A knowledge graph-based time sequence role recognition method

Abstract

This application discloses a temporal role recognition method based on knowledge graphs, relating to the field of data processing technology. It includes: acquiring first text data, second temporal record data, and third temporal event data of a target object; constructing event units for the target object and dividing them into first, second, and third temporal stages; extracting behavioral features from the first text data and mapping them to event nodes, extracting capability features from the second temporal record data and connecting them to basis nodes, and connecting the event nodes to result nodes to construct a role evidence graph; extracting continuous evidence chains and converting them into stage observation sequences; inputting the sequence inference model to generate role state trajectories, determining the role category based on the trajectory's continuity across the three stages, and outputting classification labels. This method transforms role recognition into a staged inference process based on temporal evidence chains, effectively distinguishing between continuous dominance and partial participation, thus improving the objectivity and accuracy of recognition.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to a time-series role recognition method based on knowledge graphs. Background Technology

[0002] In many scenarios requiring behavioral analysis and role evaluation of target individuals, identifying whether the target individual played a leading role in a specific time-series event is a key basis for judging their ability level and contribution. Relevant information about the target individual typically comes from various heterogeneous data sources, including first-level textual data, second-level time-series record data, and third-level time-series event data. First-level textual data describes the target individual's specific behaviors in the event in an unstructured or semi-structured form; second-level time-series record data reflects the target individual's knowledge reserves and ability development process; and third-level time-series event data defines the complete progression of the event from initiation to completion, including descriptions of each time point and its corresponding events.

[0003] Taking the evaluation of candidates for technical positions as an example, employers need to determine whether a candidate, in their past involvement in events, only undertook partial execution tasks or played a leading role throughout the entire process from requirements analysis and solution design to development, implementation, integration testing, and final delivery. This judgment directly relates to the assessment of the candidate's competency level. However, because the initial textual data submitted by candidates (such as personal resume descriptions) often contains problems such as vague descriptions, mixed responsibilities, or unclear timelines, automatically and accurately identifying the candidate's true role from massive amounts of multi-source heterogeneous data has become a challenging technical problem.

[0004] Existing automated identification methods mostly rely on extracting keywords (such as "responsible" or "leading") from the initial text data or performing static matching of specific skill tags to assign corresponding role tags to the target object. However, these methods have the following significant technical drawbacks.

[0005] First, existing methods lack a verification mechanism for the complete logical chain of "capability input - behavior execution - result output". For example, a target object may claim in the first text data that it "led the design of the distributed system architecture", but the associated second time-series record data does not contain any knowledge reserves related to the architecture design, or the subsequent third time-series event data and the first text data do not show whether the architecture design was implemented or what results it produced. Existing technology cannot effectively identify such breaks in the chain of evidence, thus easily misjudging a target object with only partial execution behavior as having a leading role, resulting in low accuracy of the identification results.

[0006] Secondly, existing technologies typically fail to decompose chronological events into a sequence of stages with clear dependencies. For example, a chronological event can usually be naturally divided into a requirements and solution formulation stage, a development and verification execution stage, and a deliverables and acceptance stage. Existing methods often ignore this chronological structure, making it difficult to distinguish whether the target object only participated in the execution behavior at a certain stage or played a dominant role throughout multiple consecutive stages of the event. This coarse-grained judgment method introduces a lot of noise into the automated screening process and cannot meet the needs of fine-grained identification of role categories.

[0007] Therefore, how to provide a method that can effectively integrate multi-source heterogeneous time-series data, construct a complete logical chain of evidence, and accurately infer the role and state of the target object at each stage of an event has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to provide a knowledge graph-based temporal role recognition method. This method constructs a three-stage event unit that integrates first textual data, second temporal record data, and third temporal event data, unifying scattered heterogeneous information into a temporal framework. It extracts a continuous chain of evidence from behavioral features, ability features, and result nodes, transforming it into a staged observation sequence. A sequence inference model is used to identify role states stage by stage, and a role state trajectory is generated by combining coverage parameters, outputting classification labels. This method transforms role recognition into a staged inference process based on temporal evidence chains, effectively distinguishing between sustained dominance and partial participation, suppressing misjudgments caused by evidence breaks and exaggerated statements, and providing interpretability of classification results traceable back to specific evidence nodes, significantly improving the objectivity and accuracy of recognition.

[0009] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a temporal role recognition method based on knowledge graphs, the method comprising: Obtain the first text data, the second time-series record data, and the third time-series event data of the target object; The first text data and the third time-series event data are segmented and matched at the event level to form target object event units, and the target object event units are divided into the first time-series stage, the second time-series stage and the third time-series stage. Behavioral features are extracted from the first text data within the target object event unit, capability features are extracted from the second time-series record data, and behavioral features are mapped to event nodes at the corresponding stage positions according to the stage division. Capability features that can explain the event nodes are connected as basis nodes, and event nodes are connected to the corresponding result nodes to construct a role evidence graph. Extract a continuous chain of evidence from the role evidence graph, consisting of a basis node, an event node, and a result node in sequence, and convert the continuous chain of evidence into a phase observation sequence corresponding to the first, second, and third time series phases. The phase observation sequence is input into the sequence inference model to identify the role state stage by stage, and the role state trajectory of the target object event unit is generated based on the continuous coverage of the role state at each stage. Based on whether the role state trajectory spans the first, second, and third time-series stages, the role category of the target object event unit is determined, and a classification label for the target object level is generated based on the role category of each target object event unit.

[0010] The beneficial effects of this invention are as follows: The knowledge graph-based temporal role recognition method of this invention constructs a three-stage event unit that integrates first text data, second temporal record data, and third temporal event data, unifying scattered heterogeneous information into a temporal framework; it extracts a continuous chain of evidence from behavioral features, ability features, and result nodes and transforms it into a stage observation sequence; it uses a sequence inference model to identify role states stage by stage, and generates role state trajectories and outputs classification labels by combining coverage parameters. This method transforms role recognition into a staged inference process based on temporal evidence chains, effectively distinguishing between pervasive dominance and local participation, suppressing misjudgments caused by evidence breaks and exaggerated statements, and providing interpretability of classification results that can be traced back to specific evidence nodes, significantly improving the objectivity and accuracy of recognition.

[0011] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a knowledge graph-based temporal role recognition method according to an embodiment of the present invention. Detailed Implementation

[0013] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0014] It should be noted that references to "an embodiment," "embodiment," "example embodiment," etc., in this specification refer to the described embodiment including specific features, structures, or characteristics; however, not every embodiment must include these specific features, structures, or characteristics. Furthermore, such expressions do not refer to the same embodiment. Moreover, when describing specific features, structures, or characteristics in conjunction with embodiments, whether or not explicitly described, it is indicated that incorporating such features, structures, or characteristics into other embodiments is within the knowledge of those skilled in the art.

[0015] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0016] In some embodiments, such as Figure 1 The diagram illustrates a flowchart of a knowledge graph-based temporal role recognition method, which includes the following specific steps: S101, obtain the first text data, the second time-series record data, and the third time-series event data of the target object.

[0017] Specifically, the system retrieves the first text data, second time-series record data, and third time-series event data of the target object from the target object database, the time-series recording system, and the time-series event management system. The first text data is typically unstructured or semi-structured text, containing descriptive information about the target object's behavioral events. The second time-series record data is structured data, containing at least the record name, record content description, and record start and end times. The third time-series event data is a sequence of events with timestamps, recording the key nodes and descriptions of each event from its start to its end.

