A method for multi-dimensional information integration and intelligent query of immigration archives

Abstract

This invention relates to the field of computer data processing technology and discloses a method for multi-dimensional information integration and intelligent querying of immigration records. The method includes: collecting and standardizing multi-source heterogeneous immigration record data to generate data to be processed; constructing an initial event knowledge graph containing element nodes, event nodes, and relationships based on the data to be processed; performing causal relationship discovery and correction on the initial event knowledge graph; mining, updating confidence levels, and completing logic for causal relationships between events by integrating data statistical analysis and business process paradigms learned from historical data; and generating a time-series causal graph; receiving users' natural language query requests and performing queries and inferences on the time-series causal graph. This invention can reveal the hidden business process logic behind the data, support intelligent question answering for complex questions such as event causes and impacts, and improve the intelligence level of queries and the interpretability of results.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, specifically to a method for integrating and intelligently querying multi-dimensional information from immigration records. Background Technology

[0002] Immigration record management involves a large amount of data from diverse sources and in various formats, such as structured records in business databases and unstructured texts like contracts and other documents. Current technology often manages this data at the level of independent storage and digitization, lacking effective methods for structured integration. This results in information from different sources describing the same business entity or event remaining isolated, making it difficult to form a unified and coherent view of the records.

[0003] Furthermore, current archival systems primarily organize information chronologically, recording the order in which events occur, but typically failing to reveal the underlying business logic and causal relationships between them. The system cannot automatically identify the steps involved in a standard business process, nor can it determine whether two consecutive events have a necessary causal connection. When data records are incomplete, this purely chronological organization method is even less able to reveal logical breakpoints within the business process.

[0004] The limitations of this data organization method directly restrict the level of querying and utilizing archival information. Users can usually only search through preset fields or keywords, and the query method is relatively fixed. For complex analytical needs such as exploring the causes behind events and predicting subsequent developments, which require reasoning based on business logic and causal relationships, existing technologies cannot provide direct support, and users need to spend a lot of effort on manual sorting and correlation analysis. Summary of the Invention

[0005] The technical problem addressed by this invention is that existing immigration record management systems typically face issues such as data isolation and lack of correlation due to diverse data sources and inconsistent formats. This scattered data lacks structured internal connections, particularly failing to effectively reveal the temporal and causal relationships between events. Therefore, traditional record management and query methods are ill-suited to support in-depth analysis and intelligent question answering of complex business processes, and cannot meet the advanced analytical needs of managers to explore the underlying causes of events and predict future developments.

[0006] To address the shortcomings of existing technologies, this invention provides a method for multi-dimensional information integration and intelligent querying of immigration records. This method solves the problems of single-dimensional data integration, difficulty in revealing logical relationships between events, relatively fixed query methods, and low level of intelligence by constructing a causal relationship network between events and incorporating business process logic for correction.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for integrating and intelligently querying multi-dimensional information from immigration records, including:

[0009] Collect raw data from diverse and heterogeneous immigration records, and perform normalization processing on the raw data to generate data to be processed;

[0010] Based on the data to be processed, an initial event knowledge graph is constructed, which includes element nodes, event nodes, participation relationship edges, and temporal relationship edges.

[0011] Perform causal relationship discovery and correction on the initial event knowledge graph to generate a time-series causal graph;

[0012] Receive the user's natural language query request, perform query and reasoning on the time-series causal graph, and generate and output the query results.

[0013] By adopting the above technical solution, its core innovation mechanism lies in the causal relationship discovery and correction process performed on the initial event knowledge graph. This process does not simply rely on the order in which events occur, but integrates data statistical patterns and domain business logic, as detailed below:

[0014] Potential causal relationship mining based on statistical analysis: The method first performs statistical analysis on event data in the initial event knowledge graph to quantify the strength of conditional probability association between any two event types. If the occurrence of one event increases the probability of the other event occurring within a subsequent specific time window, a potential causal relationship is preliminarily determined. This step constructs an initial set of causal relationship hypotheses based on statistical correlation, providing a data foundation for subsequent logical correction.

[0015] Business process paradigm learning based on sequence pattern mining: To ensure that the discovered causal relationships align with real business logic, the method learns and summarizes standard business process patterns from historical data. By converting the archival records of each core entity into a time-ordered sequence of event types and applying sequence pattern mining algorithms, recurring and common event combination patterns can be discovered. These discovered frequent subsequences constitute a business process paradigm library, which provides the business logic basis for subsequent causal relationship correction.

