A label system driven report generation system and method

Abstract

The application relates to the technical field of data processing and intelligent text generation, and discloses a label system driven report generation system and method, which comprises the following steps: receiving a report generation request, generating a key statistical index set and a bottom layer data entity label set, judging a report generation task execution mode to obtain an execution mode judgment result; when the judgment result is a cause-effect reasoning and judgment mode, a structured cause-effect reasoning framework with entity-level traceability anchor points is generated, self-recurrence text filling is performed to generate candidate natural language conclusion vocabularies; the candidate natural language conclusion vocabularies are reversely mapped into a graph node set in a knowledge graph, and a connected path between the graph node set and a bottom layer original work order entity node corresponding to the entity-level traceability anchor points is searched; the application constructs a constraint mechanism between semantic vocabularies and bottom layer source data by performing conditional branch processing according to a connected path search result, and realizes bidirectional traceability of natural language conclusions and bottom layer original work orders.

Description

Technical Field

[0001] This invention relates to the field of data processing and intelligent text generation technology, and more specifically, to a report generation system and method driven by a tag system. Background Technology

[0002] With the development of artificial intelligence and data governance technologies, automatically generating semi-structured or structured analysis reports using large language models has become a common method for improving data application efficiency in the government sector. Patent application CN121936562A discloses an intelligent report generation system and method based on knowledge graph enhancement. It uses a first-level model to convert natural language into structured query statements, retrieves results from the knowledge graph, and then uses a second-level model to generate the report. By leveraging structured knowledge and efficient querying, it improves the accuracy and efficiency of report conclusions and enables traceability of the reasoning process. Patent application CN122155247A discloses an intelligent report generation and in-depth analysis method for government administrators. It inputs unstructured data into a large language model to form a knowledge graph and then generates language paragraphs to fill templates. This solves the problem of low efficiency in analysis and judgment work caused by the difficulty of processing unstructured data for government staff.

[0003] However, existing automated report generation solutions based on large models and knowledge graphs often suffer from the risk of losing topological relationships between generated text and discrete source data when processing multi-source heterogeneous data and generating analysis reports. Specifically, when business personnel review business analysis reports automatically generated by the system, they usually need to perform drill-down verification on specific conclusions in the report that contain causal logic or situational assessments. For example, when the report text outputs the assessment that a certain type of business governance problem has been alleviated, business personnel attempt to view the specific original work orders and processing records that support this assessment through interactive operations. The front-end interface often experiences addressing redirection failures or incorrect matching to macro-level policy document nodes. These failures in tracing the source and mismatching of evidence prevent business personnel from verifying the underlying data support for the assessment conclusions, reducing the business reference value of intelligent reports and making it difficult for automatically generated reports to meet the requirements of data rigor in actual business governance. Summary of the Invention

[0004] Regarding the issues encountered by business personnel when reviewing intelligent analysis reports, such as drill-down failures and failures in tracing back to the source, the technical cause is as follows: Although the front-end data governance module has structured and stored the original business work orders and processing records and labeled them with underlying data entity tags, in the final report text generation stage, the large language model uses an autoregressive approach to predict continuously distributed tokens one by one to construct the natural language report text. During this autoregressive generation process, the topological pointer coordinates corresponding to the original underlying discrete data tags supporting the analysis conclusions are discarded and overwritten by the newly generated continuous semantic features. At the internal generation mechanism level, there is a topological break between the continuous semantic feature expression in the text generation space and the discrete entity relationship graph in the data tracing space. Conventional automatic report generation methods are essentially a one-way output mapping path from source data to text vocabulary. Due to the lack of a structured constraint mechanism that strongly binds dynamically generated semantic units with underlying source data entity identifiers and performs reverse mapping in the text autoregressive generation process, the output natural language vocabulary loses its bidirectional association with the underlying original business entities, causing the rendering interface to be unable to obtain accurate address coordinates, ultimately leading to the interruption of the penetrating entity-level tracing link.