[0018] S102, perform event-level segmentation and matching on the first text data and the third time-series event data to form target object event units, and divide the target object event units into the first time-series stage, the second time-series stage and the third time-series stage.

[0019] Specifically, the first text data and the third time-series event data are segmented and matched at the event level to form target object event units. Specifically, the matching rule uses both event name and start / end time to associate each event description in the first text data with the corresponding event in the third time-series event data. For example, if the first text data describes an event called "Backend Development of a Data Platform" with a start / end time of March 2022 to September 2022, then the third time-series event data will search for time-series event records with the same name and consistent start / end times for pairing. After successful pairing, the second time-series record data of the target object needs to be associated and filtered. The filtering condition is that only records whose start time is earlier than the end time of the second time-series phase of the associated event and whose completion time is no later than the end time of the third time-series phase of the associated event can be included in the corresponding event description. This time-series constraint ensures that the capabilities applied by the target object are already present before or during event execution, avoiding misjudging records learned after the event has ended as the basis for event capabilities. At this point, a data set that integrates event descriptions, associated records, and corresponding time-series events is defined as the target object event unit.

[0020] After obtaining the target object event unit, its internal third-sequence event data needs to be divided into stages with clear semantics. This embodiment does not simply divide the time-sequence event into three equal segments, but rather divides it based on the semantic content of the event description. Specifically, the third-sequence event data corresponding to the target object event unit is segmented according to time nodes and event semantics. Segments representing the first type of semantics (such as requirements analysis, solution design, and task decomposition) are determined as the first time-sequence stage; segments representing the second type of semantics (such as development implementation, integration testing and verification, and defect repair) are determined as the second time-sequence stage; and segments representing the third type of semantics (such as online deployment, deliverables, and acceptance defense) are determined as the third time-sequence stage. For example, in the time-sequence events, March 10, 2022, the completion of the requirements review meeting is classified as the first time-sequence stage; June 15, 2022, the completion of the payment module integration testing is classified as the second time-sequence stage; and September 1, 2022, the official system launch and handover to maintenance is classified as the third time-sequence stage. In this way, the event progression is solidified into three fixed stages: the first, the second, and the third, providing a unified timeframe for subsequent steps.

[0021] The beneficial effect of this approach is that by integrating and unifying three types of heterogeneous data—first, textual data; second, time-series record data; and third, time-series event data—into a three-stage framework at the event level, it solves the problems of data dispersion and chaotic temporal relationships in existing technologies. This allows subsequent role analysis to be conducted within a clearly defined event context with well-defined temporal relationships, laying the foundation for improving the accuracy of identification.

[0022] S103, extract behavioral features from the first text data within the target object event unit, extract capability features from the second time-series record data, and map the behavioral features to event nodes at the corresponding stage positions according to the stage division, connect the capability features that can explain the event nodes as basis nodes, connect the event nodes to the corresponding result nodes, and construct a role evidence graph.

[0023] Specifically, after constructing the event units and dividing them into stages, it is necessary to further extract key elements related to the target object's role from the text and structured records, and organize them into a computable knowledge graph, namely, a role evidence graph. The construction process of this graph mainly includes two parts: node generation and relationship edge construction.

[0024] The first step is node generation. Action semantic parsing is performed on the descriptive fragments belonging to the event within the first text data of the target object event unit. Here, a behavioral feature category table can be predefined, containing 32 common behavioral feature categories such as requirements analysis, system design, database design, interface development, unit testing, performance optimization, and deployment and maintenance. The parsing process includes extracting verb phrases and object phrases and combining them into behavioral features. For example, from the descriptive fragment "designed and implemented the database structure of the order system," the verb phrases "design" and "implementation," and the object phrases "order system" and "database structure" are extracted and mapped to the two categories "system design" and "database design" in the behavioral feature category table, thus forming two specific behavioral features. Simultaneously, result phrases, such as "completed stress testing under high concurrency scenarios," are extracted as candidate results.

[0025] Next, based on the semantic content of the descriptive fragment from which the behavioral features originate, they are divided into the first, second, or third temporal stage determined in the preceding steps, forming event nodes carrying stage location labels. For example, the behavioral features of the aforementioned system design are divided into the first temporal stage, while the database design is divided into the second temporal stage. Thus, the behavioral description in the first text data is transformed into event nodes with stage attributes.

[0026] In generating capability feature nodes, the second time-series record data associated with the target object event unit is extracted. This second time-series record data includes a record name and a record content description. Capability features are formed by matching this data with a preset capability feature category table (which can also contain 32 categories). For example, if a record is named "Principles and Practices of Distributed Systems" and its content description includes terms such as "consistency protocol" and "distributed transaction," then this record is mapped to the capability feature node "Distributed System Design."

[0027] Secondly, the construction of the relation edge is crucial. When it is determined that a certain capability feature has an explanatory relationship with the behavioral feature of a certain event node, and the completion time of the capability feature is no later than the stage in which the event node is located, a basis relation edge is constructed from the capability feature node to the event node. An explanatory relationship means that the recorded content constitutes the theoretical or practical basis for executing the behavioral feature at the knowledge level. For example, the capability feature of distributed system design can explain the event node of the microservice architecture split in the first time-series stage. The time-series constraint is guaranteed by comparing the record completion time with the stage end time. When multiple capability features can explain the same event node, this embodiment will further select the capability feature with the highest semantic overlap with the behavioral feature category of the event node and the corresponding event description in the third time-series event data as the sole basis node, thereby suppressing redundancy and noise in the evidence chain.

[0028] Similarly, when a result candidate is determined to be located after an event node in the stage sequence, and the result candidate represents the result of that event node, an output relationship edge is constructed from that event node to the result candidate. To ensure the consistency of results across stages, it is also necessary to perform consistency verification between the category of the result candidate and the stage position of the event node. The rules are as follows: event nodes in the first time sequence can only connect to result candidates representing the first type of result semantics (such as requirements, solutions, or task decomposition results); event nodes in the second time sequence can only connect to result candidates representing the second type of result semantics (such as development, integration testing, or defect repair results); and event nodes in the third time sequence can only connect to result candidates representing the third type of result semantics (such as deployment, deliverables, or acceptance results). For example, an interface development event node in the second time sequence must connect to result candidates that are products of the second time sequence, such as completed interface code of module XX or passed integration testing, and cannot be products of the third time sequence, such as system deployment reports.

[0029] After the initial construction of nodes and edges is completed, a double verification is required on the evidence structure composed of candidate capability features, event nodes, and result candidates to generate the final role evidence graph. The first verification is integrity verification, which confirms that the evidence structure contains all three types of nodes and that both relational and output relational edges exist. The second verification is stage continuity verification, which confirms that the stage positions of each node in the evidence structure do not show any reversal in the order of the first, second, and third time-series stages. Only the nodes and edges corresponding to the evidence structures that pass the double verification are included in the role evidence graph.

[0030] The beneficial effect of this approach is that the role evidence graph constructed in this invention is not a simple term co-occurrence network, but a temporal knowledge graph that integrates the triple constraints of formation conditions, stage position, and result orientation. It successfully instantiates the abstract role formation process of ability input – behavior execution – outcome output into a directed graph structure that can be queried and computed by machines. Through a rigorous verification mechanism, it effectively filters out potentially exaggerated statements and temporal discrepancies in the initial text data, ensuring that the evidence entering subsequent judgment stages has high credibility and interpretability.