[0016] Integration and correction of statistical regularities and business logic: The method compares each potential causal relationship discovered in the first step with the business process paradigm learned in the second step.

[0017] For potential causal relationships that perfectly match the business process paradigm, the system will increase their confidence level and confirm them as highly reliable causal chains.

[0018] For potential causal relationships that contradict the business process paradigm, the system will weaken their confidence level, effectively filtering out pseudo-causal relationships that only have accidental statistical correlations but lack business logic support.

[0019] For causal relationships with paradigm jumps (i.e., data shows that event A directly caused event C, and the learned business paradigm is ABC), the system will determine that it is a logical break caused by missing data records. In this case, the system will insert a hidden event node representing the missing link (event B) between events A and C, thereby logically completing the causal chain.

[0020] Through the aforementioned statistical discovery, logical learning, fusion correction, and logical completion processes, this invention upgrades an initial graph containing only temporal relationships into a logically rigorous, reliable, and complete temporal causal graph. Therefore, this invention not only achieves deep integration of immigration record data, but more importantly, it reveals and completes the implicit business process logic within the data, enabling intelligent queries oriented towards causal relationships and achieving the technical effect of accurate and highly interpretable query results.

[0021] Preferably, in the normalization process, an entity alignment operation is performed on entity records lacking a unique identifier. This operation achieves accurate matching and identity unification of records from different sources by calculating a weighted similarity score based on multiple attributes.

[0022] By adopting the above technical solution, the reliability of all subsequent analyses has been improved.

[0023] Preferably, in the construction of the initial event knowledge graph, a joint information extraction model is adopted, and a named entity recognition task is executed simultaneously to extract element nodes, and an event extraction task is executed to extract event nodes and participation relationship edges.

[0024] By adopting the above technical solutions and utilizing the parameter sharing mechanism inherent in the joint model, the overall efficiency and consistency of information extraction are improved, and the structural quality of the constructed map is guaranteed.

[0025] Preferably, in the causal relationship discovery and correction, the statistical association between events is quantified by calculating the potential causal strength score; business process paradigms are discovered from the set of event type sequences by applying a sequence pattern mining algorithm; causal confidence is updated by comparing the matching status of potential causal relationships with business process paradigms; and logical completion is performed by inserting latent event nodes by identifying causal relationship edges in paradigm jump states.

[0026] By adopting the above technical solution, a specific and quantifiable implementation method is provided for the entire process, ensuring the reproducibility and robustness of the technical solution.

[0027] Preferably, in the query and reasoning process, the user's natural language query request is parsed and constructed into a structured query graph, then the structured query graph is converted into a formal query statement and subgraph matching is performed on the temporal causal graph. Finally, the returned original query result set is aggregated, sorted, and converted into natural language text output.

[0028] By adopting the above technical solutions, a complete interactive process from natural language to machine query and back to natural language was constructed, which reduced the threshold for using the system. At the same time, the availability and transparency of the query results were improved by using causal confidence to sort the results and annotate missing links.

[0029] This invention provides a method for integrating and intelligently querying multi-dimensional information from immigration records. It offers the following advantages:

[0030] 1. This invention establishes a structured association between archival information from different sources by collecting and standardizing multi-source heterogeneous data and constructing an initial event knowledge graph containing elements, events and relationships. This solves the problem of isolated archival information caused by data dispersion and inconsistent formats, and forms a unified, multi-dimensional view of immigration archival information.

[0031] 2. This invention performs causal relationship discovery and correction on event knowledge graphs, integrating data statistical analysis and business process paradigm learning. It can identify and confirm causal relationships between events that conform to business logic, and at the same time fill in logical gaps caused by missing data records. It reveals the business processes and causal chains hidden behind archival data, solves the problem that traditional methods are difficult to conduct in-depth logical analysis, and improves the interpretability of data.

[0032] 3. Based on the constructed temporal causal graph, this invention supports the parsing and reasoning of natural language query requests. Users are no longer limited to traditional keyword or fixed condition searches, and can directly raise complex causal inquiry questions such as the cause of an event or the impact of an event. This query method solves the problems of existing query methods being mechanical and lacking in intelligence, and realizes an improvement from data retrieval to intelligent question answering. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the system functional module architecture of a method for integrating and intelligently querying multi-dimensional information in immigration records according to an embodiment of the present invention.

