system based on the Transformer Encoder, in combination with the prefix array, the word segmentation tool, and the multi-intent classification module, completes the semantic analysis and classification of the query requests submitted by the job seekers or employers.

[0105] Referring to Figure 2 , the following is a detailed analysis of each part:

[0106] Input layer:

[0107] - The user's input query statement (such as "obstetrician and gynecologist") is decomposed into basic units (tokens), such as "woman", "delivery", "department", "doctor", "life".

[0108] - The query statement and the user portrait information are concatenated through the special character `[seq]` to form an input sequence, which is used to enrich the context information (such as industry, position preference, etc.).

[0109] Transformer Encoder:

[0110] - The input sequence is fed into the Transformer Encoder model to generate the embedding vectors (embeddings) of each token, such as "woman emb", "delivery emb", etc.

[0111] - These embedding vectors retain the semantic information of each token and provide a basis for subsequent processing.

[0112] Prefix array and word segmentation tool:

[0113] - Prefix array (preSum): Used to quickly calculate the interval sum of word vectors:

[0114] - The interval sum formula is:

[0115] $$\text{Interval sum} = \text{preSum}[j] - \text{preSum}[i - 1]$$

[0116] For example, the vector of "obstetrics and gynecology department" is the average of "woman emb + delivery emb + department emb".

[0117] - Word segmentation tool: Confirm the boundary of words. For example, combine "woman", "delivery", "department" into "obstetrics and gynecology department" to generate the embedding representation of the complete word.

[0118] Multi-intent classification module:

[0119] - The embedding vectors at each non-MASK position are fed into the Softmax layer for multi-classification.

[0120] - The classification result outputs labels (such as "position", "industry", "skill", etc.), supporting the simultaneous parsing of multiple intents.

[0121] - The label numbers (such as "position: 0", "industry: 2") are associated with the classification results for standardization processing.

[0122] Output layer:

[0123] - The system finally outputs multiple category labels (Label1, Label2, etc.) to represent different intents in the query, providing the basis for subsequent personalized recommendations.

[0124] Input example:

[0125] The user input query statement is: "obstetrician", and the user profile information is: "industry: medical, position preference: doctor".

[0126] Processing flow:

[0127] Data input module:

[0128] - The query statement and user profile information are spliced to form an input sequence: "obstetrician [seq] medical doctor".

[0129] Semantic analysis module:

[0130] - The Transformer Encoder performs deep semantic analysis on the input sequence to generate embedding vectors for each token:

[0131] …

[0132] obstetrician emb, medical emb, doctor emb, [seq] emb

[0133] …

[0134] Prefix array processing module:

[0135] - Calculate the interval sum of the prefix array:

[0136] …

[0137] obstetrician emb = (obstetrician emb + medical emb + doctor emb) / 3

[0138] doctor emb = (doctor emb + [seq] emb) / 2

[0139] …

[0140] - After the boundary is confirmed by the word segmentation tool, complete word embedding representations are generated: "obstetrician emb" and "doctor emb".

[0141] Multi-intent classification module:

[0142] - Send the word embedding vectors of each non-MASK position to the Softmax layer for multi-classification:

[0143] …

[0144] obstetrician -> Label: industry

[0145] Doctor -> Label: Position

[0146] …

[0147] Personalized Matching Module:

[0148] - Dynamically adjust weights based on user profile information (industry preference: medical; position preference: doctor):

[0149] …

[0150] Industry Weight Promotion -> Preferentially Recommend Medical Related Positions

[0151] Position Weight Promotion -> Preferentially Recommend Doctor Related Positions

[0152] …

[0153] Output Results:

[0154] System Output Analysis Results:

[0155] …

[0156] Label 1: Industry -> Medical

[0157] Label 2: Position -> Doctor

[0158] …

[0159] Recommended Results Example:

[0160] According to the analysis results, recommend positions that meet the conditions for the user, such as "Obstetrics and Gynecology Doctor Position".

