Intelligent matching method and system for talent resume based on semantic retrieval

Abstract

The present application relates to the technical field of data processing, and discloses a talent resume intelligent matching method and system based on semantic retrieval, comprising: constructing a skill preposition dependence graph containing independent physical nodes and directed edges and configuring edge transmission damping coefficients; extracting resume text skill entities, continuous spans and time-to-present time to generate time sequence feature tuples and writing into an inverted index; mapping explicit skill entities in job requirement description text to the graph, constructing an asymmetric query topology tree along the bottom layer dependence and assigning static weights; extracting candidate resume from the inverted index to read time sequence features, calculating dynamic time decay rates and comprehensive proficiency values in combination with the edge transmission damping coefficients, and calculating matching scores to generate a recommendation sequence in combination with the static weights. The present application completes implicit basic skills by constructing a topology tree, quantifies basic counterproductive and skill decay phenomena by using damping coefficients and time sequence features, overcomes static label evaluation errors, and improves retrieval comprehensiveness and sorting accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for intelligent matching of talent resumes based on semantic retrieval. Background Technology

[0002] In corporate recruitment, existing resume retrieval systems use keyword comparison or vector matching. When performing the matching calculation between job requirements and job seekers' skills, the system relies heavily on the static proficiency tags filled in by job seekers to complete the scoring. The existing retrieval mechanism uses skill vocabulary as an independent feature for single-point comparison, without analyzing the prerequisite dependencies and evolution relationships of skills within the technical field. When advanced technical terms are entered for job requirements, the system can only perform literal character matching and cannot follow the technology development path to complete implicit basic skills. This results in candidates who meet the underlying technical requirements not being recalled, and there are omissions in the resume retrieval data. Meanwhile, existing matching technologies do not incorporate time decay factors for quantitative consideration. Skill proficiency changes with the accumulation of continuous use time and the extension of disuse time. Static tags alone cannot reflect the continuous application span of technical skills in work experience, nor can they assess the skill rust caused by disuse. Existing systems do not establish a correlation mechanism between the active state of basic skills and the forgetting and decay of advanced skills, causing the calculated matching score to deviate from the objective mastery level and reducing the ranking accuracy of the final search and recommendation results. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for intelligent matching of talent resumes based on semantic retrieval. This solves the problems of existing resume matching technologies failing to parse skill prerequisite dependencies and ignoring time decay factors, which leads to resume recall omissions and distorted skill proficiency assessments.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a method for intelligent matching of talent resumes based on semantic retrieval, comprising: Construct a directed acyclic skill prerequisite dependency graph containing independent physical nodes and directed edges, and configure edge transmission damping coefficients for directed edges; Extract the skill entities, duration, and date of origin from the original resume text, generate a set of time-series feature tuples, and write them into the inverted index; Map explicit skill entities in the job requirement description text to a skill prerequisite dependency graph, traverse along directed edges to the lower-level dependency direction, construct an asymmetric query topology tree and assign static topology weights. The nodes based on the asymmetric query topology tree extract candidate resume text from the inverted index, calculate the activity value based on the continuous span and time since the child node, and calculate the dynamic time decay rate of the parent node by combining the edge propagation damping coefficient of the directed edge between the parent and child nodes. Based on this, the comprehensive proficiency value is calculated, and the final matching score is calculated by combining the static topology weights to generate a recommended resume sequence.

[0005] Furthermore, a directed acyclic skill prerequisite dependency graph containing independent physical nodes and directed edges is constructed, including: Obtain the set of skill entities in the technical field, map each skill entity to an independent physical node in the graph database, assign a globally unique node identifier to each independent physical node and store it persistently. Based on the underlying technological dependencies between skill entities, when a high-level skill entity uses a basic skill entity as its prerequisite technological foundation, a directed edge is established from the independent physical node representing the high-level skill entity to the independent physical node representing the basic skill entity. Configure edge propagation damping coefficients for directed edges. The edge propagation damping coefficients quantify the extent to which the activity of basic skill entities slows down the decay rate of time forgetting of advanced skill entities. The edge propagation damping coefficients are written into the graph database as attribute data of directed edges.

[0006] Furthermore, the skill entities, duration, and time to date of the original resume text are extracted to generate a set of time-series feature tuples, including: Named entity recognition processing is performed on the original resume text to identify skill entities, start timestamps, and end timestamps. The original resume text is then segmented into a set of work experience event blocks arranged in chronological order. For each work experience event block, the difference between the end timestamp and the start timestamp is used as the duration, and the difference between the current timestamp and the end timestamp is used as the time since then. Extract the skill entities from each work experience event block, and combine the current date and duration of the same skill entity in each work experience event block to generate a set of time-series feature tuples consisting of multiple pairs of current date and duration.

