Real-time job intelligent recommendation system based on instant messaging context

Abstract

The application relates to the technical field of recruitment, and discloses a real-time position intelligent recommendation system based on an instant communication context, which comprises an instant communication access module, a real-time context processing engine, a multi-dimensional semantic understanding module, a dynamic position matching engine, a real-time recommendation pushing module, a data storage module, a permission management module, a business opportunity conversion module and a compliance processing module, and each module cooperatively works; the application solves the context split problem, realizes seamless experience of "communication is data", through efficient communication architecture of the instant communication access module and the back end, real-time capture of bidirectional chat messages of HR and candidates, customers, breaking of the split state of the communication scene and the recommendation system in the prior art, inclusion of each interaction into the recommendation consideration, complete capture of dynamic demands and implicit demands in the communication process, and avoidance of demand information omission.

Description

Technical Field

[0001] This invention belongs to the field of recruitment technology, specifically a real-time intelligent job recommendation system based on instant messaging context. Background Technology

[0002] Currently, job recommendation technologies in the recruitment industry are mainly divided into three categories: traditional search matching, static matching based on resumes, and one-way job push. Traditional search matching relies on HR to manually input keywords, which is inefficient and the matching accuracy depends on the quality of the keywords. Static matching based on resumes can only extract information and recommend jobs once after the job seeker uploads their resume, lacking dynamic interactive capabilities. One-way job push only pushes based on the user's historical behavior and cannot be combined with real-time communication scenarios.

[0003] Some existing technologies attempt to combine instant messaging tools with recruitment recommendations. For example, CN119646301A discloses a group and job recommendation method based on chat tools, which focuses on group matching rather than job recommendation in real-time communication scenarios; CN121258453A discloses a dynamic job recommendation method based on a large language model of dialogue context, which can only parse the candidate's active input and cannot capture the dynamic needs in two-way real-time communication between HR and candidates; CN121504405A discloses a recruitment recommendation system based on AI intelligent agents and instant messaging tools, which focuses on intent classification and lacks the ability to capture implicit needs in the communication process; and Beijing BitEasy's CN118780772A discloses an automated recruitment service system based on WeChat for Enterprise, which is only used for interview process management and does not involve real-time job recommendation.

[0004] The aforementioned existing technologies generally suffer from the following shortcomings: First, they are context-segmented, failing to capture real-time needs information during communication between HR and candidates / clients, especially dynamic and implicit needs in two-way interactions; second, recommendations are delayed, unable to achieve real-time "chat-and-push" responses, with high recommendation generation latency, failing to meet the immediacy needs of real-time communication scenarios; third, matching dimensions are limited, relying solely on resume text or explicit dialogue content for matching, lacking the ability to understand implicit needs; fourth, business opportunities are lost, unable to convert recruitment needs generated during HR-client communication into job recommendation tasks in real time, with response delays leading to competitors acquiring high-quality candidates first; and fifth, the integration with the instant messaging ecosystem is insufficient, failing to fully utilize the organizational structure, permission system, and message push mechanism of WeChat Work, thus failing to achieve a seamless "communication as a service" experience.

[0005] Based on this, a real-time intelligent job recommendation system based on instant messaging context is designed. Summary of the Invention

[0006] In view of the above situation and to overcome the shortcomings of the prior art, the present invention provides a real-time intelligent job recommendation system based on instant messaging context, which effectively solves the problems raised in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a real-time intelligent job recommendation system based on instant messaging context, comprising: The instant messaging access module obtains session messages based on the enterprise-level communication platform's session content archiving interface, application callback interface, or authorized message synchronization interface. The instant messaging access module is configured with authorized employee scope, customer / candidate notification and consent mechanism, message retrieval cursor, breakpoint resume, decryption process, message compensation mechanism, and data minimization processing module. It is used to establish a persistent connection with the backend, obtain chat messages and message metadata between HR and candidates / customers, and transmit the obtained message data to the backend in real time. The real-time context processing engine communicates with the instant messaging access module and includes a message queue and a stream processing unit. The message queue is used to receive and buffer message data transmitted by the instant messaging access module, and the stream processing unit is used to process the message data in the message queue in real time to complete message parsing and preliminary classification. The multi-dimensional semantic understanding module communicates with the real-time context processing engine and is used to perform entity extraction, confidence calculation, urgency scoring and implicit skill mapping based on the message data output by the real-time context processing engine. It extracts structured explicit and implicit demand information and synthesizes explicit demand, implicit demand and historical profile into a dynamic demand vector, wherein the dynamic demand vector V=[skill tag vector, location vector, salary range, years of experience, education, urgency, emotional intensity, job category, customer priority]; The dynamic job matching engine communicates with the multi-dimensional semantic understanding module and includes a pre-calculation unit, an incremental update unit, and a matching retrieval unit. The pre-calculation unit is used to pre-calculate the candidate job matching matrix based on candidate profile data and job database data. The incremental update unit is used to trigger incremental retrieval when the change in the dynamic demand vector exceeds a threshold. It only updates the matching data corresponding to the changed candidate profile or job information and does not perform a full recalculation. The matching retrieval unit is used to perform weighted fusion of the pre-calculated candidate job matrix and the dynamic demand vector to generate a job recommendation list. The real-time recommendation push module communicates with the dynamic job matching engine and the instant messaging access module respectively. It is used to push the job recommendation list to the corresponding enterprise-level communication platform chat window in the form of message cards through the enterprise-level communication platform message interface. It also provides one-click access to view resumes and one-click initiation of communication, thereby reducing recommendation response latency, generating recommendation results under cache hit conditions, and returning the candidate list within a preset time threshold. The data storage module communicates with the real-time context processing engine, the multi-dimensional semantic understanding module, and the dynamic job matching engine respectively. It is used to store message data, candidate profile data, job database data, candidate job matching matrix and system operation logs. It is configured with sensitive field desensitization, data retention period, deletion / revocation authorization mechanism and log auditing module. The access control module communicates with the instant messaging access module and is used to implement hierarchical control of system operation permissions based on the organizational structure of the enterprise-level communication platform. It ensures that different roles can only access data and functions within their corresponding permission scope and is configured with data visibility range and abnormal access alarm modules for different roles. The opportunity conversion module communicates with the real-time context processing engine, the multi-dimensional semantic understanding module, and the dynamic job matching engine. It uses a multi-factor model to identify recruitment opportunities in chat messages, automatically initiates requirement clarification, generates job drafts, and triggers an instant recommendation process after HR confirmation. The compliance processing module communicates with the instant messaging access module, data storage module, and access control module to achieve compliance control over personal information processing, including functions such as pre-processing notification, authorization scope management, candidate and customer consent recording, explanatory reasons for recommendation, and automated decision-making with manual review.