[0031] S104: Extract a continuous chain of evidence from the role evidence graph, consisting of a basis node, an event node, and a result node in sequence, and convert the continuous chain of evidence into a phase observation sequence corresponding to the first, second, and third time series phases.

[0032] Specifically, after obtaining the character evidence map, the next goal is to extract the strongest evidence from the map and convert it into a numerical vector form suitable for sequence model processing.

[0033] First, based on a relational graph convolutional network, the features of each node are aggregated and encoded along the edges of the basis relationships, output relationships, and stage sequence relationships in the role evidence graph, generating graph encoding vectors for each node. A relational graph convolutional network is a neural network specifically designed for processing heterogeneous graph data. It can aggregate information from neighboring nodes using different parameter matrices depending on the type of the edge. In this embodiment, the initial feature vector of each node is formed by concatenating its type (basis node, event node, result candidate), stage position (first, second, and third temporal stages), and its category (e.g., 32 behavioral features, 32 ability features, and 26 result categories) after one-bit effective encoding. Through propagation through two layers of relational graph convolutional layers, the final graph encoding vector of each node not only contains its own attribute information but also incorporates its contextual structure information within the evidence chain. For example, the encoding vector of an event node will incorporate its preceding ability feature information and its subsequent result information.

[0034] Then, based on the stage position of each node, the node graph encoding vectors belonging to the first, second, and third time-series stages are aggregated respectively. Aggregation can be achieved by summing all node vectors within the same stage and then averaging, or by performing max pooling. This generates the first, second, and third observation vectors. Finally, these three observation vectors are arranged according to the time sequence of the first, second, and third time-series stages, forming a 3×D tensor (D being the dimension of a single observation vector), i.e., the stage observation sequence. This sequence is the direct input to the subsequent Markov model.

[0035] The beneficial effect of this approach is that, through the encoding of graph neural networks, discrete and heterogeneous graph-structured evidence is compressed into a fixed-length, context-rich continuous vector sequence. This transformation not only preserves the core semantic and structural information in the evidence chain but also enables it to be consumed by probabilistic models such as Markov models, which are adept at handling sequential data, thus bridging the gap between symbolic knowledge representation and numerical probabilistic inference.

[0036] S105, input the stage observation sequence into the sequence inference model, identify the role state stage by stage, and generate the role state trajectory of the target object event unit based on the continuous coverage of the role state at each stage.

[0037] Specifically, after obtaining the three-stage observation sequence, a sequence inference model is needed to identify the role and state of the target object at each stage of the event. This embodiment uses a Markov model containing two types of latent variables, the first state and the second state, to accomplish this task.

[0038] First, the stage observation sequence is input into the Markov model. The model contains an observation mapping layer, typically implemented using a multilayer perceptron, which maps the observation vectors of each stage to observation scores for the first and second states. The observation scores reflect the probability of determining whether the target object is in the first or second state based solely on evidence from the current stage, without considering the correlation between preceding and subsequent stages.

[0039] Next, the model, using a transition matrix representing the probability of state transitions between adjacent stages, sequentially calculates the posterior probability vectors for each stage belonging to the first and second states. The transition matrix is ​​a 2×2 matrix containing four transition probabilities: first state → first state, first state → second state, second state → first state, and second state → second state. These probabilities are learned through training on a large amount of labeled event data, reflecting the prior knowledge that role states are usually continuous. For example, a target object in the first state in the first time series usually has a high probability of remaining in the first state in the second time series. By combining the current observation score with the probability from the previous stage, the model can make smoother sequence judgments that better align with business logic. Based on the calculated posterior probability vectors, if the posterior probability of a stage belonging to the first state is greater than a preset threshold (e.g., 0.5), the stage is marked as the first state; otherwise, it is marked as the second state.

[0040] After obtaining the binarized labels for the three stages, continuity and coverage statistics are also required. The number of stages continuously labeled as the first state is counted to obtain the continuity length; then, it is determined whether the first-state labels cover the three key positions of the first, second, and third time-series stages. These two indicators together constitute the coverage parameter.

[0041] Finally, the markers for each stage are corrected based on a preset coverage threshold to form the final role state trajectory. For example, the coverage threshold can be set to 3, requiring simultaneous coverage of the first, second, and third time-series stages. When the coverage parameter meets this threshold, the markers for each stage are directly determined as the role state trajectory; when it does not, it indicates that there is a breakpoint in the evidence of the first state throughout the entire event flow. In this case, the markers for stages not covered by the coverage parameter are corrected to the second state to form a more accurate role state trajectory.

[0042] The beneficial effect of this approach is that it does not simply classify events as a whole into single roles, but rather refines role identification to different stages of the event and uses a Markov model to characterize the temporal evolution of role states. Combined with a post-processing mechanism that considers continuous coverage length and coverage threshold, it can effectively distinguish between three role types with vastly different values ​​in actual classification scenarios: partial participation, phased responsibility, and sustained dominance. This significantly improves the granularity and accuracy of role identification.

[0043] S106. Based on whether the role state trajectory spans the first, second, and third time-series stages, determine the role category of the target object event unit, and generate a classification label for the target object level based on the role category of each target object event unit.

[0044] Specifically, after generating the role state trajectory of each target object event unit, the final judgment and tag generation stage begins.

[0045] First, determine the role category for each single event. Based on the role state trajectory of each target object event unit, check whether it completely covers the first, second, and third time-series stages. If the role state trajectory of an event is marked as the first state in all three stages, then the target object event unit is assigned a first role identifier (e.g., represented by the value 1); otherwise, if any stage is marked as the second state, then a second role identifier is assigned (e.g., represented by the value 0).

[0046] Next, target object-level aggregation is performed. Since the initial text data of a target object typically contains multiple event experiences, the identification results of all target object event units belonging to the same target object need to be arranged in chronological order. This forms a chronologically arranged sequence of 0s and 1s, which is the target object role representation. This representation not only reflects whether the target object has a first role experience, but also reflects the temporal distribution of its first role experiences.

[0047] Finally, classification labels are generated. This is a binary decision label for classification scenarios. Specifically, in response to the presence of at least one first role identifier in the target object's role representation (i.e., at least one 1 in the sequence), a first classification label is output; otherwise, a second classification label is output. This label can be directly consumed by downstream data processing systems for automated filtering or classification.

[0048] The beneficial effect of this approach is that the final output of this invention is not merely a general score or rating, but a complete and interpretable chain of results, from the event level to the target object level. Every step from the initial text data and subsequent time-series record data to the final classification label is traceable. Users can not only see the classification result for the target object, but also trace back to see which specific event the conclusion was based on, at which stages of that event it exhibited its first state, and which specific ability characteristics and result candidates supported these judgments. This greatly enhances the transparency and credibility of the automated classification results, providing a powerful decision support tool for various time-series role recognition-based application scenarios.

[0049] In another embodiment, the above method is illustrated using a recruitment scenario as an example.

[0050] Specifically, when recruiting for technical positions, a company needs to automatically filter out candidates who played a leading role in past events (projects) from a massive amount of data submitted by target candidates. The data submitted by target candidates includes: first, text data (resume text), which describes the names, start and end times, and descriptions of multiple events in which they participated; second, time-series data (course learning records), which includes course names, course content, start and end times; and third, time-series event data (project timelines), which records key time nodes and corresponding descriptions of each event.