[0161] Through the above technical solutions of the invention, the system can efficiently implement the following functions:

[0162] Deep Semantic Understanding: Accurately capture complex semantics in user queries.

[0163] Multi-intent Separation: Support simultaneous recognition of multiple intents to improve parsing ability.

[0164] Personalized Recommendation Optimization: Combine user profile to dynamically adjust matching strategy and improve recommendation accuracy.

[0165] Further, the semantic analysis module implements deep semantic analysis through the following steps:

[0166] Concatenate the user's query statement and user profile information through special character [seq] to form an input sequence;

[0167] The input sequence is vectorized using the Transformer Encoder to generate an embedding vector for each basic unit token. The prefix array preSum is used to calculate the word vector sum of any interval, which includes: after confirming the boundary of the word, the word vector of the interval [i, j] is quickly calculated using the formula preSum[j]-preSum[i-1]; the average value of the vectors in the interval is calculated to generate the embedding representation of the complete word; the context information is combined to capture the dependency between words in the query sentence through the self-attention mechanism, enhancing the context awareness ability of the semantic representation; position encoding is introduced to integrate word order information into word embedding, enabling the model to capture word order structure features. The multi-intent classification module realizes multi-intent recognition through the following steps: input the embedding vector of each non-MAS position generated by the semantic analysis module into the Softmax layer for multi-classification; the Softmax layer outputs the corresponding class label, which includes but is not limited to "position", "industry", "skill", "language", used to identify multiple intents in the query request; separate the identified multiple intents and generate independent analysis results for each intent to realize hierarchical identification of query intent; combine user portrait information to dynamically adjust the weight of the classification result, ensuring that the classification result is highly relevant to user demand.

[0168] Through deep semantic analysis, multi-intent separation and personalized recommendation optimization, efficient and accurate service support is provided for recruitment platforms and job seekers. The following are the core points of the technical solution and its implementation steps:

[0169] Technical core functions

[0170] Deep semantic understanding

[0171] Input sequence construction: concatenate the user query statement and user portrait information through special character [seq] to form a context-rich input sequence.

[0172] Transformer Encoder:

[0173] Vectorize the input sequence to generate an embedding vector for each basic unit (token).

[0174] Introduce a self-attention mechanism to capture the dependency between words in the query sentence and enhance context awareness.

[0175] Integrate position encoding to enable the model to capture word order structure features.

[0176] Prefix array calculation:

[0177] Calculate the word vector sum of any interval using the formula

[0178] Calculate the average value of the vectors in the interval to generate the embedding representation of the complete word.

[0179] Multi-intent separation

[0180] Softmax classification:

[0181] Input the embedding vector of each non-MASK position into the Softmax layer for multi-classification.

[0182] Output category labels (such as "position", "industry", "skills", etc.), supporting simultaneous recognition of multiple intents.

[0183] Hierarchical recognition:

[0184] Separate the recognized multiple intents and generate independent analysis results for each intent.

[0185] Dynamic weight adjustment:

[0186] Combine user portrait information to dynamically adjust the classification results, ensuring that the classification results are highly relevant to user needs.

[0187] Personalized recommendation optimization:

[0188] According to the analysis results, provide users with recommended content (such as positions or talents) that meet the conditions.

[0189] Adjust the recommendation strategy dynamically, prioritize matching content with higher relevance to user portrait information, and improve recommendation accuracy.

[0190] Technical implementation steps:

[0191] Semantic analysis module:

[0192] Concatenate the query statement and user portrait information to form an input sequence.

[0193] Use Transformer Encoder to generate token embedding vectors and quickly calculate the embedding representation of complete words through prefix arrays.

[0194] Self-attention mechanism and position encoding enhance context awareness and word order feature capture ability.

[0195] Multi-intent classification module:

[0196] Input the embedding vector into the Softmax layer for multi-classification, output category labels (such as "position", "industry", etc.).