[0007] Furthermore, a set of time-series feature tuples is generated and written into the inverted index, including: The time-series feature tuple set is set as a custom composite feature payload in the underlying retrieval engine, and the time-to-date and duration-paired data in the time-series feature tuple set are converted into continuous binary byte stream data using a serialization protocol. The inverted index uses the globally unique node identifier corresponding to the independent physical node of the skill entity as the key, and the document identifier of the original resume text and the continuous binary byte stream data as the corresponding value. The key-value pair structure is written into the inverted index linked list of the underlying search engine, and the document identifier and the continuous binary byte stream data form a custom payload field that is physically bound.

[0008] Furthermore, traversing along the directed edges towards the lower-level dependencies, an asymmetric query topology tree is constructed, including: Locate the explicit skill entity to the corresponding independent physical node, and add the corresponding independent physical node as the initial query entity node to the root node set; Using each initial query entity node in the root node set as the traversal origin, perform a breadth-first hierarchical expansion traversal along the directed edges in the skill prerequisite dependency graph towards the underlying basic dependencies. Read the expansion depth value of the current traversal path and compare the expansion depth value with the preset maximum level threshold. Before the expansion depth value reaches the maximum level threshold, pull the currently visited downstream independent physical nodes as implicit skill nodes into the asymmetric query topology tree. When the expansion depth value reaches the maximum level threshold, truncate the current traversal path.

[0009] Furthermore, the static topology weights are assigned, including: Assign initial baseline weights to the initial query entity nodes in the root node set; The tree hierarchy depth of the implicit skill node is determined based on the number of directed edge jumps between the implicit skill node and the initial query entity node with the fewest directed edge jumps. According to the step decay rule, the static topological weight of each independent physical node in the asymmetric query topology tree is calculated based on the initial baseline weight, the node depth decay constant, and the tree level depth.

[0010] Furthermore, candidate resume text is extracted from the inverted index based on nodes of the asymmetric query topology tree, including: Extract the globally unique node identifiers of all independent physical nodes in the asymmetric query topology tree, and input them as query terms into the underlying retrieval engine to match the corresponding set of document identifiers in the inverted index list; The product of the number of document identifiers matched by each independent physical node and the corresponding static topology weight is used as the global static parameter. For each document identifier, the global static parameters corresponding to each independent physical node that matches the document identifier are summed to obtain the initial matching score of the document identifier. Sort the documents in descending order of their initial matching scores, extract the top-ranked document identifiers and their corresponding original resume texts to form a candidate set of original resume texts.

[0011] Furthermore, the dynamic aging attenuation rate is calculated by combining the edge transmission damping coefficient, including: For the candidate's original resume text, perform topological sorting on the asymmetric query topology tree in memory to generate a computational sequence from implicit skill nodes to the root node set; Read the continuous binary byte stream data formed by serializing the time-series feature tuple set in the inverted index according to the calculation sequence, and deserialize the continuous binary byte stream data into a time-series feature tuple set. Calculate the activity value of independent physical nodes based on the pairing data of the present time and duration recorded in the time-series feature tuple set. The dynamic time-dependent decay rate of the parent node is obtained by combining the activity value of the child node with the edge transmission damping coefficient configured on the directed edge from the parent node to the child node. When an independent physical node has no child nodes or all child nodes have an activity value of 0, the dynamic time-degradation rate of the independent physical node is set to the global basic degradation constant.

[0012] Preferably, the calculation of the overall proficiency score includes: After updating the dynamic time-depreciation rate of each independent physical node in the asymmetric query topology tree, the updated dynamic time-depreciation rate is applied to perform a proficiency integral operation on the time-series feature tuple set of each independent physical node. The continuous span recorded by each independent physical node in each work experience event block is calculated and accumulated based on the corresponding dynamic time decay rate and the time since the work experience event block, to obtain the comprehensive proficiency value of each independent physical node. The comprehensive proficiency value of each independent physical node is multiplied by the corresponding static topology weight, and the product results of all independent physical nodes are summed to obtain the final matching score. The candidate original resume text set is then rearranged according to the final matching score, and a recommended resume sequence is output.

[0013] A second aspect of the present invention provides a talent resume intelligent matching system based on semantic retrieval, comprising: The graph construction module is used to build a skill prerequisite dependency graph with a directed acyclic graph structure in the graph database, and to configure edge transmission damping coefficients for the directed edges in the skill prerequisite dependency graph. The index building module is used to receive the original resume text, extract skill entities, duration and time to date, generate a set of time-series feature tuples, and serialize the set of time-series feature tuples into continuous binary byte stream data and write it into the inverted index linked list of the underlying retrieval engine. The query rewriting module is used to receive job requirement description text, extract explicit skill entity set, map explicit skill entity set to skill pre-dependency graph, traverse along directed edges in skill pre-dependency graph towards underlying dependency direction, construct asymmetric query topology tree and assign static topology weights. The scoring and retrieval module is used to extract candidate original resume text from the inverted index, read the duration and time to date of each independent physical node in the asymmetric query topology tree, calculate the dynamic time-related decay rate by combining the edge propagation damping coefficient, calculate the comprehensive proficiency value, calculate the final matching score by combining the static topology weight, and generate a recommended resume sequence based on the final matching score.