[0034] Figure 2This is a flowchart illustrating a method for integrating and intelligently querying multi-dimensional information in immigration records according to an embodiment of the present invention.

[0035] Figure 3 This is a schematic diagram of the structure of an initial event knowledge graph according to an embodiment of the present invention;

[0036] Figure 4 This is a flowchart illustrating a causal relationship correction and completion algorithm according to an embodiment of the present invention.

[0037] The module includes: 100, Data Acquisition and Preprocessing Module; 200, Time-Series Event Graph Construction Module; 300, Time-Series Causal Relationship Discovery and Correction Module; and 400, Intelligent Query and Reasoning Module. Detailed Implementation

[0038] 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.

[0039] See attached document Figure 1 The method provided in this embodiment of the invention can be run on one or more computing devices. The computing device can be general-purpose or special-purpose, and its hardware configuration includes a processor, memory, and a system bus for connecting different system components. The memory stores computer program instructions, and the processor is configured to execute these program instructions to implement the method of the present invention.

[0040] In one embodiment of the present invention, a method for multi-dimensional information integration and intelligent query of immigration files may include the following system-level functional modules: a data acquisition and preprocessing module 100, a time-series event graph construction module 200, a time-series causal relationship discovery and correction module 300, and an intelligent query and reasoning module 400.

[0041] See attached document Figure 2 The present invention provides a method for multi-dimensional information integration and intelligent query of immigration records, comprising the following steps:

[0042] S201 involves collecting raw immigration records from multiple heterogeneous sources and performing normalization processing on the raw data to generate data to be processed. This normalization process includes cleaning, aligning, and formatting data from different sources and formats.

[0043] S202, Based on the data to be processed, construct an initial event knowledge graph. This construction process includes: extracting event nodes and element nodes from the data to be processed according to a preset event element model, and generating participation relationship edges and temporal relationship edges connecting the nodes.

[0044] S203, perform causal relationship discovery and correction on the initial event knowledge graph to generate a time-series causal graph. This step specifically includes: first, mining potential causal relationships between events through statistical analysis; then, using process paradigms learned from historical data, updating the confidence level and logically completing the potential causal relationships.

[0045] S204: Receive the user's natural language query request and perform query and inference on the temporal causal graph to generate and output query results. This step includes: semantically parsing the natural language query request to generate a structured query instruction, performing subgraph matching and inference on the temporal causal graph based on the instruction, and aggregating the matching results to generate the final query result.

[0046] The following will describe in more detail the various steps and functional modules involved in the embodiments of the present invention with reference to the accompanying drawings.

[0047] In step S201, this embodiment of the invention collects and standardizes the raw data of multi-source heterogeneous immigration records. This step is the foundation for all subsequent data analysis and map construction. A specific implementation can be further broken down into the following sub-steps:

[0048] S2011 retrieves raw immigration records from one or more external data sources through a pre-defined data access interface. This interface is configured to support multiple data source types. For structured data stored in relational databases (e.g., Oracle, MySQL, SQL Server), access and data extraction can be performed via standard interfaces such as Java Database Connectivity (JDBC) or Open Database Connectivity (ODBC). For unstructured or semi-structured data existing in file format, such as text documents (.doc, .docx, .pdf) and image files (.jpg, .png, .tiff) stored on a file server, reading can be done through the file system application programming interface (API).

[0049] S2012 performs alignment and standardization processing on the acquired structured data. Due to differences in structure and field naming across database tables from different business systems or historical periods, schema alignment is necessary. Specifically, the system maintains a configurable field mapping rule base, which defines the correspondence between fields in different source data tables and standard fields in the target unified data model. During processing, the system uses this rule base to unify field names from different sources to standard names.

[0050] This step further includes performing entity alignment to address the issue of records representing the same entity (e.g., the same immigrant household) in different data tables lacking unique identifiers. When a record contains an identifier that can be used as a unique criterion, it is directly used as the alignment basis. When such identifiers are missing or inconsistent, an attribute similarity-based method is used to match entity records. An entity similarity score calculation formula is used to compare two entity records. and The similarity is quantified, and the entity similarity score is calculated using the following formula:

[0051] ;

[0052] in:

[0053] Representative entity record and Overall similarity score between them;

[0054] This refers to the number of preset attributes used for comparison, such as selecting name, address, family member relationship, etc. One attribute;

[0055] Representing the The preset weights of each attribute, and satisfying The weights can be manually set according to the importance of different attributes. For example, the weight of the ID number can be set to 1, and the weight of the name can be set to 0.8.