[0197] Separate multiple intents and generate independent analysis results to achieve hierarchical recognition.

[0198] Personalized matching module:

[0199] Dynamic adjustment of classification result weights combined with user portrait and historical behavior data.

[0200] Optimized recommendation strategy to provide precise and highly relevant recommended content for users.

[0201] Technical advantages:

[0202] Efficient semantic analysis:

[0203] Based on Transformer Encoder and prefix array calculation, fast and efficient deep semantic understanding is achieved.

[0204] Multi-intent processing capability:

[0205] Supports multi-intent separation in complex queries to meet users' diverse needs.

[0206] Personalized recommendation optimization:

[0207] Combines user portrait information and behavior data dynamically to improve matching accuracy and service efficiency.

[0208] Strong adaptability:

[0209] Can quickly adapt to query analysis needs in different fields (such as recruitment, education, e-commerce, etc.).

[0210] Recommendation result example:

[0211] User input query: "obstetrician", combined with user portrait information (industry: medical; position preference: doctor).

[0212] System analysis output:

[0213] text

[0214] Label 1: Industry -> Medical

[0215] Label 2: Position -> Doctor

[0216] Recommended result: Recommend suitable positions for users, such as "obstetrician position".

[0217] The invention solves the problems of shallow semantic understanding, weak multi-intent processing capability, and insufficient personalized matching in traditional recruitment systems through deep learning technology. The system can efficiently analyze complex query requests, accurately identify multiple intents, and dynamically optimize the recommendation strategy based on user portraits, providing efficient, precise, and intelligent service experience for recruitment platforms and job seekers.

[0218] The personalized matching module realizes personalized recommendation through the following steps: obtaining the query request and user portrait information of the user, including but not limited to the industry to which the user belongs, position preference, skill specialty; dynamically adjusting the weight of each category according to the category label output by the multi-intent classification module, and preferentially matching the category with higher correlation with the user portrait information; using the user historical behavior data, including browsing records and click habits, to build a user interest model, and further optimizing the recommendation strategy; combining the adjusted category weight and the user interest model to generate personalized recommendation results that meet the user's needs; sorting and filtering the generated personalized recommendation results, and selecting the highest confidence results to output to the user end. The prefix array processing module realizes efficient calculation through the following steps: constructing a prefix array preSum: preSum[j] represents the cumulative vector sum from the start of the sequence to position j; quickly calculating the cumulative sum of any interval [i, j]: using the formula preSum[j]-preSum[i-1] to calculate the cumulative sum of all token embedding vectors in the interval [i, j] in O(1) time; generating the embedding representation of the word: taking the average of the cumulative vector in the interval [i, j] to obtain the embedding representation of the corresponding complete word; combining the word segmentation tool to confirm the word boundary: accurately positioning the start and end positions of each word through word segmentation to ensure that the interval calculation result accurately corresponds to the semantic information of the word. The system further includes a user portrait enhancement module for optimizing user portrait information in the following ways: dynamically updating the user portrait according to the user's historical behavior data; concatenating the user portrait information and the query request through a special character [seq] to form a more comprehensive input sequence; in the semantic analysis and multi-intent classification process, the user portrait information is used as a context reference to improve the understanding ability of the query intent; adjusting the weights of the industry, position, and skill attributes in the user portrait to adapt to the personalized needs in different scenarios. The system further includes a multi-label mapping module for label processing of the intent categories in the query request, specifically including: mapping the intent categories in the query request with a pre-defined label set, which includes but is not limited to "position", "job level", "industry", "product", "professional", "company", "school", "language", and "skill"; assigning a unique label number to each intent category and associating it with the output result of the multi-intent classification module; supporting the output of multi-label query analysis results to meet the expression needs of multiple intents in user queries; through the mapping relationship between the label number and the classification result, the processing efficiency of the subsequent personalized recommendation module is optimized.The system further comprises a label prioritization module for prioritizing the category labels output by the multi-intent classification module, specifically including: assigning an initial weight to the category labels according to the user portrait information; dynamically adjusting the priority of the category labels according to the historical behavior data of the user; sorting the multiple category labels output by the multi-intent classification module in descending order of priority; and taking the sorted priority labels as the input of the personalized recommendation module to optimize the relevance and accuracy of the recommendation results. Figure 2 As shown, a query analysis system based on Transformer Encoder is combined with a prefix array and a word segmentation tool for deep semantic analysis and multi-intent classification. The specific meanings are as follows:

[0219] Input layer (Transformer Encoder)

[0220] The input is the query sentence submitted by the user, for example, "gynecologist".

[0221] The query is split into basic units (tokens), such as "woman", "production", "department", "doctor", and "person", and the corresponding embedding vectors (embedding) are generated through the Transformer Encoder, such as "woman emb", "production emb", etc.

[0222] Word segmentation tool and prefix array

[0223] The word segmentation tool is used to identify the boundaries of words, such as combining "woman", "production", and "department" into "woman production department" and combining "doctor" and "person" into "doctor".

[0224] The prefix array (preSum) is used to quickly calculate the interval sum of the word vector:

[0225] For example, the vector of "woman production department" is the average of "woman emb + production emb + department emb".

[0226] Through preSum, the word vector of any interval can be efficiently obtained, avoiding repeated calculation.

[0227] Intermediate layer (embedding and sequence processing)

[0228] After word segmentation and prefix array processing, the embedding representation of each complete word is generated, such as "woman production department emb" and "doctor emb".

[0229] Special character [seq] is used to concatenate user query and user portrait information (such as industry, position, etc.), enriching the context information.

[0230] Output layer (Softmax classification)

[0231] The embedding vector of each non-MASK position is multi-classified by a Softmax layer, outputting the corresponding class label (Label1, Label2, etc.).

[0232] The label corresponds to the query intent, such as "position", "skill", etc., supporting multi-intent simultaneous parsing.

[0233] User portrait association

[0234] User portrait data (such as position, industry, skill, etc.) is spliced into the query through special characters, further enhancing personalized parsing capabilities.

[0235] As shown in Figure 3 , the figure demonstrates the technical method of calculating the interval sum of word vectors in the query statement using prefix and array (preSum), with the following specific meanings:

[0236] Definition and function of prefix sum array:

[0237] - preSum[j] and preSum[i-1] are marked in the figure, representing the cumulative vector sum from the start of the sequence to position j and position i-1, respectively.

[0238] - Through the calculation of prefix sum, the vector sum of any interval [i, j] can be quickly obtained, with the formula:

[0239] Interval sum = preSum[j] - preSum[i-1]

[0240] This method avoids repeated calculations and improves processing efficiency.

[0241] Generation of word vectors and boundary confirmation:

[0242] - Each word in the query is generated into a corresponding embedding vector (token embedding) through a deep learning model (such as BERT).

[0243] - The blue part in the figure represents the previous word, and the red part represents the current processing interval [i, j].

[0244] - The word segmentation tool is used to identify the boundaries of each word, ensuring that the vector sum within the interval results in a complete word vector.

[0245] Average calculation of interval vector:

[0246] - For interval [i, j], take the average of the sum of all token vectors to generate the word vector corresponding to the interval. For example, the overall embedding representation of "obstetrician" can be calculated through the interval sum in the figure.

[0247] Multi-intent classification:

[0248] - Each non-MASK position (i.e., valid position) vector will be multi-classified by a softmax layer to identify multiple intents in the query.

[0249] - This method can effectively separate different intents in complex queries and improve parsing ability.

[0250] The figure shows a method for quickly calculating word vectors in query statements through prefix and array, and realizes efficient and multi-intent parsing by combining tokenization tools and soft max classification. This solution can improve semantic understanding ability and optimize information processing efficiency.