[0014] This invention provides a method and system for intelligent matching of talent resumes based on semantic retrieval. It has the following beneficial effects: 1. This invention constructs a skill pre-dependency graph with a directed acyclic graph structure, traversing and expanding the explicit skill entities in the job requirement description text along the directed edges towards the underlying dependency direction to build an asymmetric query topology tree; during retrieval and matching, based on the underlying technical dependency logic, it automatically associates and completes the job requirements with the required implicit basic skills, avoiding matching omissions caused by isolated retrieval of a single skill tag, and improving the comprehensiveness of resume retrieval data.

[0015] 2. This invention configures edge transmission damping coefficients for directed edges in the skill prerequisite dependency graph, and combines the duration and time elapsed extracted from the original resume text to calculate the dynamic time decay rate of the parent node by combining the activity value of the child node with the edge transmission damping coefficient. The calculation process establishes a parameter transmission mechanism that delays the natural decay of advanced skill proficiency based on the activity of basic skills. It can calculate the candidate's objective skill mastery level based on the time distribution of work experience, and overcome the evaluation error caused by relying on static proficiency labels filled in by job seekers.

[0016] 3. This invention applies a dynamic time decay rate to perform a proficiency integral operation on the time-series feature tuple set of each independent physical node to obtain a comprehensive proficiency value, and then multiplies the comprehensive proficiency value with the static topology weight to obtain the final matching score. The scoring process takes into account the importance of each skill in the job requirement topology and the amount of time accumulated by the candidate in applying the skills, and outputs a recommended resume sequence based on the objective level of skill mastery and job fit, thereby improving the ranking accuracy of talent matching results. Attached Figure Description

[0017] Figure 1 This is the timing interaction diagram of the present invention; Figure 2 This is a schematic diagram of the method steps of the present invention; Figure 3 This is a schematic diagram of the map and index construction process of the present invention; Figure 4 This is a schematic diagram of the query rewriting and scoring process of the present invention; Figure 5 This is a schematic diagram illustrating the recommended resume sorting effect of the present invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see the appendix Figure 1 This invention provides a semantic retrieval-based intelligent resume matching system, the system comprising: The graph construction module establishes a directed acyclic graph structure containing multiple skill entities and prerequisite dependencies within the graph database, and configures edge-transmitting damping coefficients for the directed edges in the directed acyclic graph structure. The index building module receives the original resume text, extracts skill entities and time-series parameters, converts the original resume text into time-series feature tuples, and serializes the time-series feature tuples into byte stream data and stores them in the inverted index payload of the underlying retrieval engine. The query rewriting module receives the job requirement description text from external input, extracts the set of explicit skill entities from the job requirement description text, maps the set of explicit skill entities to the skill prerequisite dependency graph, and traverses along the directed edges in the skill prerequisite dependency graph to construct an asymmetric query topology tree with static topological weights. The scoring and retrieval module receives the asymmetric query topology tree, retrieves candidate documents in the underlying retrieval engine, performs static coarse ranking calculation and dynamic damped fine ranking calculation in sequence, summarizes the scores of each node and outputs a recommended resume sequence. The graph construction module interacts with the index construction module and the query rewriting module through a network interface, and provides the skill prerequisite dependency graph to the index construction module and the query rewriting module; the index construction module stores the processed data in a non-volatile storage medium for the scoring and retrieval module to call at runtime.

[0020] Please see the appendix Figure 2 This invention provides a method for intelligent matching of talent resumes based on semantic retrieval, the method comprising the following steps: S1, obtain the preset technical field skill entities and the underlying technical dependency rules between skill entities, initialize and construct the skill prerequisite dependency graph in the graph database; S2, the input original resume text is segmented into work experience event blocks, the duration and time to date corresponding to each work experience event block are extracted as time series parameters, and the skill entities in each work experience event block are extracted. Multiple sets of time series parameters corresponding to the same skill entity are mapped and combined to generate a set of time series feature tuples. The set of time series feature tuples is used as a composite feature payload and written into the inverted index linked list of the underlying retrieval engine in a key-value pair structure. S3, extract the initial query entity set from the job requirement description text, map the initial query entity set as the root node set to the skill prerequisite dependency graph, perform a hierarchical traversal along the directed edges from the root node set to the lower-level dependency direction, construct an asymmetric query topology tree with a hierarchical structure, and assign static topology weights according to the tree-like hierarchy depth of independent physical nodes in the asymmetric query topology tree. S4, the retrieval engine receives the asymmetric query topology tree, extracts candidate documents from the inverted index and performs coarse sorting to obtain a specified number of candidate raw resume text sets; For each original resume text in the candidate original resume text set, a connectivity traversal is performed on the asymmetric query topology tree in an in-memory computing environment. The temporal parameters of the child nodes in the composite feature load are read, the dynamic time decay rate acting on the parent node is calculated, the comprehensive proficiency value of each node is calculated based on the dynamic time decay rate, and the final matching score is calculated in combination with the static topology weight. The recommended resume sequence is generated based on the final matching score.