[0008] Preferably, the instant messaging access module establishes communication with the backend using a WebSocket long connection, and the obtained message metadata includes sender identity information, receiver identity information, message sending timestamp, session ID, and message type; The message types include text messages, voice messages, and file messages. Voice messages are transmitted together with text messages after being processed into text, while file messages only capture the file name and type information.

[0009] Preferably, the real-time context processing engine uses a Kafka message queue, and the stream processing unit uses the Flink stream processing framework. The processing steps of the stream processing unit include message deduplication, message format standardization, and message classification. The message classification includes two categories: dialogue between HR and candidates, and dialogue between HR and customers.

[0010] Preferably, the explicit requirement extraction of the multi-dimensional semantic understanding module adopts the Named Entity Recognition (NER) model, and the extracted structured explicit requirement information includes skills, work location, salary range, years of work experience and educational requirements; The implicit requirement extraction includes urgency identification, sentiment analysis, and implicit skill inference. Urgency identification is based on message text keywords, message sending time, and message sending frequency, and outputs an urgency score in the range of 0-1 through weighted calculation. Sentiment analysis is based on a sentiment analysis model and outputs sentiment polarity and sentiment intensity. Sentiment polarity includes positive and negative. Implicit skill inference is based on a contextual reasoning model, which infers specific skill requirements from fuzzy descriptions and maps them to skill tags.

[0011] Preferably, the input features of the multi-dimensional semantic understanding module include standardized message text, message metadata, historical session data, and candidate / customer basic profiles. The output format is structured explicit requirement data, implicit requirement feature values, and dynamic requirement vector. Explicit and implicit requirements are synthesized into a dynamic requirement vector through weighted fusion, and the weights are dynamically adjusted based on the business scenario.

[0012] Preferably, the pre-computation unit of the dynamic job matching engine stores the candidate job matching matrix in a cache index; the recommendation score S of the matching retrieval unit is S = α × candidate - job basic matching score + β × similarity between demand vector and job vector + γ × urgency weight + δ × conversation feedback weight - ε × permission / visibility penalty, where α, β, γ, δ, and ε are weight coefficients that can be dynamically adjusted according to the business scenario. The generated job recommendation list is sorted from high to low matching degree. Each recommended job includes job name, core requirements, salary range, and matching degree score. The conversation feedback weight is updated in real time based on user click / ignore feedback.

[0013] Preferably, the multi-factor model of the business opportunity conversion module includes recruitment keywords, dialogue roles, customer historical recruitment behavior, job entity density in the conversation, completeness of fields such as number of recruits / location / salary, message frequency, urgency score, whether there are already similar positions, and whether the demand clarification state machine is triggered. Among them, recruitment keywords are subordinate factors, including "recruiting", "recruiting", "job position", "urgent recruitment", and "searching for people". The requirement clarification process involves the system automatically sending preset clarification questions to the message sender. These preset clarification questions include the job title, number of openings, urgency level, core requirements, and work location.

[0014] Preferably, the data storage module uses a combination of relational and non-relational databases. The relational database stores job database data, candidate basic information, and system permission data, while the non-relational database stores chat message data, candidate profile data, and candidate job matching matrix. The cache index uses Redis for caching, and the matching retrieval unit is also configured with a vector retrieval unit and a fallback retrieval unit. The vector retrieval unit is implemented using Elasticsearch, and the fallback retrieval unit is used to switch to real-time computing mode when the cache fails.

[0015] Preferably, it also includes a fault degradation module, which communicates with the real-time context processing engine and the dynamic job matching engine. When the WebSocket long connection fails, it automatically switches to the HTTP long polling mechanism; when the stream processing unit fails, it automatically switches to the batch processing mode; and when the cache index fails, it automatically switches to the vector retrieval unit real-time calculation mode.

[0016] Preferably, the message card pushed by the real-time recommendation push module includes job title, company name, core skill requirements, salary range, matching degree, and operation buttons. The operation buttons include "View Resume", "Initiate Communication", and "Save Job". Clicking the operation button can directly jump to the corresponding function page without switching applications.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention solves the problem of context fragmentation and achieves a seamless experience of "communication as data". Through the efficient communication architecture between the instant messaging access module and the backend, it captures two-way chat messages between HR and candidates / customers in real time, breaking the fragmentation between communication scenarios and recommendation systems in the existing technology. It incorporates every interaction into recommendation considerations, fully captures dynamic and implicit needs in the communication process, and avoids the omission of demand information. 2. This invention achieves real-time recommendation response and reduces recommendation latency. It adopts the core logic of "pre-calculation + incremental update". The candidate job matching matrix is ​​pre-calculated and cached. Incremental updates are only performed when the requirements or data change, without the need for full recalculation, which greatly reduces the computational pressure on the system. Under the condition of cache hit, recommendation results can be generated quickly, realizing "chatting and pushing", which meets the immediacy requirements of real-time communication scenarios. 3. This invention improves matching accuracy by simultaneously extracting explicit and implicit requirements through a multi-dimensional semantic understanding module, generating a dynamic requirement vector through weighted fusion, and combining it with a weighted fusion matching algorithm to clarify the matching logic and weight coefficients, making the recommendation process interpretable. At the same time, it optimizes the recommendation weight in real time through conversation feedback, improving the accuracy of job matching and avoiding the bias caused by single-dimensional matching. 4. This invention enables real-time conversion of recruitment opportunities, reducing opportunity loss. It uses a multi-factor model to identify recruitment opportunities and constructs a closed-loop process of "opportunity identification → demand clarification → job creation → instant recommendation". Compared with the existing simple keyword matching method for opportunity identification, it has higher identification accuracy and stronger creativity. It can quickly transform recruitment needs generated in the communication between HR and customers into recommendation tasks, preventing high-quality candidates from being preempted by competitors. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0019] In the attached diagram: Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart illustrating the overall workflow of the system of the present invention; Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] Depend on Figures 1-2 The present invention relates to a real-time intelligent job recommendation system based on instant messaging context, comprising an instant messaging access module, a real-time context processing engine, a multi-dimensional semantic understanding module, a dynamic job matching engine, a real-time recommendation push module, a data storage module, a permission management module, a business opportunity conversion module, and a compliance processing module, wherein the modules work together.