[0051] The system first performs data acquisition and event unit construction. For target object A, its first text data contains two event descriptions: event P1 (development of an e-commerce platform order system, March 2022 to August 2022) and event P2 (construction of a data analysis platform, September 2022 to January 2023). The system uses the rule of matching both event name and start / end time to establish a relationship between the event descriptions in the first text data and the corresponding event timelines in the third time-series event data. Simultaneously, the system performs time-series filtering on the second time-series record data of target object A, incorporating events such as the distributed system design course completed in February 2022 and the database performance optimization course completed in May 2022 into event units that meet the time-series constraints. For example, the start time of the distributed system design course is earlier than the end time of the second time-series phase of event P1, and the completion time is no later than the end time of the third time-series phase of event P1; therefore, this course is incorporated into the unit of event P1. Thus, two target object event units are formed for target object A.

[0052] Next, the system divides each event unit into stages. Taking event P1 as an example, the system reads the time nodes and event descriptions from its third-sequence event data: requirements review completed on March 10, 2022; solution design completed on March 20, 2022; development and integration testing conducted from April to July 2022; online deployment completed on August 1, 2022; and project acceptance defense completed on August 15, 2022. The system segments according to the semantics of the events, determining the requirements review and solution design segments as the first time-sequence stage (initial stage), the development and integration testing segments as the second time-sequence stage (implementation stage), and the online deployment and acceptance defense segments as the third time-sequence stage (delivery stage). Thus, event P1 is solidified into a framework containing three consecutive time-sequence stages.

[0053] When constructing the role evidence graph, the system performs semantic analysis on the first text data description fragment corresponding to event P1. From the description of being responsible for the architecture design of the order system and completing the core code implementation, the verb phrases "architecture design" and "code implementation," and the object phrases "order system" and "core code" are extracted and combined into behavioral features, which are mapped to the two categories of system architecture design and core module development, respectively. At the same time, result phrases such as "the system was successfully launched and supported the peak traffic of Double Eleven" are extracted as result candidates. Based on the semantic description, the architecture design behavior is divided into the first time sequence stage, and the core module development behavior is divided into the second time sequence stage, forming event nodes with stage position labels.

[0054] The system further processes the associated second-series record data. The course content of the Distributed System Design course includes terms such as microservice decomposition and distributed transactions, which are mapped to the capability feature node of distributed system architecture capability. The system determines that this capability feature has an interpretive relationship with the architecture design event node in the first time-series phase, and that the course completion time (February 2022) is earlier than the end time of the first time-series phase (March 20, 2022). Therefore, a basis relationship edge is constructed from this capability feature to this event node. Simultaneously, the system verifies the phase order and result direction of the result candidates: the result candidate for successful system deployment is located in the third time-series phase, later than the core module development event node in the second time-series phase, and represents the result of that event node. Therefore, an output relationship edge is constructed from the core module development event node to this result candidate.

[0055] After verifying the completeness of the evidence structure and the continuity of the stages, the system confirmed that the distributed system architecture capability → architecture design event → (no direct output, not constituting a complete link) does not meet the completeness requirement. However, in the link of distributed system architecture capability → core module development event → system successful launch result candidate, based on the capability characteristics of the node in the first time sequence stage, the event node in the second time sequence stage, and the result candidate in the third time sequence stage, the stage positions advance sequentially without reversal. Having passed dual verification, it is included in the valid link of the role evidence graph. Therefore, the role evidence graph is completed, as shown in the appendix. Figure 1 As shown in the figure, each node and its relational edges are clearly displayed.

[0056] Subsequently, the system extracts the continuous chain of evidence and converts it into a phased observation sequence. A relational graph convolutional network is used to aggregate and encode the features of each node along the edges of the dependent relation, the output relation, and the phase sequence relation, generating a graph encoding vector for each node. The encodings of nodes belonging to the first temporal phase are then aggregated into the first observation vector, those belonging to the second temporal phase into the second observation vector, and those belonging to the third temporal phase into the third observation vector, arranged chronologically to form the phased observation sequence. This sequence preserves the structural and semantic information of the evidence chain in the form of numerical vectors.

[0057] The system inputs the observation sequence of this stage into a Markov model with the first state (dominant state) and the second state (participating state) as latent variables. The model converts each observation vector into a state observation score through an observation mapping layer, and calculates the posterior probability of each stage in sequence with the transition matrix. For event P1, the calculation results show that the probability of the first time series stage belonging to the first state is 0.82, the second time series stage is 0.91, and the third time series stage is 0.76, all exceeding the threshold of 0.5, and all three stages are marked as the first state. Statistical analysis shows that the continuous length covers all three stages, the coverage parameter meets the coverage threshold (set to 3), and the finally generated role state trajectory is a continuous sequence of three first states.

[0058] Based on the fact that the role's state trajectory completely spans the first, second, and third time-series stages, the system determines event P1 as the first role category (dominant event) and assigns it a first role identifier. For event P2, its role state trajectory shows it is only marked as the first state in the second time-series stage, and as the second state in both the first and third time-series stages; therefore, the system determines it as the second role category (participating event) and assigns it a second role identifier. The system arranges all event identifier results for target object A chronologically to form a target object role representation. Since this representation contains at least one first role identifier, the system ultimately outputs a first category label for target object A, indicating that it meets the initial job screening criteria.

[0059] In this embodiment, the system successfully and accurately identified event P1 as the dominant event and event P2 as a participating event from two events involving target object A. Even though target object A used words such as "responsible" in the description of both events in the first text data, the system was still able to reveal, through evidence chain verification and phase coverage analysis, that it only participated in the implementation phase of event P2 and did not dominate the entire process. This demonstrates the technical advantage of this method in distinguishing between continuous dominance and partial participation. The output classification labels have high interpretability and traceability, providing reliable technical support for automated screening decisions.

[0060] For example, project records in the resume text can be represented as ,in, This indicates the number of project records in the resume text. , Indicates the first The project name recorded in each project. Indicates the project start time. Indicates the project end time. This indicates a project description, representing course records as... ,in, Indicates the number of course records. , Indicates the course name. Indicates course content, Indicates the start time of the course. Indicate the course completion time, and represent the project timeline as follows: ,in, Indicates the number of timelines in the project. , Indicates the project name. Indicates the order in chronological order. Each time point Indicates the first Description of the items corresponding to each time point. This indicates the number of time points included in the project's timeline; Record each project Construct matching set ,when At that time, the timeline of the single project will be recorded as ,when At that time, take And record the corresponding project timeline as ,when At that time, no project correspondence is established, thus obtaining the project description. Corresponding project timeline The one-to-one correspondence is maintained, and project-level segmentation only retains project records that simultaneously satisfy the pairing rules, including both the project name and the project start and end time. The semantic word set of the initial stage is defined as the requirement proposal, solution design, and task decomposition. The semantic word set for the implementation phase is defined as development, implementation, joint debugging, verification, and defect repair. The semantic term set for the delivery phase is defined as the online deployment results delivery, acceptance, and defense. ,right Each item description Perform stage semantic determination to obtain stage marker sequence ,in, , Indicates the initial stage marker. Indicates the implementation stage marker, The delivery stage markers are used to indicate the delivery stage. When the same item description contains multiple stage semantic terms, the item description is first divided into sub-descriptions according to punctuation and conjunctions. Then, stage markers are assigned to each sub-description, while keeping the original time nodes unchanged. The implementation stage end time is read according to the stage marker sequence. and delivery phase end time Reconstruct a subset of courses Course subset Only course records with a start time earlier than the end time of the implementation phase and a completion time no later than the end time of the delivery phase are retained, so that the course records incorporated into the project description are within the same project progress time boundary as the corresponding project. The project timeline with stage markers is divided into continuous segments to obtain a set of time segments. ,in, Indicates the number of time segments. , Indicates the start time of the time segment. Indicates the end time of the time segment. Indicates a fixed stage position. This represents the set of item descriptions corresponding to this time segment. When adjacent item descriptions have the same stage marker, they are merged into the same time segment. When adjacent item descriptions have different stage markers but are consecutive in time, they are formed into different time segments. The time segment is determined as the initial stage time segment, and The time segment is determined as the implementation phase time segment, and The time segment is determined as the delivery phase time segment only when When a project simultaneously contains initiation phase time segments, implementation phase time segments, and delivery phase time segments, construct the project phase sequence. ,in, Indicates the initial stage. Indicates the implementation stage. Indicates the delivery phase, project phase sequence It is not a simple segmentation of the project timeline, but a fixed expression of the three stages in the project process, so that technical actions, course capabilities and project outputs have a locatable sequential relationship within the same stage framework; Define the candidate project unit as ,in, Provide a project description. Provide course records that align with the project's timeline. Provide project timeline, Provide time segments with fixed stage positions. Provides a sequence of project phases, including the initiation phase, implementation phase, and delivery phase, and candidate project units. The scenario input is organized around individual projects, rather than simply stacking resume texts, course records, and project timelines side by side. Instead, it is organized by project phase sequence. To establish the organizational framework, project descriptions, course records, and project timelines are constrained to the same project scope, with multiple candidate project units forming the input set. ,in, The input set represents the number of projects that satisfy the project correspondence and three-phase integrity conditions. Each Both can be directly used for the phased mapping of technical actions and the construction of role evidence diagrams.