[0021] Please see the appendix Figure 3 The S1 provided by the present invention may include the following sub-steps in a specific implementation: S11, the graph construction module obtains the set of skill entities in the technical field, maps each skill entity to an independent physical node in the graph database, and uses independent physical nodes, edges and attributes for data representation and storage. The implementation form of the graph database includes attribute graph database or resource description framework triplet library. During the mapping process, the graph construction module assigns a globally unique node identifier to each skill entity and persistently stores the text specification name of the skill entity as an attribute field of an independent physical node.

[0022] S12, the graph construction module establishes directed connections between independent physical nodes and forms a directed acyclic graph structure based on the underlying technical dependencies between skill entities. When an advanced skill entity uses a basic skill entity as its prerequisite technical foundation, the graph database establishes directed edges from the independent physical nodes representing the advanced skill entity to the independent physical nodes representing the basic skill entity. The skill prerequisite dependency graph includes a single connected graph or multiple independent connected subgraphs, and each independent connected subgraph is decoupled and separated.

[0023] S13, When the graph construction module establishes a directed edge, it configures the edge transmission damping coefficient for the directed edge and writes the edge transmission damping coefficient as the attribute data of the directed edge into the graph database. The edge transmission damping coefficient quantifies the extent to which the activity of basic skill entities slows down the decay rate of time forgetting of advanced skill entities. The graph construction module configures the edge transmission damping coefficient according to the preset hierarchical transmission rules in the system. The value range of the edge transmission damping coefficient is set in the open interval from 0 to 1. The value is set according to the iteration cycle of the technical field, the continuous span of the historical original resume text and the time since then. The closer the edge transmission damping coefficient is to 1, the stronger the damping transmission effect of basic skill entities on the attenuation of advanced skill entities. The skill prerequisite dependency graph is defined at the logical calculation level as follows: ; In the formula, For skill prerequisite dependency graph; It is a collection of independent physical nodes contained in the skill prerequisite dependency graph; It is the set of directed edges formed between independent physical nodes; This is the set of weighting functions that transmit damping coefficients to edges configured on directed edges.

[0024] In specific implementations, S2 provided by the present invention may include the following sub-steps: S21, the index building module receives the input raw resume text, performs named entity recognition processing on the raw resume text, calls the sequence labeling neural network structure composed of a bidirectional encoder representation model and a conditional random field, maps the character sequence contained in the raw resume text into a word vector matrix of a specific dimension, and inputs the word vector matrix into the sequence labeling neural network structure, the sequence labeling neural network structure outputs character-level entity boundaries and entity category labels, the index building module identifies skill entities, start timestamps and end timestamps according to entity boundaries and entity category labels, and segments the raw resume text into a set of discrete work experience event blocks arranged in time sequence; For each work experience event block, the index building module extracts the corresponding start and end timestamps, uses the difference between the end and start timestamps as the duration of the work experience event block, and uses the difference between the current timestamp and the end timestamp of the current system as the elapsed time of the work experience event block.

[0025] S22, the index building module extracts skill entities from each work experience event block based on the entity boundaries and entity category labels output by the sequence labeling neural network structure, compares the skill entities with the text specification names of independent physical nodes in the skill prerequisite dependency graph, and obtains the corresponding globally unique node identifiers. The same skill entity within the same work experience event block is retained only once; the index building module counts the number of work experience event blocks corresponding to a skill entity, and arrays the present time and duration of the skill entity in each work experience event block to generate a set of time-series feature tuples consisting of multiple pairs of present time and duration. The internal structure of the time-series feature tuple set is defined algebraically as follows: ; In the formula, This is the set of generated time-series feature tuples; This represents the number of work experience event blocks corresponding to the current skill entity. For the current skill entity number The time elapsed since the beginning of each work experience event block; For the first The continuous span recorded in each work experience event block; In this embodiment, to iterate through the index of the number of work experience event blocks corresponding to the current skill entity, The value range is from 1 to Positive integers.

[0026] S23, the index building module sets the generated time-series feature tuple set as a custom composite feature payload in the underlying retrieval engine, and uses a serialization protocol to convert the time-to-date and duration pairing data in the time-series feature tuple set into continuous binary byte stream data. The time-to-date and duration are stored as continuous memory storage blocks through continuous binary byte stream data, and a feature vector representation containing both duration and time-to-date attributes is formed on the physical storage medium.