[0022] The instant messaging access module obtains session messages based on the session content archiving interface, application callback interface, or authorized message synchronization interface of enterprise-level communication platforms (such as WeChat Work, Lark, DingTalk). This module is configured with authorized employee scope, customer / candidate notification and consent mechanisms, message retrieval cursors, breakpoint resumption, decryption processes, message compensation mechanisms, and data minimization processing modules. Its core function is to establish a WebSocket long connection with the backend server to achieve persistent communication and obtain chat messages and message metadata between HR and candidates / customers. The message metadata includes sender identity information (enterprise-level communication platform account, department), receiver identity information, message sending timestamp, session ID, and message type (text, voice, file). For voice messages, the module has a built-in voice-to-text plugin that converts voice messages into text messages in real time before transmitting them along with the original text messages. For file messages, only the file name and type information are captured; the file's internal content is not read to ensure data security. All obtained message data is transmitted in real-time to the backend real-time context processing engine via the WebSocket long connection to ensure the stability and compliance of data transmission.

[0023] The real-time context processing engine is the core data processing hub of the system, communicating with the instant messaging access module. It comprises two parts: a message queue and a stream processing unit. The message queue receives and buffers message data transmitted from the instant messaging access module, enabling asynchronous message transmission and buffering to prevent system lag due to instantaneous message spikes, while ensuring no message loss or duplication. The stream processing unit consumes message data from the message queue in real time and performs real-time processing. The processing includes three steps: first, message deduplication, removing duplicate messages based on session ID and message timestamp; second, message format standardization, converting different types of messages (text, speech converted to text) into a standard text format for easier subsequent processing; and third, message classification, categorizing messages into two types based on the sender and receiver: HR-candidate dialogue and HR-customer dialogue, which are then tagged and transmitted to the multi-dimensional semantic understanding module. Preferably, the message queue uses Kafka, and the stream processing unit uses the Flink stream processing framework.

[0024] The multi-dimensional semantic understanding module communicates with the real-time context processing engine. Its core function is to parse the requirement information in message data, including explicit and implicit requirements. This addresses the problems of insufficient algorithmic detail and "black box AI" in existing technologies. Specifically, it includes three core processes: explicit requirement extraction, implicit requirement extraction, and requirement modeling. The explicit requirement extraction adopts the Named Entity Recognition (NER) model. The input features include standardized message text, message metadata, historical session data, and candidate / customer basic profiles. The standardized message text is parsed to extract the structured explicit requirement information, which specifically includes skills (such as Java, Python), work location (such as Beijing, Shanghai), salary range (such as 8k-12k, 15k-20k), years of work experience (such as 1-3 years, 3-5 years), and education requirements (such as bachelor's degree, master's degree). The output format is structured explicit requirement data. Implicit requirement extraction comprises three core processes: urgency identification, sentiment analysis, and implicit skill inference. Urgency identification is based on keywords in the message text (e.g., "urgently hiring," "starting this week," "starting as soon as possible"), message sending time (sending outside of working hours suggests urgency), and message sending frequency (high-frequency, continuous sending suggests urgency). An urgency score in the 0-1 range is output through weighted calculation; a higher score indicates a more urgent need. The urgency score formula is achieved through weighted summation, with weight coefficients calibrated according to the business scenario. Sentiment analysis, based on a pre-trained sentiment analysis model, identifies tone preferences in conversations between HR and candidates / clients (e.g., "preferred," "not very suitable," "hope to start as soon as possible"), outputting sentiment polarity (positive or negative) and sentiment intensity. Sentiment intensity is used to adjust the priority of subsequent job matching. Implicit skill inference, based on a contextual reasoning model, infers specific skill requirements from vague descriptions (e.g., "has similar project experience," "has done related system development") and maps them to standard skill tags. For example, "has done e-commerce platform development" infers the need for Java development and database design skills. Demand modeling involves weighted fusion of structured explicit demand information extracted from explicit demands and implicit demand information (urgency score, emotional polarity, and implicit skills) extracted from implicit demands to construct a dynamic demand vector. This vector, V = [skill tag vector, location vector, salary range, years of experience, education level, urgency, emotional intensity, job category, and customer priority], is used for subsequent matching calculations with job information. The fusion weights of explicit and implicit demands are dynamically adjusted based on the business scenario.