[0061] Read project descriptions from candidate project units. Corresponding course records Corresponding project timeline and project phase sequence ,in, Indicates the initial stage. Indicates the implementation stage. Indicates the delivery phase, for Segmenting by punctuation and conjunctions yields a sequence of project description segments. , This indicates the number of project description fragments, for each project description fragment. Perform part-of-speech tagging and dependency parsing to extract verb phrases. object phrase and task result phrases Only retain project description fragments that directly correspond to project goals, task breakdown, feature implementation, or deliverables, and exclude verb phrases. With object phrase Matching is performed in a preset table of technical action categories. The number of verb terms and object terms that match for each category is counted, and the category with the highest total number of matches is selected as the technical action category. , If the total number of hits is the same, the categories of those that hit both verb and object terms are retained. If they are still the same, the category with the earlier hit in the preset technical action category table is selected, thus obtaining the technical action. , the task result phrase Match the preset output category table and obtain candidate descriptions of project outputs according to the same rules. , Indicates the candidate description category for project outputs; from Description of Reading Phase Phase End Time and item category codes ,in, , This indicates that the request has been made. Indicate the design scheme. This indicates task decomposition. Indicates development and implementation. This indicates joint testing and verification. Indicates defect repair. This indicates that the deployment has been completed. Indicates the delivery of results. This indicates the acceptance defense, which will cover each technical action. and Perform semantic matching, when the technical action corresponds When representing the semantics of a matter, the technical action is positioned at a specific stage. , Indexing technical actions within the same stage location by project description fragments. Sort in ascending order to get the order in which project events occur within their respective phases. Based on this, project event candidates are generated. Construct rules for corresponding items Each row consists of a technical action category code and an event category code, constructing corresponding rules. Each row consists of a technical action category code and a project output candidate description category code, when At that time, the project event candidates are kept in a consistent correspondence with the event descriptions in the project timeline. ,satisfy ,and or The stage position is later than When the descriptions of project event candidates and project output candidates form a consistent connection, project event candidates that meet the consistency correspondence judgment or the continuity consistency judgment are assigned a reservation mark. If not satisfied, assign , only Write project event candidates to project events ; Course records For each course record, extract the course name, course content, and course completion time to obtain... ,in, , Indicates the number of course records. Indicates the course name. Indicates course content, Indicate the course completion time, including the course name. With course content Match the terms in the preset course ability category table, count the number of matched terms in each category, and select the category with the highest number of matched terms as the course ability category. , Then the course content Matching is performed in the preset semantic vocabulary to obtain the semantic code of the course content. Construct a preset interpretation mapping rule Each row consists of a course competency category code and a technical action category code, constructing the item correspondence rules. ,in Each line consists of a semantic code for course content and a code for an item category. , and At that time, course capabilities are linked to project events and identified as project criteria. When the same project event corresponds to multiple course capabilities that meet the conditions, the course content is statistically analyzed. Project description fragment Description of each stage Based on the number of common technical terms, only the course with the largest number of common technical terms will be retained as the basis for the project. If the number of common technical terms is the same, the course with the closest completion time and no later than [previous course] will be retained. The course competence serves as the basis for the project. Therefore, the project basis that can explain the project events is limited to course competence nodes that simultaneously satisfy the constraints of corresponding formation conditions, semantic correspondence of events, and sequential stage. From project output candidate descriptions Project Timeline Extract project outputs from the deliverables to obtain ,in, , Indicates the quantity of project output. Indicates the output category, Indicates the stage or position of the project's output. This indicates the order in which project outputs appear in their respective stages, and constructs result mapping rules. The third column is the stage position code. When the stage position code is... At that time, only results of requirement submissions, solution design results, or task decomposition results are allowed, when the stage position is coded as At that time, only development and implementation results, integration and verification results, or defect repair results are allowed, when the stage position is coded as follows: At that time, only deployment results, deliverables, or acceptance defense results are allowed. The closed-loop decision value of project basis, project events, and project outputs is expressed as: And calculate using the following formula: ; In the formula, Indicates by the first The project is based on, the first The project event and the first The criteria for determining whether the output of a project constitutes a closed loop. Indicates the first Reservation markers for individual project events This represents an indicator function, which takes the value when the condition is true. When the condition is not met, the value is taken as... , Indicates the course competency category code, Indicates the technical action category code, This indicates the preset interpretation mapping rules. This represents the semantic encoding of the course content. This indicates the item category code corresponding to the current stage of a project event. Indicates the rules corresponding to the items. Indicates the course completion time. This indicates the end time of the phase corresponding to the location of the project event. Indicates the output category code, Indicates the result mapping rules, Indicates the stage of project output. Indicates the stage or position of a project event. This indicates the order in which project outputs appear within their respective stages. This indicates the order in which project events occur within their respective phases. Represents logical OR, This represents logical AND, when At that time, establish an effective link ,when At that time, no link is established; After deduplicating the project data, project events, and project outputs in all valid links, a node set is formed. The explanatory relationships between the project and the project events are used to form an edge set. The connection between project events and project outputs is formed into an edge set. The sequential relationships formed by the stage positions within the same valid link are used to form an edge set. This constitutes a character evidence diagram. Generate a node feature vector for each node. , among which the former The dimension is a single valid code for the stage position, and the middle is... The dimension is a single valid encoding for the node type, followed by The dimension is a valid code for the technical action category, and then... The last valid code is a single-bit code for the course competency category. Each dimension represents a valid output category code. The technical action category features and output category features of project nodes are filled with zeros. The course ability category features and output category features of project event nodes are filled with zeros. The technical action category features and course ability features of project output nodes are filled with zeros. All node feature vectors are stacked in node order to form a node feature matrix. Graph coding networks As input, a two-layer relation graph convolutional layer is used to aggregate the features of adjacent nodes according to the explanatory relation edges, the continuation relation edges, and the sequential relation edges, respectively. The first layer will... Dimensional node feature mapping is Dimensional graph encoding matrix The second layer will The dimensional graph encoding matrix is ​​mapped to 3D node-level graph encoding matrix Under the constraints of the project phase sequence, a role evidence diagram is formed by connecting project evidence, project events, and project outputs, which specifically corresponds to a node feature matrix. With the set of three types of relation edges , , This structure is directly used as input to the graph coding network.