[0027] S24, When the index building module builds the inverted index inside the underlying retrieval engine, it uses the globally unique node identifier of the skill entity as the key of the inverted index, the document identifier of the current original resume text and the continuous binary byte stream data as the corresponding value, and writes them into the inverted index linked list of the underlying retrieval engine according to the key-value pair structure. The document identifier and the continuous binary byte stream data form a physically bound custom payload field. The custom payload field is directly mounted on the inverted index list, so that the time-series feature tuple set can be directly read in the subsequent retrieval and calculation stage.

[0028] Please see the appendix Figure 4 In specific implementations, S3 provided by the present invention may include the following sub-steps: S31, The query rewrite module receives the job requirement description text from external input, performs entity extraction operation on the job requirement description text, and obtains the set of explicit skill entities clearly recorded in the job requirement description text. The entity extraction operation calls a sequence labeling neural network structure composed of a bidirectional encoder representation model and a conditional random field to map the character sequence contained in the job requirement description text into a word vector matrix of a specific dimension. The word vector matrix is ​​then input into the sequence labeling neural network structure, and the explicit skill entity set is extracted based on the character-level entity boundaries and entity category labels output by the sequence labeling neural network structure. The query rewriting module compares each explicit skill entity in the explicit skill entity set with the text specification name of the independent physical node in the skill prerequisite dependency graph preloaded into the underlying retrieval engine cache space. The explicit skill entities that are successfully matched are located to the corresponding independent physical nodes, and the corresponding independent physical nodes are added to the root node set as the initial query entity nodes. The root node set constitutes the zero-level vertex set of the query rewriting extension.

[0029] S32, the query rewriting module uses each initial query entity node in the root node set as the traversal origin, and performs a breadth-first hierarchical expansion traversal along the directed edges in the skill prerequisite dependency graph towards the underlying basic dependencies. During the hierarchical unfolding traversal, the query rewriting module uses a hash table in the system memory to record the globally unique node identifiers of the independent physical nodes that have been visited. The independent physical nodes that have been recorded will no longer be repeatedly fetched and calculated. The query rewriting module will fetch the currently visited downstream independent physical nodes as implicit skill nodes into the asymmetric query topology tree. The query rewriting module reads the expansion depth value of the current traversal path and compares it with the system's preset maximum level threshold. The maximum level threshold is set to a positive integer between 2 and 5 based on the historical retrieval recall distribution. Before the expansion depth value reaches the maximum level threshold, the query rewriting module pulls the currently accessed downstream independent physical nodes as implicit skill nodes into the asymmetric query topology tree. When the expansion depth value reaches the maximum level threshold, the query rewriting module truncates the current traversal path, forming an asymmetric query topology tree with a multi-branch hierarchical structure.

[0030] S33, The query rewriting module assigns static topological weights based on the tree-like hierarchy depth of each independent physical node in the asymmetric query topology tree. The static topological weights represent the baseline importance of the skill entity corresponding to the independent physical node in the current job requirement context. The tree hierarchy depth of the initial query entity node in the root node set is set to 0, and an initial baseline weight is assigned to the initial query entity node; the tree hierarchy depth of the implicit skill node is determined based on the number of directed edge hops between the implicit skill node and the nearest initial query entity node, and the static topology weight is calculated according to the step decay rule. The calculation mechanism of the static topology weight is defined algebraically as follows: ; In the formula, For the asymmetric query topology tree, the first Static topology weights corresponding to each independent physical node; In this embodiment, the initial baseline weights set for the system are... The value is set to a constant positive number 1; In this embodiment, the preset node depth attenuation constant is used. The value range is set to positive real numbers in the open interval from 0 to 1; For the first The tree-like hierarchy depth of an individual physical node in the asymmetric query topology tree; In this embodiment, the index for an independent physical node in the asymmetric query topology tree is used. The value range is from 1 to a positive integer representing the total number of independent physical nodes contained in the asymmetric query topology tree.

[0031] After the query rewrite module completes the static topology weight calculation for each independent physical node, it submits the asymmetric query topology tree with static topology weight attributes to the execution link of the underlying retrieval engine.

[0032] In specific implementations, S4 provided by the present invention may include the following sub-steps: S41, the scoring retrieval module receives the asymmetric query topology tree with static topological weight attributes submitted by the query rewriting module, extracts the globally unique node identifiers of all independent physical nodes in the asymmetric query topology tree, and inputs the globally unique node identifiers as query terms into the underlying retrieval engine. The underlying search engine matches the corresponding set of document identifiers in the inverted index list. The scoring search module uses global static parameters to perform the first-stage decoupling baseline coarse ranking calculation on the matched set of document identifiers. The global static parameters are the product of the number of document identifiers matched by independent physical nodes in the underlying search engine and the static topology weight. The scoring search module sums the global static parameters for each matched document identifier to obtain the corresponding initial matching score, and sorts them in descending order according to the initial matching scores. It then extracts a preset number of document identifiers from the top of the sorted list and extracts the corresponding original resume text. In this embodiment, the preset number is set to a positive integer between 500 and 2000 to form a set of candidate original resume texts for the second stage of calculation.