[0025] The dynamic job matching engine, communicating with the multi-dimensional semantic understanding module, is the core module for achieving accurate job recommendations. It employs a "pre-computation + incremental update + weighted fusion" logic to address the issues of recommendation lag and unclear matching logic in existing technologies. It includes a pre-computation unit, an incremental update unit, and a matching retrieval unit. The pre-computation unit pre-calculates the candidate-job matching matrix based on the candidate profile data and job database data in the data storage module. This matrix stores the initial matching degree between all candidates and all jobs. The matching degree calculation is based on the candidate's basic profile (resume information) and the job's basic requirements (job description). After the pre-computation is completed, the matching matrix is ​​stored in the cache index to ensure fast retrieval. As a preferred implementation, the cache index uses Redis caching. The incremental update unit monitors changes in the dynamic demand vector in real time. When the change in the dynamic demand vector exceeds a preset threshold, incremental retrieval is triggered. At the same time, it monitors changes in candidate profile data and job database data. When information such as a candidate's skills and salary expectations changes, or information such as job requirements and salary changes, only the matching data corresponding to the change is incrementally updated, rather than being fully recalculated. This greatly reduces the computational pressure on the system and ensures the real-time performance of the matching matrix. The matching and retrieval unit receives a dynamic demand vector output by the multi-dimensional semantic understanding module. Based on this vector, it retrieves the candidate job matching matrix in the cached index and calculates a recommendation score using a weighted fusion algorithm. The recommendation score S = α × candidate - job base matching score + β × similarity between demand vector and job vector + γ × urgency weight + δ × session feedback weight - ε × permission / visibility penalty, where α, β, γ, δ, and ε are weight coefficients that can be dynamically adjusted according to the business scenario. The top-K jobs with the highest matching degree (K can be set by HR, with a default value of 5) are selected to generate a job recommendation list, which is sorted from high to low matching degree. Each recommended job posting includes the job title, company name, core requirements, salary range, and matching score. Simultaneously, the session feedback weight is updated in real-time based on user clicks / ignores, optimizing subsequent recommendation accuracy. This module reduces recommendation response latency, generates recommendation results under cache hit conditions, and returns a candidate list within a preset time threshold, avoiding the problem of insufficient experimental data to support absolute performance indicators. As a preferred implementation, the matching retrieval unit is also configured with a vector retrieval unit and a fallback retrieval unit. The vector retrieval unit is implemented using Elasticsearch, and the fallback retrieval unit is used to switch to real-time computing mode in case of cache failure.

[0026] The real-time recommendation push module communicates with both the dynamic job matching engine and the instant messaging access module. Its core function is to instantly push the job recommendation list generated by the dynamic job matching engine to the corresponding enterprise-level communication platform chat window. The push method uses message cards from the enterprise-level communication platform. Each message card includes the job title, company name, core skill requirements, salary range, match rate, and three operation buttons ("View Resume," "Initiate Communication," and "Save Job"). HR can directly jump to the candidate's resume page by clicking "View Resume," send a message directly to the candidate by clicking "Initiate Communication," and add the job to their favorites by clicking "Save Job," achieving a seamless "communication-as-recommendation" experience without switching applications. Push timing is divided into two types: first, when HR communicates with a candidate, an updated job recommendation list is instantly pushed after each valid demand message is captured; second, when HR communicates with a client, a matched candidate job recommendation list is instantly pushed after a business opportunity is converted.

[0027] The data storage module communicates with the real-time context processing engine, the multi-dimensional semantic understanding module, and the dynamic job matching engine, respectively. It uses a combination of relational databases (such as MySQL) and non-relational databases (such as MongoDB) to store all data during system operation, while ensuring data security and privacy compliance. Relational databases are used to store job database data (job title, company information, core requirements, salary range, work location, etc.), candidate basic information (name, contact information, education, years of work experience, etc.), and system permission data (user roles, operation permissions, etc.), ensuring structured storage and efficient querying of data; Non-relational databases (such as MongoDB) are used to store chat message data, candidate profile data (basic information + real-time contextual requirements information), and system operation logs; cached indexes (such as Redis) are used to store the candidate job matching matrix to ensure fast response of matching retrieval; All data is stored in encrypted form and backed up regularly. It is also equipped with sensitive field anonymization, data retention period, deletion / revocation authorization mechanism and log audit module to ensure data security and integrity and comply with privacy compliance requirements.

[0028] The access control module communicates with the instant messaging access module and implements hierarchical control of system operation permissions based on the organizational structure of the enterprise-level communication platform. The system pre-defines three user roles: Super Administrator, HR Administrator, and Regular HR. The Super Administrator has full system operation permissions, including user role assignment and system parameter configuration. The HR Administrator has permissions for job management, candidate management, and recommendation settings. Regular HR personnel only have basic permissions such as viewing chat context, receiving job recommendations, and viewing resumes. The system also includes modules for data visibility range and abnormal access alerts for different roles. The access control module synchronizes with the organizational structure information of the enterprise-level communication platform in real time. When the organizational structure of the enterprise-level communication platform changes, the system automatically updates user permissions to ensure the accuracy and real-time nature of access control.

[0029] The opportunity conversion module communicates and connects with the real-time context processing engine, the multi-dimensional semantic understanding module, and the dynamic job matching engine. Its core function is to achieve real-time conversion of recruitment opportunities, constructing a closed-loop process of "opportunity identification → requirement clarification → job creation → instant recommendation." It optimizes opportunity identification from simple keyword matching to a multi-factor model, enhancing creativity. Opportunity Identification: A multi-factor model is used to monitor HR-customer conversations transmitted by the real-time context processing engine. The multi-factors include recruitment keywords, conversation roles, customer's historical recruitment behavior, density of job entities in the conversation, completeness of fields such as number of hires / location / salary, message frequency, urgency score, whether similar positions already exist, and whether the demand clarification state machine is triggered. Recruitment keywords are subordinate factors, including "hiring", "recruiting", "job position", "urgent hiring", and "searching for people". When the comprehensive score of the multi-factors exceeds the preset threshold, the opportunity identification process is immediately triggered, and the conversation is marked as an opportunity conversation. Requirements Clarification: The system automatically sends pre-set clarification questions to the client, including "What positions are you recruiting for?", "How many people are you recruiting for?", "How urgent is the position?", "What are the core skills required?", and "Where is the work location?". After the client replies, the multi-dimensional semantic understanding module parses the reply and extracts the structured requirements information. Job creation: Based on the parsed requirements information, the system automatically generates a job draft, which includes information such as job title, core requirements, salary range, work location, and urgency. The draft is then pushed to the corresponding HR's enterprise communication platform window for HR confirmation. Instant Recommendations: Once HR confirms a job draft, the system immediately triggers a dynamic job matching engine to search the candidate database for candidates who match the job, generating a job recommendation list. This list is then pushed to the chat window between HR and the client by the real-time recommendation push module, enabling instant conversion of business opportunities.