[0056] and Representing records respectively and The Middle The attribute value of each attribute;

[0057] It is the first The similarity calculation function corresponding to each attribute.

[0058] The specific implementation of this function depends on the attribute. The function is determined by the data type. For example, for text attributes (such as name or address), the function can use string similarity measures such as edit distance (Levenshtein distance) or Jaccard similarity coefficient; for numeric or date attributes, the function can calculate based on the normalized difference between the two attribute values. When the value exceeds the preset similarity threshold, the system determines the record. and They point to the same entity and assign it a unified entity ID.

[0059] S2013 performs normalization processing on the acquired unstructured and semi-structured data. For image files, especially scanned copies of contracts, agreements, and handwritten forms, the system invokes the Optical Character Recognition (OCR) engine module. This OCR engine module performs layout analysis, text region detection and recognition on the input image, converting the text content in the image into a machine-readable string sequence. The specific implementation of the OCR engine is well-known in the field and will not be described in detail here.

[0060] For text documents and text content output by the OCR engine, the system performs natural language preprocessing. This process includes: first, cleaning the text by removing headers, footers, page numbers, and other formatting tags unrelated to the main text; then, using a sentence segmentation model to divide large blocks of text into lists of sentences, which serve as the basic processing units for subsequent information extraction. Through these sub-steps, step S201 ultimately outputs a unified, entity-aligned, and structured dataset to be processed, laying the foundation for subsequent graph construction.

[0061] See attached document Figure 3 In step S202, this embodiment of the invention constructs an initial event knowledge graph based on the data to be processed output in step S201. This graph uses events as its organizational core, structurally representing various elements in the archival information and their dynamic relationships. Specific implementation of this step may include:

[0062] First, the data structure of the initial event knowledge graph constructed in this embodiment of the invention is formally defined. This graph is a directed attribute graph, composed of a set of nodes and a set of edges. The node set includes two types:

[0063] Element nodes and event nodes.

[0064] The data structure of an element node may include a unique identifier, node type (e.g., person, geographic location, organization, document, asset), and specific attributes (e.g., name and ID number for a person node; title and document number for a document node).

[0065] The data structure of an event node may include a unique identifier, event type (e.g., signing a compensation agreement, holding a mobilization meeting), and event attributes (e.g., start time, end time, location, event description).

[0066] The edge set also includes two types: participation relationship edges and temporal relationship edges. Participation relationship edges are directed edges that connect feature nodes and event nodes, and are labeled with role information to indicate the role in which the feature participated in the event. Examples of role information include "contracting entity", "land parcel involved", "meeting location", "policy basis", etc. Temporal relationship edges are also directed edges that connect two event nodes and are used to indicate the chronological order of the two events. The types of temporal relationships can include "before", "begins with", "ends with", "contains", "overlaps", etc.

[0067] S2021 involves extracting feature nodes and event nodes from the text content of the data to be processed, and establishing participation relationship edges between them. This extraction process can be implemented using a joint information extraction model. A specific implementation method is:

[0068] A sequence labeling model based on a pre-trained language model is used to perform the Named Entity Recognition (NER) task to extract feature nodes. For example, a model combining a BERT model and a Conditional Random Field (CRF) layer takes a sequence of lexical terms in a sentence as input and outputs a labeled sequence to annotate the input sentence, identify fragments in the text that refer to predefined entity types such as person names, place names, and organization names, and instantiate these identified entities as feature nodes in the graph.

[0069] Based on the identified feature nodes, an event extraction (EE) task is performed to extract event nodes and participation relationships. This task further includes two sub-tasks: event type identification and event argument role extraction. Event type identification is used to determine whether a predefined event exists in the sentence and to identify its specific type. Event argument role extraction is used to find the feature entities participating in the event (i.e., the feature nodes extracted by the aforementioned NER task) from the sentence after identifying the event, and to assign a preset semantic role to each feature entity. For example, for the sentence "A certain immigrant household signed a compensation agreement in the office on a certain day of a certain month of a certain year", the event extraction task will identify the event of type "signing an agreement" and extract the arguments "a certain immigrant household" (role: signatory), "a certain day of a certain month of a certain year" (role: time), "office" (role: location), and "compensation agreement" (role: object). Based on this, the system generates an event node and establishes participation relationship edges between the event node and feature nodes such as "a certain immigrant household", "office", and "compensation agreement".