[0033] S42, for each original resume text in the candidate original resume text set, the scoring and retrieval module performs topological sorting processing on the asymmetric query topology tree in memory, generating a calculation sequence from implicit skill nodes to the root node set. The scoring and retrieval module reads the continuous binary byte stream data mounted on the inverted index linked list sequentially according to the calculation sequence, and deserializes the continuous binary byte stream data into a set of time-series feature tuples. When no corresponding data exists on an independent physical node, the activity value and comprehensive proficiency value of the independent physical node are set to 0. The scoring and retrieval module calculates the activity value of independent physical nodes based on the paired data of the present time and duration recorded in the time-series feature tuple set. The activity value is obtained by summing the duration recorded in each work experience event block according to the global basic attenuation constant and the present time corresponding to the work experience event block. The activity value is then combined with the edge propagation damping coefficient configured on the directed edge from the parent node to the child node to calculate the dynamic time-related attenuation rate of the parent node. The algebraic expression of the dynamic correction process is defined as follows: ; In the formula, For the calculated parent node The dynamic aging decay rate; In this embodiment, the global baseline attenuation constant is preset for the system. The value range is set to positive real numbers in the open interval from 0 to 1; It is an exponential function with the natural constant as its base; The parent node in the asymmetric query topology tree; Parent node in the asymmetric query topology tree All child nodes pointed to by the directed edge The set that constitutes; Parent node in the asymmetric query topology tree The child node pointed to by the directed edge; From the parent node Point to child node The edge transmission damping coefficient of the directed edge configuration, in this embodiment... The value range is set to positive real numbers in the open interval from 0 to 1; child node The number of work experience event blocks corresponding to the corresponding skill entity; To traverse child nodes In this embodiment, the index of the number of corresponding work experience event blocks is... The value range is from 1 to Positive integers; child node No. The continuous span recorded in each work experience event block; is the natural constant, that is, the base of the natural logarithm; child node No. The time elapsed since the beginning of each work experience event block.

[0034] When an independent physical node has no child nodes or all child nodes have an activity value of 0, the dynamic time-degradation rate of the independent physical node is set to the global basic degradation constant.

[0035] S43, after the scoring and retrieval module updates the dynamic time-effect decay rate of each independent physical node in the asymmetric query topology tree, it applies the updated dynamic time-effect decay rate to perform a proficiency integral operation on the time-series feature tuple set of each independent physical node. The scoring and retrieval module calculates and accumulates the duration recorded by each independent physical node in each work experience event block based on the dynamic time-effect decay rate corresponding to the independent physical node and the time since the work experience event block, to obtain the comprehensive proficiency value of the independent physical node. The calculation logic of the comprehensive proficiency value is defined algebraically as follows: ; In the formula, For the asymmetric query topology tree, the first The overall proficiency value calculated from each independent physical node; For the first The number of work experience event blocks contained in the skill entity corresponding to each independent physical node; For traversing the first In this embodiment, the index represents the number of event blocks corresponding to each independent physical node's work history. The value range is from 1 to Positive integers; For the first The first independent physical node The continuous span recorded in each work experience event block; is the natural constant, that is, the base of the natural logarithm; For the first Dynamic aging attenuation rate of an independent physical node after prior dynamic damping correction; For the first The first independent physical node The time elapsed since the beginning of each work experience event block; In this embodiment, the index for an independent physical node in the asymmetric query topology tree is used. The value range is from 1 to a positive integer representing the total number of independent physical nodes contained in the asymmetric query topology tree.

[0036] S44, the scoring and retrieval module multiplies the comprehensive proficiency value of each independent physical node with the static topological weights assigned in the asymmetric query topology tree, and sums the product results corresponding to all independent physical nodes in the asymmetric query topology tree to obtain the final matching score of the current original resume text in the current job requirement context. The scoring and retrieval module rearranges the candidate original resume text set into a list according to the final matching score, outputs a recommended resume sequence, and provides it to an external data interface.

[0037] Please see the appendix Figure 5 To aid in understanding the technical solution of this invention, the following provides application examples of a semantic retrieval-based intelligent matching method and system for talent resumes.

[0038] Taking a software development job recruitment scenario as an example, the graph construction module builds a skill prerequisite dependency graph covering software development technologies in the graph database. The graph contains directed edges from microservice framework nodes to object-oriented programming language nodes. The graph construction module configures edge transmission damping coefficients on the directed edges. The system receives the input raw resume text, and the index construction module segments the raw resume text into work experience event blocks sorted by time. It extracts microservice framework entities and object-oriented programming language entities from the work experience event blocks, calculates the continuous application span and time to date corresponding to the extracted entities, and serializes the continuous application span and time to date into low-level feature payloads and mounts them into the inverted index.