[0030] The compliance processing module communicates with the instant messaging access module, data storage module, and access control module to achieve compliant control over personal information processing, adhering to the requirements of the Personal Information Protection Law and addressing the issue of superficial privacy compliance in existing technologies. Specifically, it includes: pre-processing notification, clearly informing candidates and clients of the purpose, method, and scope of personal information processing; authorization scope management, strictly controlling access and use permissions for personal information; candidate and client consent records, retaining proof of consent from candidates and clients regarding personal information processing; explainable recommendation reasons, clarifying the core basis for job recommendations to HR and users, avoiding the "black box" of automated decision-making; and manual review of automated decisions, providing a manual review entry for automatically generated recommendation results and business opportunity identification results to ensure the rationality of decisions. Simultaneously, in conjunction with the data storage and access control modules, it implements sensitive field anonymization, data retention period control, deletion / revocation authorization mechanisms, and log auditing to reduce privacy compliance risks.

[0031] The fault degradation module communicates with the real-time context processing engine and the dynamic job matching engine to handle system failures and ensure stable system operation. When the WebSocket long connection fails, it automatically switches to the HTTP long polling mechanism, where the client periodically requests new messages from the server. Although the real-time performance is slightly reduced, message capture is guaranteed to be uninterrupted. When the stream processing unit fails, it automatically switches to batch processing mode as a temporary supplement. When the cache index fails, it automatically switches to the vector retrieval unit's real-time calculation mode to ensure the normal operation of the recommendation function.

[0032] As a preferred embodiment, the Kafka+Flink+Redis+Elasticsearch+BERT combined technology solution in this invention can be iteratively replaced according to the latest technology.

[0033] Example 1

[0034] This invention relates to a real-time intelligent job recommendation system based on instant messaging context, comprising an instant messaging access module, a real-time context processing engine, a multi-dimensional semantic understanding module, a dynamic job matching engine, a real-time recommendation push module, a data storage module, a permission management module, a business opportunity conversion module, and a compliance processing module. These modules work collaboratively, and the specific implementation process is as follows: 1) Implementation of the instant messaging access module: Based on the enterprise-level communication platform's session content archiving interface, application callback interface, and authorized message synchronization interface, the authorized employee scope is configured to include internal HR personnel. Personal information processing notices are pushed to clients and candidates to obtain written or electronic consent certificates. A message retrieval cursor is used to record the retrieval progress, enabling breakpoint resumption. When message transmission is interrupted, retrieval can resume from the cursor's position, avoiding message loss. The DecryptData interface provided by the enterprise-level communication platform is used to decrypt encrypted messages, and a message compensation mechanism is configured to retrieve unsuccessfully retrieved messages a second time. Data minimization processing is adopted, capturing only recruitment-related chat messages and necessary metadata, excluding irrelevant information. Persistent communication is established with the backend via a WebSocket long connection to obtain chat messages between HR, candidates, and clients in real time. Voice messages are converted to text using a built-in speech-to-text plugin, and only the name and type of file messages are extracted. All message data is transmitted to the real-time context processing engine in real time.

[0035] 2) Implementation of the real-time context processing engine: Kafka message queue is used as the message queue to receive and cache message data transmitted by the instant messaging access module. The message retention time is set to 7 days to ensure message traceability. Flink stream processing framework is used as the stream processing unit to consume messages in Kafka message queue in real time. The processing includes: deduplicating messages based on session ID and message timestamp to avoid duplicate processing; converting text messages and voice messages after text conversion to UTF-8 encoding standard text format; and classifying messages into two categories, HR-candidate dialogue and HR-customer dialogue, according to the enterprise-level communication platform identity of the sender and receiver, and transmitting them to the multi-dimensional semantic understanding module after marking.

[0036] 3) Implementation of the multi-dimensional semantic understanding module: A BERT-based NER model is used as the explicit requirement extraction model. Input features include standardized message text, message sending timestamp, session ID, and candidate / customer basic profile. Output is structured explicit requirement data, including skills, work location, salary range, etc. Urgency identification uses a weighted summation formula: Urgency score = 0.4 × keyword weight + 0.3 × sending time weight + 0.3 × sending frequency weight. Keyword weight is set based on the frequency of occurrence of keywords such as "urgently hiring" and "as soon as possible." Sending time weight is set to 1 for non-working hours (18:00-9:00 the next day) and 0.5 for working hours. Sending frequency weight is set based on the number of messages sent within one hour (≥5 messages = 1, 3-4 messages = 0.7, 1-2 messages = 0.7). 0.3); Sentiment analysis uses a pre-trained RoBERTa sentiment analysis model to output sentiment polarity (positive / negative) and sentiment intensity in the 0-1 range; Implicit skill inference uses a prompt-based contextual reasoning model, based on a preset prompt (e.g., "Extract core skills from 'have experience in e-commerce platform development':"), to infer specific skills from fuzzy descriptions and map them to a standard skill tag library; Demand modeling uses a weighted fusion approach, combining explicit and implicit demands into a dynamic demand vector V=[skill tag vector, location vector, salary range, years of experience, education level, urgency, sentiment intensity, job category, customer priority], where the explicit demand weight is set to 0.6 and the implicit demand weight is set to 0.4.