[0070] S2022, based on the generated set of event nodes, calculates and generates temporal relationship edges between event nodes. Each event node contains one or more time attributes, such as "occurrence time," "start time," and "end time." By considering any two event nodes... and timestamp attribute and By comparing them, the temporal relationship between them can be determined. For example, if The end time is earlier than The start time is at and Establish a temporal relationship edge between them, with the type being "prior" and the direction being "from". point to If the time intervals of two events overlap, overlapping temporal relationship edges can be established. After completing the above steps, an initial event knowledge graph is constructed. This graph integrates entity and event information from the archives and includes basic participation and temporal relationships, providing a structured data foundation for subsequent causal relationship discovery.

[0071] In step S203, this embodiment of the invention processes the initial event knowledge graph generated in step S202 to discover and correct causal relationships between events, ultimately generating a time-series causal graph containing high-quality temporal and causal relationships. This step is the core part of the method of this invention, and its specific implementation may include:

[0072] S2031 involves statistical analysis of all event nodes in the initial event knowledge graph to uncover potential causal relationships between different event types. The principle behind this process is that if the occurrence of one type of event increases the probability of another type of event occurring within a subsequent specific time window, then a potential causal relationship exists between the two types of events. A potential causal strength scoring formula is used to evaluate any two event types. and The strength of the causal relationship between them is quantified, and the formula for the potential causal strength score is as follows:

[0073] ;

[0074] in:

[0075] Represents the event type Pointer to event type Potential causal strength score;

[0076] The type of reference is An event instance, The type of reference is An event instance;

[0077] It is an event instance. The timestamp of the occurrence;

[0078] It is a predefined time window length, which is a configurable system parameter;

[0079] Represents conditional probability, indicating that in cases of type The event at time Under the premise that it occurs, in the subsequent The type observed within the time window is The probability of an event occurring. This probability value is calculated by analyzing all data in the dataset. The event instances of the type are iterated and counted to estimate the value;

[0080] Representative event type The baseline probability of occurrence. This probability value is calculated over the entire data time span for the event type. The average frequency of occurrence per unit time.

[0081] Calculation The value of . If the value is greater than 1, it indicates that the event . The occurrence of the event The occurrence of this phenomenon has a statistically facilitating effect. For all calculated scores exceeding the preset causality threshold... Event type The system will establish a "potential causal" relationship edge between corresponding event instances that satisfy the sequential relationship in the initial event knowledge graph, and assign a score to each edge. Used as the initial confidence level for this edge.

[0082] S2032, Learning Business Process Paradigms from Historical Archive Data. A business process paradigm refers to a standardized workflow consisting of a series of ordered events that is repeatedly executed in immigration record management. The specific implementation of this learning process is as follows:

[0083] First, the event data in the initial event knowledge graph is organized according to the timeline of preset core element nodes (e.g., each "immigrant household" element node or "project plot" element node). All events associated with the same core element node are sorted by their occurrence time, transforming them into a set of event type sequences. Each sequence represents a complete business case, for example... .

[0084] Then, sequence pattern mining algorithms are applied to the set of event type sequences to discover frequently occurring subsequences. For example, the PrefixSpan algorithm is used to process the sequence set, which can effectively discover frequent sequence patterns that meet a preset minimum support threshold. The minimum support threshold is used to ensure that the mined patterns are sufficiently prevalent in the dataset. The specific implementation of sequence pattern mining algorithms (such as PrefixSpan) is well-known in the art and will not be described in detail here.

[0085] Finally, all the frequently discovered sequence patterns are used as business process paradigms and stored in a process paradigm library. This process paradigm library provides domain knowledge and business logic basis for subsequent causal relationship correction steps.

[0086] In S2033, the business process paradigm learned in S2032 is compared with the potential causal relationships discovered in S2031 to perform confidence updates and pruning on the potential causal relationships. This step uses highly deterministic business process knowledge to examine and correct statistically uncertain associations.