[0062] Represent the character evidence diagram as follows: ,in, Represents a sequence of nodes. , Indicates the number of nodes. This represents the set of explanatory edges relating the project to the project events. This represents the set of edges representing the connections between project events and project outputs. This represents the set of edges indicating the sequential relationships within a project phase sequence, where each node is represented as... ,in, Indicates the node type. Indicates the project based on nodes, Indicates project event nodes, Indicates the project output node, Indicates the stage position. Indicates the starting stage position. Indicates the stage of implementation. Indicates the delivery stage location. This indicates the technical action category code, and the project event node is taken from... to One of the codes, non-project event nodes take , This indicates the course competency category code, and the project is selected based on the node. to One of the codes, not the project's node-based code. , This indicates the output category code, and the project output node is taken from... to One of the codes, taken from non-project output nodes. Then read the project phase sequence. ,in, Indicates the initial stage. Indicates the implementation stage. Indicates the delivery stage, used to fix the aggregation order of stages; Set along the explanatory relation edge and the set of edges with succession relationship Enumerate the three-node paths to obtain the candidate link sequence. ,in, , Indicates the number of candidate links. , This indicates that the project is based on the node index. Indicates the index of the project event node. This represents the index of the node output by the project. Integrity checks are performed on each candidate link, and the check content is... , , ,and , For link candidates that pass the integrity check, the continuity check of the execution phase is performed, and the check content is as follows: ,as well as ,when At that time, it was required ,when At that time, it was required And the stage position is according to The project proceeds sequentially through its phases, retaining only candidate links that simultaneously satisfy both integrity and phase continuity checks as continuous evidence chains. ,in, , Indicates the number of consecutive chains of evidence. Therefore, a continuous chain of evidence is defined as a three-node link that simultaneously includes project basis, project events, and project outputs, and whose explanatory relationship, successive relationship, and phase progression sequence are all valid. By removing duplicate nodes from the entire chain of continuous evidence by index, we obtain the node sequence of the continuous evidence subgraph. ,in, , This indicates the number of nodes in the continuous evidence subgraph, and writes the explanatory relationship edges that appear in the continuous evidence chain into the graph. Write the edges of the successive relationships that appear in the continuous chain of evidence into Write the sequential relationship edges that satisfy the stage order in the continuous chain of evidence into Forming a continuous evidence subgraph For each node Reading inherited properties , , , and And generate node feature vectors ,in, The stage position feature is represented by a three-dimensional one-dimensional effective encoding, with the starting stage position written as... The implementation stage position is written as The delivery stage location is written as , Indicates node type characteristics; the project is written based on the node. Project event nodes are written as Project output nodes are written as , Indicates the category characteristics of technical actions, only when At the time Dimensional writing The remaining dimensions are written , Indicates the characteristics of course competency categories, only when At the time Dimensional writing The remaining dimensions are written , Indicates output category characteristics, only when At the time Dimensional writing The remaining dimensions are written All node feature vectors are sorted according to The node feature matrix is ​​obtained by stacking the nodes in ascending order. ; Represent the graph coding network as ,in, This represents the relational transformation parameters and linear mapping parameters of the two-layer relational graph convolutional layer. The input graph encoding network, the first relational graph convolutional layer, gathers the features of adjacent nodes with explanatory relational edges, the features of adjacent nodes with inheriting relational edges, and the features of adjacent nodes with sequential relational edges for each node, and performs a linear mapping with the current node's own 96-dimensional node features. The first relational graph convolutional layer contains... A fully connected neuron is used to obtain the first-layer graph encoding matrix. ,in, The second layer of the relational graph convolutional layer uses As input, the same relational propagation method is used for re-encoding. The second relational graph convolutional layer contains... A fully connected neuron is used to obtain a node-level graph encoding matrix. ,in, Indicates the first The node-level graph encoding vector of each node is then used to construct node subsequences according to their stage positions. ,in, For each stage position Read all corresponding node-level graph encoding vectors Summing is performed dimension by dimension along the sixty-four-dimensional scale, and then divided by the number of nodes at that stage. The stage observation vector is obtained. ,when At that time, Set as a 64-dimensional zero vector, according to Arranged in order , and Forming a phased observation sequence ,in, It is a three-step sequence with a sequence length of . Each sequence element is A dimensional vector can be directly used as input to a Markov model.