[0037] 4) Implementation of the dynamic job matching engine: The pre-computation unit calculates the initial matching degree based on candidate profile data and job database data using cosine similarity, generates a candidate job matching matrix, and stores it in Redis cache with a cache expiration time of 1 hour, refreshing it periodically; the incremental update unit sets the threshold for the change in dynamic demand vector to 0.2 (cosine distance between vectors). When the change exceeds the threshold, incremental retrieval is triggered, updating only the matching data corresponding to the change; the matching retrieval unit uses a weighted fusion algorithm to calculate the recommendation score S = 0.3 × candidate-job basic matching score + 0.3 × demand vector and job vector similarity + 0.2 × urgency weight + 0.15 × session feedback weight - 0.05 × permission / visibility penalty, where the session feedback weight is updated in real time based on user click (+0.1), ignore (-0.05), favorite (+0.15) behavior; the 5 jobs with the highest matching degree are selected to generate a recommendation list, ensuring that the recommendation results are returned within the preset time threshold under the condition of Redis cache hit. As a preferred implementation, the vector retrieval unit is implemented using Elasticsearch, while the fallback retrieval unit directly calculates the matching degree in real time based on candidate profiles and job information when Redis cache fails.

[0038] 5) Implementation of the real-time recommendation push module: Using an enterprise-level communication platform message interface, the job recommendation list is pushed in the form of message cards. The message card includes the job name, company name, core skill requirements, salary range, matching degree, and three operation buttons: "View Resume", "Initiate Communication", and "Save Job". Button jump is implemented through the enterprise-level communication platform JSAPI without switching applications. The push timing is set as follows: when HR communicates with candidates, for each valid requirement message (including explicit or implicit requirements) captured, an updated recommendation list is pushed immediately; when HR communicates with clients, after the business opportunity is converted (HR confirms the job draft), a matching recommendation list is pushed immediately.

[0039] 6) Implementation of the data storage module: MySQL is used as the relational database to store job database data, candidate basic information, and system permission data. Data is encrypted and stored (AES-256 encryption) with regular backups (daily backup at midnight). MongoDB is used as the non-relational database to store chat message data, candidate profile data, and system operation logs. Sensitive fields are anonymized (e.g., the middle 4 digits of the phone number are hidden, and only the range of salary is displayed). Data retention is set to 1 year, and data is automatically deleted upon expiration. Redis is used as the cache index to store the candidate job matching matrix. A cache cluster is set up to ensure high availability. A log auditing module is configured to record all data access and operation behaviors, and alarms are triggered for abnormal access (e.g., unauthorized access, high-frequency access).

[0040] 7) Implementation of the access control module: Based on the enterprise-level communication platform's organizational structure API, the module synchronizes enterprise organizational structure information in real time, and presets three roles: super administrator, HR administrator, and ordinary HR, assigning corresponding permissions; it configures the data visibility range for different roles: the super administrator can view all data, the HR administrator can view the positions and candidate data of the departments under their supervision, and ordinary HR can only view the conversations and recommendation data they are responsible for; it configures an abnormal access alarm module, and immediately pushes alarm messages to the super administrator when unauthorized access, high-frequency access, or abnormal IP access occurs.

[0041] 8) Implementation of the opportunity conversion module: The factor weights of the multi-factor model are set as follows: recruitment keywords (0.15), dialogue roles (0.1), customer historical recruitment behavior (0.2), job entity density in the conversation (0.15), field completeness (0.15), message frequency (0.1), urgency score (0.1), whether there are similar positions (0.03), whether the demand clarification state machine is triggered (0.02). When the comprehensive score is ≥0.6, opportunity identification is triggered. Demand clarification issues adopt adaptive push. Based on the extracted demand information, only unclear issues are pushed (e.g., if the job name has been extracted, the job name clarification issue is not pushed). After the job draft is generated, it is pushed to the HR enterprise-level communication platform window. HR can modify and confirm it. After confirmation, the dynamic job matching engine is immediately triggered to generate a recommendation list and push it.

[0042] 9) Implementation of the compliance processing module: Upon system login and initial interaction with candidates / customers, a personal information processing notice is pushed, clearly stating the processing purpose (job recommendation, business opportunity conversion), method (message parsing, demand extraction), and scope (chat messages, resume information, etc.), obtaining user consent and retaining records; The automated decision-making manual review module sets a review threshold. When the matching degree of the recommendation list is too large (the difference between the highest and lowest matching degree is ≥0.5) or the comprehensive score of business opportunity identification is between 0.6 and 0.7, manual review is triggered, and the HR administrator conducts the review and confirmation; A deletion / revocation authorization mechanism is configured, and candidates / customers can apply to delete personal information and revoke authorization. After the application, the system completes the data deletion and revocation of permissions within 24 hours.

[0043] 10) Implementation of the fault degradation module: Real-time monitoring of the operating status of system components. When the WebSocket long connection fails, it automatically switches to the HTTP long polling mechanism with a polling interval of 1 second. When the Flink stream processing unit fails, it automatically switches to batch processing mode, processing a batch of messages every 5 minutes. When the Redis cache fails, it automatically switches to Elasticsearch real-time computing mode to ensure the normal operation of the recommendation function, and pushes fault alarm messages to the super administrator for timely troubleshooting.