[0087] See attached document Figure 4 The overall process of this algorithm can be described as follows: The system traverses each potential causal relationship edge established in S2031 in the initial event knowledge graph. For any connection event instance... (type is) ) and event instances (type is) The system searches for potential causal edges in the process paradigm library generated by S2032.

[0088] The purpose of this search is to determine the sequence of event types. The matching relationship with learned business process paradigms. The search results are categorized into three states:

[0089] 1. The first state is normalization compliance: There exists at least one business process normal form in the process normal form library, where the event type... and The occurrence of directly adjacent event pairs indicates the existence of subsequences within the paradigm. .

[0090] 2. The second state is a paradigm jump: not found in the process paradigm library. The direct adjacent patterns exist, but there is at least one paradigm, where and They appear in non-directly adjacent locations, that is, they exist in the paradigm. Such a subsequence.

[0091] 3. The third state is a normal form mismatch: no such mismatch was found in any of the normal forms in the process normal form library. and They appear in the same logical chain (neither directly adjacent nor indirectly adjacent).

[0092] After determining the paradigm matching state of each potential causal edge, the system updates the confidence level of that edge based on that state. This confidence update process aims to strengthen causal relationships that conform to business logic and weaken causal relationships that do not. The initial confidence level of potential causal edge is adjusted using a causal confidence update formula, which is:

[0093] ;

[0094] in:

[0095] The representative, after correction, is based on the event type. point to The updated causal confidence;

[0096] It is the initial potential causality strength score calculated in S2031;

[0097] It is a paradigm adjustment factor, the value of which is uniquely determined by the aforementioned paradigm matching state.

[0098] Paradigm Adjustment Factor The rules for determining the value are as follows:

[0099] If potential causality If the matching status is normal form conformance, then It is assigned a preset value greater than 1. (For example, (to strengthen the causal relationship).

[0100] If potential causality If the matching state is a paradigm jump, then It is assigned a default value equal to 1. (Right now This measure aims to maintain the original statistical confidence of the relationship without imposing penalties, allowing it to be properly processed by subsequent steps in S2034.

[0101] If potential causality If the matching status is non-normal, then It is assigned a preset value between 0 and 1 (For example, This is to weaken the statistical correlation that lacks business logic support.

[0102] Through the above update mechanism, the confidence level of all causal relationships conforming to the business process paradigm is increased, while the confidence level of causal relationships that only have statistical correlation but lack business process support is decreased. After completing the confidence level update for all potential causal edges, the system updates all the confidence levels. Pruning below the preset confidence threshold The relation edges are removed from the graph.

[0103] S2034, after completing the confidence update and pruning of the causal relationship edges, this embodiment of the invention further identifies and completes the logical jumps existing in the graph.

[0104] The specific implementation of this completion mechanism is as follows: the system traverses all states in the graph that are determined to be paradigm jump states in S2033, and their updated confidence scores are as follows: Above the pruning threshold The causal relationship.

[0105] Since these edges have already been marked with S2033, this step does not require performing a paradigm library search again. For any instance of such a connection event... (type is) ) and event instances (type is) The system confirms that the paradigm jump edge represents a paradigm jump edge. The non-direct adjacency relationship.

[0106] When the system detects such non-directly adjacent normal form matches, it determines the event instance. and There is a logical jump in the causal chain between them. At this point, the system will... and Insert one or more implicit event nodes between them. For example, for In this paradigm, the system will generate a new implicit event node. This hidden event node It is assigned a special type, such as process missing, and may contain an attribute to record the type of event it represents that is missing in the data (e.g., record). .

[0107] Insert the hidden event node At the same time, the system will remove the original [item / entity]. point to Establish the causal relationship edge, and create two new directed relation edges: one from... point to Another one from point to To ensure that the confidence information of the causal chain is accurately transmitted, the system updates the causal confidence of the original edges. Assigned to the first new edge (i.e., from) point to The first edge represents the starting point and core confidence level of the entire causal inference. The second new edge (i.e., from...) point to The edge (e.g., the starting edge) is assigned a specific confidence level (e.g., 1.0) because it represents a deterministic logical step within a known business paradigm. Thus, the overall confidence level of the completed causal path is determined by the confidence level of its starting edge. A single representative thus resolves the ambiguity in information transmission.