[0063] For example, the phased observation sequence Input the Markov model in a fixed order of initiation phase, implementation phase, and delivery phase, where, Indicates the first Stage observation vectors for each stage, Indicates the initial stage. Indicates the implementation stage. Represents the delivery phase, phase observation vector The sixty-four components are obtained by summing the node-level graph encoding vectors at the same stage position dimension by dimension and dividing by the number of nodes at that stage. Therefore, all sixty-four components are numerical values ​​that can be directly input into the Markov model. The latent variables of the Markov model are set as the set of character states. ,in, Indicates the dominant state. Indicates participation status, will the first The role state vector for each stage is represented as follows: ,in, Indicates the first The state values ​​of each stage belong to the dominant state. Indicates the first Each stage belongs to the state value of the participating state; An observation mapping layer is used to process the observation vectors at each stage. A phased mapping is performed, and the observation mapping layer consists of two consecutive multi-layer sensing mechanisms. The first layer is... Dimensional input, The output of each fully connected neuron will Compressed into a 32D compressed representation The second layer is Dimensional input, The output of each fully connected neuron will Convert to state observation score vector ,in, Indicates the first Each stage corresponds to a state observation score of the dominant state. Indicates the first The state observation scores for each stage correspond to the participating states. Three sets of state observation score vectors are arranged according to the project stage sequence. The fully connected weights, biases, and initial state priors of the observation mapping layer are also considered. and state transition matrix The training samples were obtained using manually labeled project-dominant state trajectories. During training, the manually labeled data for each stage was written as a two-dimensional, one-dimensional effective label. Forward computation was performed to obtain the role-state vectors for the three stages. The cross-entropy between the prediction results and the manually labeled data was calculated stage by stage and accumulated. Gradient descent was used to update the fully connected weights, biases, initial state priors, and state transition matrices. After each update, the cross-entropy was calculated. The two-dimensional values ​​are normalized to a sum. The initial state prior value, and Each row is unified as the sum. State transition values; A state transition layer is used to perform sequence inference on the three sets of state observation score vectors. In the initial stage, the initial state is used as a priori. and Add the corresponding components, then divide the sum of the two dimensions by the sum of the sums of the two dimensions respectively, to obtain... The implementation phase will As row vectors and state transition matrices Multiply to obtain a two-dimensional stage prediction value, then multiply by... Add the corresponding components and normalize them to get The delivery phase will and Perform the same recursion, then with Add the corresponding components and normalize them to get This recursive process makes the role status of the implementation phase affected by the initiation phase, and the role status of the delivery phase affected by the implementation phase, so that the role status of the three phases reflects the sequential relationship in the project phase sequence. The threshold for determining the dominant state is expressed as: In this embodiment, ,when At that time, the first Each stage is marked as the dominant state, denoted as […]. ,when At that time, the first Each stage is marked as the participation status, denoted as . Arranged according to the project phase sequence to obtain the phase marker sequence The phase length and phase coverage that are continuously in a dominant state are calculated as the dominant coverage parameter. And calculate according to the following formula: ; In the formula, This indicates the starting position of a segment that is continuously in a dominant state. This indicates the end position of a segment that is continuously in a dominant state. Indicates the starting position of the candidate continuous segment. Indicates the end position of the candidate continuous segment. This indicates selecting the candidate continuous segment that maximizes the comparison value. This represents an indicator function, which takes the value when the condition is true. When the condition is not met, the value is taken as... , Indicates the first Stage markers for each stage, This represents the product of the stage markers within a candidate continuous segment, used to ensure that all stages in the candidate continuous segment are in a dominant state. The validity marker indicating the selected consecutive segments. Indicates the length of a phase in which the state remains dominant. Indicates whether the initial phase is overridden by the dominant state. This indicates whether the implementation phase is overridden by the dominant state. Indicates whether the delivery phase is covered by the dominant state. This indicates the number of key stage locations covered by the dominant state. This indicates the stage position index. When there is no dominant state segment in the stage marker sequence, it will be... Set as and will , , , and Set all to ; The critical phase coverage threshold is expressed as: The critical phase coverage threshold is defined as the number of critical positions in the project phase sequence that must be covered by the dominant state. In this embodiment, the initiation phase, implementation phase, and delivery phase are all designated as critical positions, therefore, the threshold is set to... The state output layer reads the stage marker sequence. and dominant coverage parameters ,when At that time, directly Determined as the dominant state trajectory of the project ,in, , Indicates the first This stage is the dominant state. Indicates the first This stage belongs to the participation state, when At that time, the state output layer follows , , Each of the three phases was verified, and the phases not covered by the dominant coverage parameter were marked as participating, thus obtaining the project dominant state trajectory. The Markov model receives an input of length . Phase observation sequence Each sequence element is The observation mapping layer outputs three two-dimensional state observation score vectors, the state transition layer outputs three two-dimensional role state vectors, and the state output layer outputs vectors of length [missing information]. Project dominant status trajectory Read the project's dominant state trajectory Dominant Coverage Parameters and project phase sequence ,in, This indicates the character's status during the initial phase. This indicates the role status markers during the implementation phase. This indicates the role status flag during the delivery phase, with possible values. Indicates the dominant state, with values... Indicates participation status. Indicates the length of a phase in which the state remains dominant. Indicates whether the initial phase is covered by the phase coverage area. Indicates whether the implementation phase is covered by the phase coverage scope. Indicates whether the delivery phase is covered by the phase coverage scope, with possible values. Indicates overwrite, value Indicates that it is not covered. Indicates the number of stage locations covered by the stage coverage area. Indicates the initial stage. Indicates the implementation stage. Indicates the delivery phase, according to Read in a fixed order , , This yields the role status arrangement corresponding to the initiation phase, implementation phase, and delivery phase, and simultaneously... The length of the stage that is continuously in the dominant state is directly extracted from the data. and phase coverage ; Perform a phase-by-phase coverage check on the phase coverage against the initiation phase, implementation phase, and delivery phase, and read the initiation phase. ,when When it is determined that the initial stage has been overwritten, If the initial phase is not covered, read the implementation phase. ,when When the implementation phase has been covered, If the implementation phase is not covered, read the delivery phase. ,when When it is determined that the delivery phase has been covered, If a delivery phase is not covered, a continuity check is performed between the length of the continuously dominant phase and the project phase sequence. Since the project phase sequence fixedly includes three phases: initiation, implementation, and delivery, the continuity check uses three direct values ​​to judge simultaneously: [Reading...] Is it equal to Read Is it equal to Read Is it equal to Read Is it equal to Only when , , , , , , When both are established, the project's dominant state trajectory is determined to meet the project leadership condition. To ensure that the determination result can be directly written into the project unit, a project leadership condition determination value is set. ,in, When all six checks mentioned above are true, Set as If any one of the above six checks fails, then... Set as ; The project identification result is represented as ,in, and will As a discrete identifier value written to the corresponding project, the dominant project identifier is specifically defined as The write results indicate that the corresponding project is covered by the phase coverage scope in all three phases: initiation, implementation, and delivery, and the project's dominant state trajectory continuously holds the dominant state in all three phases. The participating project identifier is specifically defined as... The write result indicates that at least one phase of the project—the initiation phase, implementation phase, and delivery phase—was not covered by the phase coverage scope, or that the project's dominant state trajectory did not continuously hold the dominant state across all three phases. Mapped to At that time, no new text features, graph structure features, and probability scores are introduced; only the following are used: The binary result completes the assignment of the project identifier. At that time, Set as and assign a leading project identifier, when At that time, Set as The participating projects are assigned identification tags. Thus, the leading project identification tag and the participating project identification tag are not manually interpreted text, but machine-writable binary identifiers directly obtained based on the project's leading state trajectory, stage coverage, and the length of the stage in which it is continuously in the leading state. Will With the corresponding candidate project unit To establish connections and determine project leadership. ,in, Indicates the candidate project unit. This indicates the outcome of project leadership corresponding to the candidate's project unit. hour, The record shows the main project identifier, when hour, Records the project participation identifier and project leadership outcome. It only includes a one-to-one correspondence between candidate project units and their corresponding project identifier values, so that the determination of project leadership can be directly applied to specific candidate project units.

[0064] Reading project leadership results Project Dominant Status Trajectory and project phase sequence ,in, Indicates the first Each candidate project unit contains a candidate number. and project timeline position , Indicates the project identification result. Indicates the leading project identifier. Indicates participation in the project. , , These represent the role status markers for the initiation phase, implementation phase, and delivery phase, respectively. , , These represent the initiation phase, implementation phase, and delivery phase, respectively. , and Project unit with the same candidate Perform project-level association to generate project-level association results. In the project-level association results, the three role status flags are always written in a fixed order of start phase, implementation phase, and delivery phase, thus ensuring that the role arrangement of the same project in the project phase sequence can be directly read. For all project-level association results By candidate number The results are aggregated, and for project-level associations with candidate numbers, they are arranged according to their position on the project timeline. Arrange the candidates in ascending order, and let the candidate number be... The number of projects is Read the first result in order according to the sorting results. The project identification results for each project are denoted as... ,in, , will all Arranged chronologically to form a sequence of candidate-level project identifiers. Each element in the candidate-level project identifier sequence is a binary project identifier result for a single project, thus directly reflecting the distribution of the candidate's dominant project identifier or participating project identifier across various projects. The coverage of the project leadership status trajectory corresponding to each project in the candidate-level project identifier sequence is written into the candidate project leadership representation. The candidate's project leadership representation is specifically defined as a two-dimensional binary matrix arranged in chronological order of the projects. The first of the matrix The row indicates the candidate number is The Each item is written in a fixed order, with the seven columns of the matrix. ,in, , , They represent the first The role status markers of each project in the initiation, implementation, and delivery phases. , , They represent the first The coverage results of each project in the initiation, implementation and delivery phases are shown. Thus, the first column in the candidate project leadership representation represents the project-level role conclusion, the second to fourth columns represent the project leadership status trajectory, and the last three columns from the fourth to the seventh column represent the phase coverage conclusion. Existing technical personnel can directly read the leadership information of each project according to the row and column positions of the matrix. Characterizing the candidate's project leadership Mapped to initial screening tags for technical positions The initial screening labels for technical positions are specifically defined as candidate-level binary labels. ,in, This indicates that the candidate meets the initial screening criteria for the technical position. This indicates that the candidate does not meet the initial screening criteria for the technical position; during mapping, the data is scanned line by line in project timeline order. When there exists any row that satisfies the first column Fifth column Column 6 Column 7 When determining whether a candidate project has a leading project identifier covering the initiation, implementation, and delivery phases, it will... Set as When all rows do not satisfy the above four positions simultaneously taking values At that time, Set as In this way, the initial screening labels for technical positions are not manually interpreted statements, but rather binary judgment results directly calculated from the project identification results in the candidate's project leadership representation and the three-stage coverage results.