[0039] The query rewriting module receives external input job requirement description text containing microservice framework entities, extracts the microservice framework entities from the job requirement description text and maps them to the root node of the skill prerequisite dependency graph. Starting from the root node, the query rewriting module performs a level-order traversal along the directed edges towards the lower-level dependencies, and pulls the object-oriented programming language nodes visited during the traversal as implicit skill nodes into the asymmetric query topology tree. The query rewriting module assigns static topology weights to each node according to the tree-like hierarchical depth in the asymmetric query topology tree, and submits the asymmetric query topology tree with static topology weight attributes to the underlying retrieval engine. The underlying retrieval engine matches the corresponding document identifier set in the inverted index linked list and extracts the corresponding candidate original resume text set.

[0040] The scoring and retrieval module performs topological sorting on the asymmetric query topology tree in memory for candidate original resume texts. It reads the continuous application span and time since the beginning of the application for the corresponding object-oriented programming language entities to calculate the activity value. The scoring and retrieval module combines the calculated activity value with the edge propagation damping coefficient configured on the directed edges to calculate the dynamic time-to-earnings decay rate of the microservice framework entities. The scoring and retrieval module applies the calculated dynamic time-to-earnings decay rate to perform proficiency integration on the continuous application span and time since the beginning of the application for each entity to obtain a comprehensive proficiency value. The scoring and retrieval module multiplies the comprehensive proficiency value with the static topology weight and sums them to obtain the final matching score. The scoring and retrieval module reorders the candidate original resume text set according to the final matching score and outputs a recommended resume sequence.

[0041] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result using substantially the same method falls within the scope of the present invention.

Claims

1. A method for intelligent matching of resumes based on semantic search, characterized in that, include: Construct a directed acyclic skill prerequisite dependency graph containing independent physical nodes and directed edges, and configure edge transmission damping coefficients for directed edges; Extract the skill entities, duration, and date of origin from the original resume text, generate a set of time-series feature tuples, and write them into the inverted index; The explicit skill entities in the job requirement description text are mapped to the skill prerequisite dependency graph. The graph is traversed along the directed edges to the lower-level dependency direction to construct an asymmetric query topology tree and assign static topology weights. Based on the nodes of the asymmetric query topology tree, candidate resume texts are extracted from the inverted index. The activity value is calculated based on the continuous span and the time since the child node. The dynamic time decay rate of the parent node is calculated by combining the edge propagation damping coefficient of the directed edge between the parent and child nodes. Based on this, the comprehensive proficiency value is calculated. The final matching score is calculated by combining the static topology weight and a recommended resume sequence is generated.

2. The intelligent talent resume matching method based on semantic retrieval according to claim 1, characterized in that, The directed acyclic skill prerequisite dependency graph containing independent physical nodes and directed edges includes: Obtain the set of skill entities in the technical field, map each skill entity to an independent physical node in the graph database, assign a globally unique node identifier to each independent physical node and store it persistently. Based on the underlying technical dependencies between skill entities, when a high-level skill entity uses a basic skill entity as its prerequisite technical foundation, a directed edge is established from the independent physical node representing the high-level skill entity to the independent physical node representing the basic skill entity. Configure an edge transmission damping coefficient for the directed edge. The edge transmission damping coefficient is used to quantify the extent to which the activity of basic skill entities slows down the decay rate of time forgetting of advanced skill entities. The edge transmission damping coefficient is written into the graph database as attribute data of the directed edge.

3. The intelligent talent resume matching method based on semantic retrieval according to claim 1, characterized in that, The process involves extracting skill entities, duration, and time elapsed from the original resume text to generate a set of time-series feature tuples, including: Named entity recognition processing is performed on the original resume text to identify skill entities, start timestamps and end timestamps, and the original resume text is segmented into a set of work experience event blocks arranged in chronological order. For each work experience event block, the difference between the end timestamp and the start timestamp is used as the continuous span, and the difference between the current timestamp and the end timestamp is used as the time since then. Extract the skill entities from each work experience event block, and combine the current date and duration of the same skill entity in each work experience event block to generate a set of time-series feature tuples consisting of multiple pairs of current date and duration.

4. The intelligent talent resume matching method based on semantic retrieval according to claim 3, characterized in that, The process of generating a set of time-series feature tuples and writing them into an inverted index includes: The time-series feature tuple set is set as a custom composite feature payload in the underlying retrieval engine, and a serialization protocol is used to convert the time-to-date and duration-paired data in the time-series feature tuple set into a continuous binary byte stream. The inverted index uses the globally unique node identifier corresponding to the independent physical node of the skill entity as the key, the document identifier of the original resume text and the continuous binary byte stream data as the corresponding values, and writes them into the inverted index linked list of the underlying retrieval engine according to the key-value pair structure. The document identifier and the continuous binary byte stream data form a custom payload field that is physically bound together.