[0108] Through this completion mechanism, the method of this invention ensures that the final generated temporal causal graph not only reflects statistical causality at a macroscopic level, but also completes the missing business process links at a microscopic level, making its logical chain more complete. After completing all steps S2031 to S2034, the system outputs the final temporal causal graph for use in the subsequent query step S204.

[0109] In step S204, this embodiment of the invention provides intelligent query and inference services to users based on the time-series causal graph output in S203, which already contains high-quality time series and causal relationships. Specific implementation of this step may include:

[0110] S2041, the system performs intent decomposition on the user-input natural language query, converting it into a structured query graph. This step transforms the natural language query into a graph-understandable structured instruction. A specific implementation involves the system receiving the user's text query (e.g., "Query the reason why the user signed the agreement" or "Query what happened after the mobilization meeting").

[0111] The text query is input into a natural language understanding module. This module can employ a sequence labeling model fine-tuned based on a pre-trained language model (such as BERT) to perform named entity recognition and intent classification tasks. Named entity recognition is used to extract key element node information (such as "user") and event type information (such as "signing an agreement" or "mobilization meeting") from the query. Intent classification is used to identify the user's query purpose, for example, whether it is a causal query based on keywords such as "reason" or "why," or a temporal query based on keywords such as "before" or "after."

[0112] After identifying the key information mentioned above, the system performs an entity linking operation, matching the entities mentioned in the text (such as "user") with the corresponding unique element nodes in the graph.

[0113] Finally, based on the identified entity nodes, event types, and query intent (sequential or causal), the system constructs a structured query graph in memory. This query graph is a subgraph pattern containing variables; for example, for querying the reason why a user signed an agreement, its structured query graph can be defined as: finding an event node. and an event node , making and There exists a causal edge between them, and at the same time The type is "signing an agreement", and There is a participation relationship with the node "user".

[0114] S2042, perform graph inference based on subgraph matching to execute the structured query graph generated in S2041. Specifically, this step involves the system converting the structured query graph into a formal query statement that can be executed in a graph database (e.g., Neo4j or JanusGraph).

[0115] One specific implementation involves converting the query graph into a Cypher query language or SPARQL query statement. The graph database's query engine executes this query statement on the time-series causal graph output by S203. This execution process is a subgraph matching process, where the query engine searches the entire graph for all subgraph instances that satisfy the constraints of the structured query graph.

[0116] In this process, the query engine utilizes the causal and temporal relationship edges constructed by this invention for traversal. For example, if the query intent is "...the reason for...", the query engine will focus on traversing those that were retained or enhanced in S203 and have a confidence level higher than the threshold. The causal relationship edges, specifically, match incoming causal relationship edges pointing to the target event node. If the query intent is "the impact of...", the query engine matches outgoing causal relationship edges originating from the target event node. If the query intent is "before...", the query engine matches temporal relationship edges such as "prior to..." pointing to the target event node. During traversal and matching, if the query path contains an implicit event node of type "process missing" generated in S2034, the system can explicitly mark the missing business process in the subsequent result presentation stage. The query engine returns all matched subgraph instances (i.e., paths and nodes) as the original query result set to subsequent processing steps.

[0117] S2043 aggregates and presents the raw query result set returned in S2042. Since subgraph matching returns a large number of raw, unorganized path instances, this step aims to transform these instances into a user-friendly and easy-to-understand final answer.

[0118] One specific implementation of this step is as follows: the system aggregates and sorts the original query result set. For causal queries (e.g., "the reason for..."), the system will utilize the updated causal confidence calculated in S2033. All matched causal paths are sorted in descending order. The system will prioritize displaying those with higher... The system identifies the event node pointed to by the causal path of the value. Simultaneously, the system will group the results, merging multiple matching paths pointing to the same answer node (e.g., the same causal event) into a single result and displaying its highest confidence level as the final confidence level of the merged result.

[0119] After aggregation sorting, the system converts the structured results (i.e., subgraph instances) into natural language text. For example, a subgraph... point to An instance of a causal subgraph can be converted into a form such as: "event[ The cause of the occurrence of [the description] is: [ [Description], the credibility of this causal relationship is [ The text is titled "[value]". During this conversion process, the system will specifically handle latent event nodes identified in S2042 that are of type "process missing". If a line contains , and The query path contains this type of node (e.g. Represents missing event types If the event is specified, the generated text can be adjusted to: "Event [ The description led to [ [Description], but the standard business process between the two ([ The description of [ ] is missing in the original data. The confidence level of this causal inference is . "

[0120] As another presentation method, the system can also visualize the matched and aggregated subgraph instances on the user interface, such as displaying the start and end points of the query and the causal or temporal paths between them in the form of a graph, and highlighting key nodes or confidence information.