[0065] The technical features of the above embodiments can be arbitrarily integrated. For the sake of brevity, not all possible integrations of the technical features in the above embodiments are described. However, as long as the integration of these technical features does not contradict each other, they should be considered to be within the scope of this specification.

[0066] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A temporal role recognition method based on knowledge graphs, characterized in that, The method includes: Obtain the first text data, the second time-series record data, and the third time-series event data of the target object; The first text data and the third time-series event data are segmented and matched at the event level to form a target object event unit, and the target object event unit is divided into a first time-series stage, a second time-series stage and a third time-series stage. Behavioral features are extracted from the first text data within the target object event unit, capability features are extracted from the second time-series record data, and the behavioral features are mapped to event nodes at the corresponding stage positions according to the stage division. The capability features that can explain the event nodes are connected as basis nodes, and the event nodes are connected to the corresponding result nodes to construct a role evidence graph. Extract a continuous chain of evidence from the role evidence graph, consisting of the basis node, the event node, and the result node in sequence, and convert the continuous chain of evidence into a stage observation sequence corresponding to the first time series stage, the second time series stage, and the third time series stage. The stage observation sequence is input into the sequence inference model to identify the role state stage by stage, and the role state trajectory of the target object event unit is generated based on the continuous coverage of the role state at each stage. Based on whether the role state trajectory spans the first time series stage, the second time series stage, and the third time series stage, the role category of the target object event unit is determined, and a classification label for the target object level is generated based on the role category of each target object event unit.

2. The method according to claim 1, characterized in that, The step of dividing the target object event unit into a first timing phase, a second timing phase, and a third timing phase includes: Using the matching rule of both event name and event start and end time, each event description in the first text data is associated with the corresponding event in the third time-series event data; The records in the second time-series record data whose start time is earlier than the end time of the second time-series phase of the associated event and whose completion time is no later than the end time of the third time-series phase of the associated event are incorporated into the corresponding event description to obtain the target object event unit. The third time-series event data corresponding to the target object event unit is segmented according to time-series nodes and event semantics. The fragments representing the first type of semantics are determined as the first time-series stage, the fragments representing the second type of semantics are determined as the second time-series stage, and the fragments representing the third type of semantics are determined as the third time-series stage.

3. The method according to claim 1, characterized in that, The construction of the role evidence graph includes: Semantic parsing is performed on the description fragments of the same target object event unit in the first text data, and verb phrases and object phrases are extracted and combined to form the behavioral features, and the resulting phrases are extracted and marked as result candidates; Based on the semantic content of the descriptive fragment from which the behavioral features originate, the behavioral features are divided into the first time-series stage, the second time-series stage, or the third time-series stage, forming event nodes carrying stage location labels; Extract the record name and record content from the second time-series record data to form the capability feature. When it is determined that the capability feature has an explanatory relationship with the behavioral feature of the event node, and the completion time of the capability feature is not later than the stage of the event node, construct the basis relationship edge from the capability feature to the event node. When it is determined that the result candidate is located after the event node in the stage order and represents the result of the event node, an output relationship edge from the event node to the result candidate is constructed; Using the capability features, event nodes, and result candidates that simultaneously possess the aforementioned basis relationship edge and the aforementioned output relationship edge as nodes, and using the aforementioned basis relationship edge, output relationship edge, and stage sequence relationship as edges, the role evidence graph is generated.

4. The method according to claim 3, characterized in that, The method of constructing the relational edge from the capability feature to the event node further includes: Verify whether the recorded content corresponding to the capability feature constitutes the formation condition of the behavior represented by the event node, and that the completion time of the capability feature is not later than the end time of the stage in which the event node is located. When multiple capability features meet the above verification conditions, the capability feature with the highest semantic overlap with the behavioral feature category of the event node and the corresponding event description in the third time-series event data is selected as the sole basis node.

5. The method according to claim 3, characterized in that, The construction of the output relationship edge from the event node to the result candidate also includes: A consistency check is performed on the category of the result candidate and the stage position of the event node, so that: the event node located in the first time sequence stage is connected only to the result candidate representing the first type of result semantics; the event node located in the second time sequence stage is connected only to the result candidate representing the second type of result semantics; and the event node located in the third time sequence stage is connected only to the result candidate representing the third type of result semantics.

6. The method according to claim 3, characterized in that, The generation of the character evidence diagram further includes: Perform integrity checks and phase continuity checks on the evidence structure composed of candidate capability characteristics, event nodes, and result candidates; The integrity check is used to confirm that the evidence structure contains three types of nodes and that the edge relationships are complete. The stage continuity check is used to confirm that the stage positions of each node in the evidence structure do not appear in reverse order along the sequence of the first time-series stage, the second time-series stage, and the third time-series stage. Only nodes and edges corresponding to evidence structures that pass double verification are included in the role evidence graph.

7. The method according to claim 1, characterized in that, The step of converting the continuous chain of evidence into a phased observation sequence includes: Based on the relational graph convolutional network, the features of each node are aggregated and encoded along the basis relation edges, output relation edges and stage sequence relation edges in the role evidence graph to generate the graph encoding vector of each node. Based on the stage position of the node, the node graph encoding vectors belonging to the first time series stage, the second time series stage, and the third time series stage are aggregated respectively to generate the first observation vector, the second observation vector, and the third observation vector. The first observation vector, the second observation vector, and the third observation vector are arranged in chronological order to form the stage observation sequence.

8. The method according to claim 1, characterized in that, The generation of the role state trajectory of the target object event unit includes: The stage observation sequence is input into a Markov model containing two types of latent variables: the first state and the second state. The observation vectors of each stage are mapped to observation scores for the first state and the second state using an observation mapping layer. By combining the transition matrix that characterizes the state transition probability of adjacent stages, the posterior probability vector of each stage belonging to the first state and the second state is calculated in turn, and each stage is marked as the first state or the second state accordingly. The number of stages consecutively marked as the first state is counted, and whether the first state marking covers the first time-series stage, the second time-series stage, and the third time-series stage is determined, and a coverage parameter is generated. The markers for each stage are corrected based on a preset coverage threshold to form the character state trajectory.

9. The method according to claim 8, characterized in that, The step of correcting the markers at each stage based on a preset coverage threshold to form the character state trajectory includes: When the coverage parameter meets the coverage threshold, the markers of each stage are directly determined as the character state trajectory; When the coverage parameter does not meet the coverage threshold, the stage not covered by the coverage parameter is marked as a second state to form the role state trajectory.

10. The method according to claim 1, characterized in that, The generation of classification tags for the target object hierarchy based on the role category of each target object event unit includes: Based on whether the role state trajectory of each target object event unit completely covers its first time phase, second time phase and third time phase, assign a first role identifier or a second role identifier to each target object event unit. The identification results of all target object event units belonging to the same target object are arranged in chronological order to form a target object role representation; In response to the presence of at least one first role identifier in the target object role representation, a first category label is output.

Images