[0044] The overall workflow of this system is as follows: 1. The instant messaging access module obtains chat messages and metadata between HR and candidates and customers based on the authorized interface of the enterprise-level communication platform, and transmits them to the real-time context processing engine through WebSocket long connection, while executing compliance processing procedures (notification, consent recording, etc.). 2. The real-time context processing engine performs deduplication, standardization, and classification on message data, and transmits the processed message data to the multi-dimensional semantic understanding module; 3. The multi-dimensional semantic understanding module extracts explicit and implicit requirements from the message data, merges them to generate a dynamic requirement vector, and transmits it to the dynamic job matching engine; 4. The dynamic job matching engine retrieves the candidate job matching matrix from the cached index based on the dynamic demand vector, generates a job recommendation list through a weighted fusion algorithm, and transmits it to the real-time recommendation push module. 5. The real-time recommendation and push module pushes the job recommendation list to the corresponding enterprise-level communication platform chat window in the form of message cards for HR to view and operate; 6. The opportunity conversion module uses a multi-factor model to monitor the dialogue messages between HR and customers in real time, identify recruitment opportunities, and trigger the instant recommendation process after the needs are clarified and the job is created. 7. The data storage module stores all data during system operation in real time and performs operations such as data anonymization, backup, and log auditing; the access control module implements hierarchical access control; the compliance processing module ensures that personal information processing is compliant; and the fault degradation module monitors the system's operating status in real time, responds to various component failures, and ensures stable system operation.

[0045] The following is a sample dialogue to illustrate the specific working process of the system: The client said, "We urgently need to hire two Java developers with e-commerce backend experience this week, preferably in Shanghai, with a budget of around 25k." The instant messaging access module obtains the message and metadata (sender is the client, receiver is HR, sending time is weekday 10:30, session ID is XXX), and after decryption and data minimization, it is transmitted to the real-time context processing engine. The real-time context processing engine deduplicatizes and standardizes the message, classifies it as a dialogue between HR and customers, and transmits it to the multi-dimensional semantic understanding module. The multi-dimensional semantic understanding module extracts explicit requirements through a NER model: skills (Java, e-commerce backend), work location (Shanghai), salary range (around 25k), and number of openings (2). It calculates a score based on urgency: keywords "urgent hiring" and "this week" have a weight of 0.4, sending time during working hours has a weight of 0.3, and sending frequency (1 message) has a weight of 0.3. The urgency score is calculated as 0.4 × 1 + 0.3 × 0.5 + 0.3 × 0.3 = 0.64. Sentiment analysis outputs a positive result with a sentiment intensity of 0.8. Implicit skills inference is derived from "e-commerce backend," indicating the need for database design and distributed system development skills. The results are then fused to generate a dynamic requirement vector V. The dynamic job matching engine is based on the dynamic demand vector V, retrieves the candidate job matching matrix in the Redis cache, calculates the recommendation score S, filters out the 5 jobs with the highest matching degree, and generates a recommendation list. The multi-factor model comprehensive score of the opportunity conversion module = 0.15×1 (including the keyword "urgently hiring") + 0.1×1 (customer role) + 0.2×0.8 (customer has recruitment history in the past 3 months) + 0.15×0.9 (high job entity density) + 0.15×0.8 (high field completeness) + 0.1×0.3 (message frequency 1) + 0.1×0.64 (urgency score) + 0.03×0 (no similar jobs) + 0.02×0 (clarification state machine not triggered) = 0.72 ≥ 0.6, triggering opportunity identification; the system automatically pushes clarification questions to the customer: "What are the core skills required for this position?" "Is the number of recruits confirmed to be 2?"; The client replied, "Core skills also include MySQL, and the number of positions is confirmed at 2." The multi-dimensional semantic understanding module extracted and supplemented the explicit requirements (MySQL skills), updated the dynamic requirement vector V, and the dynamic job matching engine updated the recommendation list. The business opportunity conversion module generated a job draft (Job title: E-commerce backend development engineer, number of positions: 2, work location: Shanghai, salary: around 25k, core skills: Java, e-commerce backend, MySQL, urgency: 0.64), and pushed it to HR for confirmation. After HR confirms the job draft, the dynamic job matching engine searches again and generates a final recommendation list, which is then pushed to the chat window between HR and the client by the real-time recommendation push module for HR and the client to view.

[0046] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A real-time intelligent job recommendation system based on instant messaging context, characterized in that: include: The instant messaging access module obtains session messages based on the enterprise-level communication platform's session content archiving interface, application callback interface, or authorized message synchronization interface. The instant messaging access module is configured with authorized employee scope, customer / candidate notification and consent mechanism, message retrieval cursor, breakpoint resume, decryption process, message compensation mechanism, and data minimization processing module. It is used to establish a persistent connection with the backend, obtain chat messages and message metadata between HR and candidates / customers, and transmit the obtained message data to the backend in real time. The real-time context processing engine communicates with the instant messaging access module and includes a message queue and a stream processing unit. The message queue is used to receive and buffer message data transmitted by the instant messaging access module, and the stream processing unit is used to process the message data in the message queue in real time to complete message parsing and preliminary classification. The multi-dimensional semantic understanding module communicates with the real-time context processing engine and is used to perform entity extraction, confidence calculation, urgency scoring and implicit skill mapping based on the message data output by the real-time context processing engine. It extracts structured explicit and implicit demand information and synthesizes explicit demand, implicit demand and historical profile into a dynamic demand vector, wherein the dynamic demand vector V=[skill tag vector, location vector, salary range, years of experience, education, urgency, emotional intensity, job category, customer priority]; The dynamic job matching engine communicates with the multi-dimensional semantic understanding module and includes a pre-calculation unit, an incremental update unit, and a matching retrieval unit. The pre-calculation unit is used to pre-calculate the candidate job matching matrix based on candidate profile data and job database data. The incremental update unit is used to trigger incremental retrieval when the change in the dynamic demand vector exceeds a threshold. It only updates the matching data corresponding to the changed candidate profile or job information and does not perform a full recalculation. The matching retrieval unit is used to perform weighted fusion of the pre-calculated candidate job matrix and the dynamic demand vector to generate a job recommendation list. The real-time recommendation push module communicates with the dynamic job matching engine and the instant messaging access module respectively. It is used to push the job recommendation list to the corresponding enterprise-level communication platform chat window in the form of message cards through the enterprise-level communication platform message interface. It also provides one-click access to view resumes and one-click initiation of communication, thereby reducing recommendation response latency, generating recommendation results under cache hit conditions, and returning the candidate list within a preset time threshold. The data storage module communicates with the real-time context processing engine, the multi-dimensional semantic understanding module, and the dynamic job matching engine respectively. It is used to store message data, candidate profile data, job database data, candidate job matching matrix and system operation logs. It is configured with sensitive field desensitization, data retention period, deletion / revocation authorization mechanism and log auditing module. The access control module communicates with the instant messaging access module and is used to implement hierarchical control of system operation permissions based on the organizational structure of the enterprise-level communication platform. It ensures that different roles can only access data and functions within their corresponding permission scope and is configured with data visibility range and abnormal access alarm modules for different roles. The opportunity conversion module communicates with the real-time context processing engine, the multi-dimensional semantic understanding module, and the dynamic job matching engine. It uses a multi-factor model to identify recruitment opportunities in chat messages, automatically initiates requirement clarification, generates job drafts, and triggers an instant recommendation process after HR confirmation. The compliance processing module communicates with the instant messaging access module, data storage module, and access control module to achieve compliance control over personal information processing, including functions such as pre-processing notification, authorization scope management, candidate and customer consent recording, explanatory reasons for recommendation, and automated decision-making with manual review.