[0251] As shown in Figure 4 The method for automatically parsing query requests submitted by job seekers or employers as described above includes the following steps: receiving a query request: receiving a query request submitted by a job seeker or employer, and concatenating the query statement and user portrait information through special characters to form an input sequence; deep semantic parsing: using a deep learning model to perform deep semantic parsing on the input sequence, decomposing the query statement into multiple basic units, and generating corresponding embedding vectors; multi-intent classification: input the embedding vectors generated by the semantic parsing module into the multi-intent classification module, perform multi-classification on the input, identify multiple intents in the query request and output corresponding category labels; dynamic updating of user portrait: dynamically updating user portrait information based on user historical behavior data to reflect the latest needs and preferences of users; personalized recommendation generation: combining updated user portrait information and multi-intent classification results, adjusting the weights of each category, generating personalized recommendation results that meet user needs, and sorting and filtering the results, output to the user end.

[0252] The actual case is interpreted as follows:

[0253] Case interpretation

[0254] Input example:

[0255] The user input query statement is: "obstetrician", and the user portrait information is: "industry: medical, job preference: doctor".

[0256] Processing flow:

[0257] The data input module concatenates the query statement and the user portrait to form an input sequence: "obstetrician [seq] medical doctor".

[0258] The semantic parsing module uses Transformer Encoder to generate embedding vectors for each token, and calculates the embedding representation of the complete word through the prefix array, such as "obstetrician emb" and "doctor emb".

[0259] The multi-intent classification module identifies two category labels: "Industry: Medical" and "Position: Doctor".

[0260] The personalized matching module dynamically adjusts the weight according to the user portrait, and preferentially recommends the doctor position in the medical related position.

[0261] Output result:

[0262] The system finally outputs the analysis result:

[0263] Label 1: Industry -> Medical

[0264] Label 2: Position -> Doctor

[0265] Recommended result example: Recommend the user with the qualified position such as "Obstetrics and Gynecology Doctor Position".

[0266] This method realizes efficient analysis and accurate matching of complex query requests through deep semantic analysis, multi-intent classification and personalized recommendation optimization, and provides intelligent service support for recruitment platforms and job seekers.

[0267] The above content is a further detailed description of the present application combined with specific embodiments, and cannot be considered as limiting the specific implementation of the present application to these descriptions. For ordinary technical personnel in the data processing technology field to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be considered as belonging to the protection scope determined by the claims submitted by the present application.

Claims

1. A system for automatically parsing query requests submitted by job seekers or recruiters, characterized in that: The data input module is used to receive query requests submitted by job seekers or recruiters, and to concatenate the query statement with user profile information using the special character [seq] to form an input sequence. The user profile information includes the user's industry, job preferences, and skills. In the process of semantic parsing and multi-intent classification, the user profile information is used as a contextual reference to improve the ability to understand query intent; The semantic parsing module performs deep semantic parsing on the input sequence based on the Transformer Encoder model, decomposing it into several basic unit tokens and generating corresponding embedding vectors. The prefix array processing module calculates the interval sum of word vectors using the prefix array `preSum`, including: Use word segmentation tools to identify word boundaries; The word vectors for any interval [i,j] are quickly calculated using the formula preSum[j]−preSum[i−1], and the average value of the vectors within the interval is calculated to generate the embedding representation of the complete word. The multi-intent classification module inputs the generated embedding vector of each non-MASK location into the Softmax layer for multi-classification, which is used to identify multiple intents in the query request and output the corresponding category label; The personalized matching module combines user profile information and historical behavior data to dynamically adjust the weight of classification results, optimize matching strategies, and generate personalized recommendation results that meet user needs. Among them, the user profile is dynamically updated based on the user's historical behavior data, and the weight of each category is adjusted by combining the updated user profile information and multi-intent classification results to ensure that the classification results are highly relevant to user needs. By leveraging users' historical behavior data, including browsing history and click habits, we can build user interest models and further optimize recommendation strategies. By combining the adjusted category weights and user interest models, personalized recommendation results that meet user needs are generated. The generated personalized recommendation results are sorted and filtered, and the results with the highest confidence are selected and output to the user.