5. The intelligent talent resume matching method based on semantic retrieval according to claim 1, characterized in that, The traversal along the directed edges towards the lower-level dependencies to construct the asymmetric query topology tree includes: The explicit skill entity is located to the corresponding independent physical node, and the corresponding independent physical node is added to the root node set as the initial query entity node; Using each initial query entity node in the root node set as the traversal origin, a breadth-first hierarchical expansion traversal is performed along the directed edges towards the underlying basic dependencies in the skill prerequisite dependency graph. Read the expansion depth value of the current traversal path and compare the expansion depth value with the preset maximum level threshold. Before the expansion depth value reaches the maximum level threshold, pull the currently visited downstream independent physical node as an implicit skill node into the asymmetric query topology tree. When the expansion depth value reaches the maximum level threshold, truncate the current traversal path.

6. The intelligent talent resume matching method based on semantic retrieval according to claim 5, characterized in that, The allocation of static topology weights includes: Assign initial baseline weights to the initial query entity nodes in the root node set; The tree hierarchy depth of the implicit skill node is determined based on the number of directed edge hops between the implicit skill node and the initial query entity node with the fewest directed edge hops. According to the step decay rule, the static topological weight of each independent physical node in the asymmetric query topology tree is calculated based on the initial baseline weight, the node depth decay constant, and the tree hierarchy depth.

7. The intelligent talent resume matching method based on semantic retrieval according to claim 1, characterized in that, Based on the nodes of the asymmetric query topology tree, candidate resume text is extracted from the inverted index, including: Extract the globally unique node identifiers of all independent physical nodes in the asymmetric query topology tree, and input them as query terms into the underlying retrieval engine to match the corresponding set of document identifiers in the inverted index linked list; The product of the number of document identifiers matched by each independent physical node and the corresponding static topology weight is used as the global static parameter. For each document identifier, the global static parameters corresponding to each independent physical node that matches the document identifier are summed to obtain the initial matching score of the document identifier. Arrange the initial matching scores in descending order, extract a preset number of document identifiers from the top of the list, and extract the corresponding original resume text to form a candidate original resume text set.

8. The intelligent talent resume matching method based on semantic retrieval according to claim 5, characterized in that, The calculation of the dynamic aging rate of the parent node by combining the edge-transmitted damping coefficient of the directed edge between parent and child nodes includes: For the candidate original resume text, perform topological sorting processing on the asymmetric query topology tree in memory to generate a computational sequence from implicit skill nodes to the root node set; Read the continuous binary byte stream data formed by serializing the time-series feature tuple set in the inverted index according to the calculation sequence, and deserialize the continuous binary byte stream data into a time-series feature tuple set. Calculate the activity value of the independent physical node based on the pairing data of the present time and duration recorded in the time-series feature tuple set. The dynamic age decay rate of the parent node is obtained by combining the activity value of the child node with the edge transmission damping coefficient configured on the directed edge of the parent node pointing to the child node. When an independent physical node has no child nodes or the activity value of all child nodes is 0, the dynamic time-related decay rate of the independent physical node is set as the global basic decay constant.

9. The intelligent talent resume matching method based on semantic retrieval according to claim 8, characterized in that, The calculated overall proficiency score includes: After updating the dynamic time-effect decay rate of each independent physical node in the asymmetric query topology tree, the updated dynamic time-effect decay rate is applied to perform a proficiency integral operation on the time-series feature tuple set of each independent physical node. The continuous span recorded by each independent physical node in each work experience event block is calculated and accumulated based on the corresponding dynamic time decay rate and the time since the work experience event block, to obtain the comprehensive proficiency value of each independent physical node. The comprehensive proficiency value of each independent physical node is multiplied by the corresponding static topology weight, and the product results of all independent physical nodes are summed to obtain the final matching score. The candidate original resume text set is then rearranged according to the final matching score, and a recommended resume sequence is output.

10. A talent resume intelligent matching system based on semantic retrieval, characterized in that, include: The graph construction module is used to build a skill prerequisite dependency graph with a directed acyclic graph structure in the graph database, and to configure edge transmission damping coefficients for the directed edges in the skill prerequisite dependency graph. The index building module is used to receive the original resume text, extract skill entities, duration and time to date, generate a set of time-series feature tuples, and serialize the set of time-series feature tuples into continuous binary byte stream data and write it into the inverted index linked list of the underlying retrieval engine. The query rewriting module is used to receive job requirement description text, extract explicit skill entity set, map the explicit skill entity set to the skill pre-dependency graph, traverse along the directed edges in the skill pre-dependency graph towards the underlying dependency direction, construct an asymmetric query topology tree and assign static topology weights. The scoring and retrieval module is used to extract candidate original resume text from the inverted index linked list based on independent physical nodes in the asymmetric query topology tree, calculate the activity value of the child node based on the continuous span and the time since the child node, calculate the dynamic time-lapse rate of the parent node by combining the edge propagation damping coefficient of the directed edge between the parent and child nodes, calculate the comprehensive proficiency value based on the dynamic time-lapse rate, calculate the final matching score by combining the static topology weight, and generate a recommended resume sequence based on the final matching score.

Images