2. The real-time intelligent job recommendation system based on instant messaging context according to claim 1, characterized in that, The instant messaging access module uses a WebSocket long connection to establish communication with the backend, and the obtained message metadata includes sender identity information, receiver identity information, message sending timestamp, session ID and message type; The message types include text messages, voice messages, and file messages. Voice messages are transmitted together with text messages after being processed into text, while file messages only capture the file name and type information.

3. The real-time intelligent job recommendation system based on instant messaging context according to claim 1, characterized in that, The real-time context processing engine uses a Kafka message queue, and the stream processing unit uses the Flink stream processing framework. The processing steps of the stream processing unit include message deduplication, message format standardization, and message classification. The message classification includes two categories: dialogue between HR and candidates, and dialogue between HR and customers.

4. The real-time intelligent job recommendation system based on instant messaging context according to claim 1, characterized in that, The explicit requirement extraction of the multi-dimensional semantic understanding module adopts the Named Entity Recognition (NER) model. The extracted structured explicit requirement information includes skills, work location, salary range, years of work experience, and educational requirements. The implicit requirement extraction includes urgency identification, sentiment analysis, and implicit skill inference. Urgency identification is based on message text keywords, message sending time, and message sending frequency, and outputs an urgency score in the range of 0-1 through weighted calculation. Sentiment analysis is based on a sentiment analysis model and outputs sentiment polarity and sentiment intensity. Sentiment polarity includes positive and negative. Implicit skill inference is based on a contextual reasoning model, which infers specific skill requirements from fuzzy descriptions and maps them to skill tags.

5. The real-time intelligent job recommendation system based on instant messaging context according to claim 4, characterized in that, The input features of the multi-dimensional semantic understanding module include standardized message text, message metadata, historical session data, and candidate / customer basic profiles. The output format is structured explicit requirement data, implicit requirement feature values, and dynamic requirement vectors. Explicit and implicit requirements are synthesized into dynamic requirement vectors through weighted fusion, and the weights are dynamically adjusted based on the business scenario.

6. The real-time intelligent job recommendation system based on instant messaging context according to claim 1, characterized in that, The pre-computation unit of the dynamic job matching engine stores the candidate job matching matrix in a cache index; The recommendation score S of the matching retrieval unit is calculated as follows: S = α × candidate - job basic matching score + β × similarity between demand vector and job vector + γ × urgency weight + δ × conversation feedback weight - ε × permission / visibility penalty, where α, β, γ, δ, and ε are weight coefficients that can be dynamically adjusted according to the business scenario. The generated job recommendation list is sorted from high to low matching degree. Each recommended job includes job name, core requirements, salary range, and matching degree score. The conversation feedback weight is updated in real time based on user click / ignore feedback.

7. The real-time intelligent job recommendation system based on instant messaging context according to claim 1, characterized in that, The multi-factor model of the business opportunity conversion module includes recruitment keywords, dialogue roles, customer historical recruitment behavior, job entity density in the conversation, completeness of fields such as number of recruits / location / salary, message frequency, urgency score, whether there are already similar positions, and whether the demand clarification state machine is triggered. Among them, recruitment keywords are subordinate factors, including "recruiting", "recruiting", "job position", "urgent recruitment" and "searching for people". The requirement clarification process involves the system automatically sending preset clarification questions to the message sender. These preset clarification questions include the job title, number of openings, urgency level, core requirements, and work location.

8. The real-time intelligent job recommendation system based on instant messaging context according to claim 1, characterized in that, The data storage module uses a combination of relational and non-relational databases. The relational database stores job database data, candidate basic information, and system permission data, while the non-relational database stores chat message data, candidate profile data, and candidate job matching matrix. The cache index uses Redis for caching. The matching retrieval unit is also configured with a vector retrieval unit and a fallback retrieval unit. The vector retrieval unit is implemented using Elasticsearch, and the fallback retrieval unit is used to switch to real-time computing mode when the cache fails.

9. The real-time intelligent job recommendation system based on instant messaging context according to claim 1, characterized in that, It also includes a fault degradation module that communicates with the real-time context processing engine and the dynamic job matching engine. When the WebSocket long connection fails, it automatically switches to the HTTP long polling mechanism; when the stream processing unit fails, it automatically switches to the batch processing mode; and when the cache index fails, it automatically switches to the vector retrieval unit real-time calculation mode.

10. The real-time intelligent job recommendation system based on instant messaging context according to claim 1, characterized in that, The message card pushed by the real-time recommendation push module includes job title, company name, core skill requirements, salary range, matching degree, and operation buttons. The operation buttons include "View Resume", "Initiate Communication", and "Save Job". Clicking the operation button will directly jump to the corresponding function page without switching applications.

Images