2. The system for automatically parsing query requests submitted by job seekers or recruiters according to claim 1, characterized in that: The semantic parsing module implements deep semantic parsing through the following steps: The user's query and user profile information are concatenated using the special character [seq] to form the input sequence; The input sequence is vectorized using the Transformer Encoder to generate an embedding vector for each basic unit token. The word vector sum for any interval is calculated using the prefix array `preSum`, specifically including: After confirming the word boundaries, the word vectors for the interval [i,j] are quickly calculated using the formula preSum[j]−preSum[i−1]. The average value of the vectors within the interval is used to generate the embedding representation of the complete word; By combining contextual information, the dependency relationships between words in the query statement are captured through a self-attention mechanism, thereby enhancing the context-awareness of semantic representation; By introducing positional encoding and incorporating word order information into word embeddings, the model is able to capture word order structure features.

3. The system for automatically parsing query requests submitted by job seekers or recruiters according to claim 1, characterized in that: The multi-intent classification module achieves multi-intent recognition through the following steps: The embedding vector for each non-MAS location generated by the semantic parsing module is input into the Softmax layer for multi-class classification. The Softmax layer outputs corresponding category labels, which include "job title", "industry", "skills" and "language", used to identify multiple intents in the query request; The system separates the identified multiple intents and generates independent parsing results for each intent, thereby achieving hierarchical identification of query intents.

4. The system for automatically parsing query requests submitted by job seekers or recruiters according to claim 1, characterized in that: The prefix array processing module achieves efficient calculation through the following steps: Construct a prefix array preSum: preSum[j] represents the cumulative vector sum from the start of the sequence to position j; Fast calculation of the cumulative sum of any interval [i,j]: Using the formula preSum[j] - preSum[i-1], the cumulative sum of all token embedding vectors in the interval [i,j] is calculated in O(1) time; Word embedding representation: The average value of the cumulative vector within the interval [i,j] is used to obtain the embedding representation of the corresponding complete word; Combine word segmentation tools to confirm word boundaries: accurately locate the start and end positions of each word through word segmentation to ensure that the interval calculation results correspond precisely to the semantic information of the words.

5. The system for automatically parsing query requests submitted by job seekers or recruiters according to claim 1, characterized in that: The system further includes a multi-tag mapping module for tagging the intent categories in query requests, specifically including: Map the intent category in the query request to a predefined set of tags, including "job title", "job level", "industry", "product", "major", "company", "school", "language", and "skill". Assign a unique label number to each intent category and associate it with the output of the multi-intent classification module; It supports query parsing results with multiple tags to meet the user's need to express multiple intents in the query; By mapping the tag numbers to the classification results, the processing efficiency of the subsequent personalized recommendation module can be optimized.

6. The method implemented in a system for automatically parsing query requests submitted by job seekers or recruiters as described in any one of claims 1-5, characterized in that, Includes the following steps: Receive query requests: Receive query requests submitted by job seekers or recruiters, and concatenate the query statement with user profile information using special characters to form an input sequence; Deep semantic parsing: Utilizes deep learning models to perform deep semantic parsing on input sequences, decomposing query statements into multiple basic units and generating corresponding embedding vectors; Processing the prefix array: Calculate the interval sum of word vectors using the prefix array `preSum`, including: Use word segmentation tools to identify word boundaries; The word vectors for any interval [i,j] are quickly calculated using the formula preSum[j]−preSum[i−1], and the average value of the vectors within the interval is calculated to generate the embedding representation of the complete word. Multi-intent classification: The generated embedding vector of each non-MASK location is input into the Softmax layer for multi-class classification to identify multiple intents in the query request and output the corresponding category label; Personalized matching steps: Combining user profile information and historical behavior data, the weight of classification results is dynamically adjusted to optimize the matching strategy and generate personalized recommendation results that meet user needs.

Images