[0121] Finally, the system will present the generated natural language text or visual graph as the final query result to the user.

Claims

1. A method for multi-dimensional information integration and intelligent query of immigration archives, characterized in that, include: S1. Collect raw data from multi-source heterogeneous immigration files, and perform normalization processing on the raw data to generate data to be processed; S2. Based on the data to be processed, construct an initial event knowledge graph, which includes element nodes, event nodes, and participation relationship edges and temporal relationship edges connecting the element nodes and the event nodes. S3. Perform causal relationship discovery and correction on the initial event knowledge graph to generate a time-series causal graph; The causal relationship correction in step S3 includes: All events associated with the same core element node in the initial event knowledge graph are sorted by occurrence time and converted into a set of event type sequences; A sequence pattern mining algorithm is applied to the set of event type sequences to discover frequently occurring subsequences, and these frequently occurring subsequences are stored as business process paradigms in the process paradigm library. For each potential causal relationship edge, retrieve the matching relationship between the type sequence of the two events connected by the potential causal relationship edge and the business process paradigm in the process paradigm library, and classify the matching relationship into three paradigm matching states: paradigm conformity, paradigm jump, or paradigm mismatch. The causal confidence update formula is used to adjust the initial confidence of the potential causal relationship edge according to the paradigm adjustment factor corresponding to the paradigm matching state, thereby generating the updated causal confidence. Identify all causal relationship edges in the graph that are classified as paradigm jump states and whose updated causal confidence is higher than a preset confidence pruning threshold; For the identified causal relationship edge, a latent event node is inserted between the two event instances connected by the causal relationship edge, the identified causal relationship edge is removed, and a new relationship edge is established from the starting event instance to the latent event node and from the latent event node to the ending event instance. S4. Receive the user's natural language query request, perform query and reasoning on the time-series causal graph, and generate and output the query results.

2. The method for multi-dimensional information integration and intelligent query of emigrant archives according to claim 1, characterized in that, In step S1, the normalization process includes: For entity records lacking a unique identifier, an entity alignment operation is performed. The entity alignment operation uses an entity similarity score calculation formula to quantify the similarity between two entity records. When the calculated similarity exceeds a preset similarity threshold, it is determined that the two entity records point to the same entity.

3. The method of claim 1, wherein, In step S2, constructing the initial event knowledge graph includes: A joint information extraction model is used to perform named entity recognition on the text content of the data to be processed, and element nodes are extracted. Based on the aforementioned element nodes, the event extraction task is performed using the joint information extraction model to extract event nodes and participation relationship edges.

4. The method for multi-dimensional information integration and intelligent query of emigrant archives according to claim 1, characterized in that, In step S3, the causal relationship discovery includes: The potential causal strength score is generated by quantifying the causal association strength between any two event types in the initial event knowledge graph using the potential causal strength score formula. For event type pairs whose potential causal strength scores exceed a preset causal threshold, a relationship edge of type potential causality is established between the corresponding event instances that satisfy the sequential relationship in the initial event knowledge graph.

5. The method for multi-dimensional information integration and intelligent query of emigrant archives according to claim 1, characterized in that, In step S4, the execution of query and reasoning includes: Semantic parsing is performed on the natural language query request to identify the element nodes, event types, and query intent in the query; A structured query graph is constructed based on the element nodes, the event types, and the query intent.

6. The method for multi-dimensional information integration and intelligent query of emigrant archives according to claim 5, characterized in that, In step S4, the execution of query and reasoning further includes: The structured query graph is converted into a formal query statement that can be executed in a graph database; The formal query statement is executed on the temporal causal graph to perform subgraph matching, and the causal relationship edges or temporal relationship edges are selectively traversed according to the query intent to return the original query result set.

7. The method according to claim 6, wherein, In step S4, generating and outputting the query results includes: The original query result set is aggregated and sorted to generate a sorted structured result; The sorted structured results are converted into natural language text for output. If the query path contains hidden event nodes, the missing business processes are marked in the output text.

Images