[0005] To overcome the shortcomings of existing technologies where irreversible topological breaks between generative text and discrete source data lead to a loss of source tracing confidence, this invention provides a tag-driven report generation system and method. By pre-constructing a structured causal reasoning framework with entity-level source tracing anchors under a causal reasoning judgment model, and performing reverse mapping and connectivity path search on candidate natural language conclusion words to a knowledge graph during the autoregressive text filling stage of the government hotline large model, conditional branching processing is performed on the text generation stream based on objective connectivity path search results. This establishes a structured bidirectional mapping path between continuous semantic words and the underlying original work orders, avoiding the overwriting of underlying data entity tag pointer coordinates by unidirectional decoding output, and ensuring the data verification connectivity and source tracing accuracy of the intelligent analysis report text.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A report generation method driven by a tag system includes: The system receives and parses a report generation request to obtain a set of key statistical indicators and a set of underlying data entity labels. It then determines the report generation task execution mode and generates a report generation task execution mode judgment result, which is either a descriptive summary mode or a causal reasoning judgment mode. When the report generation task execution mode judgment result is a causal reasoning judgment mode, a structured causal reasoning skeleton with entity-level source tracing anchors is generated. A structured causal reasoning skeleton with entity-level traceability anchors is input into the government hotline big data model for autoregressive text filling to generate candidate natural language conclusion terms. The candidate natural language conclusion terms are then mapped inversely to a set of graph nodes in a knowledge graph. The connection path between the set of graph nodes and the underlying original work order entity nodes corresponding to the entity-level traceability anchors is searched. Based on the search results of the connection path, conditional branching processing is performed on the candidate natural language conclusion terms to achieve bidirectional traceability between the candidate natural language conclusion terms and the underlying original work orders.

[0007] A tag-based report generation system is provided for implementing the aforementioned tag-based report generation method. The system includes: The skeleton generation module is used to receive report generation requests and parse them to obtain a set of key statistical indicators and a set of underlying data entity labels. It determines the report generation task execution mode and generates a report generation task execution mode judgment result. The report generation task execution mode judgment result is either a descriptive summary mode or a causal reasoning judgment mode. When the report generation task execution mode judgment result is a causal reasoning judgment mode, it generates a structured causal reasoning skeleton with entity-level source tracing anchors. Text Filling Module: Input the structured causal reasoning skeleton with entity-level source anchors into the government hotline big model for autoregressive text filling to generate candidate natural language conclusion words; and back-map the candidate natural language conclusion words to a set of graph nodes in the knowledge graph. Closed-loop verification module: used to search for the connection path between the graph node set and the underlying original work order entity node corresponding to the entity-level traceability anchor point. Based on the connection path search results, conditional branch processing is performed on the candidate natural language conclusion words to realize bidirectional traceability between the candidate natural language conclusion words and the underlying original work order.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the technical bottleneck of irreversible topological disconnection between the continuous semantic features of generative text and the discrete entity relationship graph. By pre-generating a structured causal reasoning skeleton with entity-level source anchors during the execution of the causal reasoning judgment mode, it ensures that the topological pointer coordinates of the underlying data labels are preserved before autoregressive text filling. In the text generation stage of the government hotline large-scale model, the system reverse-maps the generated candidate natural language conclusion words to a set of graph nodes within the knowledge graph, and performs a targeted search on the connection path between the graph node set and the underlying original work order entity nodes pointed to by the entity-level source anchors. The objective search results of the connection path are used to perform conditional branching processing on the candidate natural language conclusion words, constructing a structured binding and mapping verification closed loop between the text generation space and the data source space. This processing mechanism effectively prevents the model decoding from overwriting the coordinates of the underlying discrete data labels during the one-way output mapping process, establishes a bidirectional traceable mapping path between the natural language conclusion words and the underlying original supporting work orders, and ensures the data verification connectivity and source traceability confidence of the intelligent analysis report text. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of a report generation method driven by a tag system according to the present invention; Figure 2 This is a flowchart illustrating the task breakdown process for report generation in this invention. Figure 3 This is a schematic diagram illustrating the extraction of key indicators and entity labels in this invention; Figure 4 This is a schematic diagram of the structured causal reasoning framework in this invention; Figure 5 This is a schematic diagram of a bidirectional traceability scenario in this invention; Figure 6 This is a functional block diagram of a report generation system driven by a tag system in this invention. Detailed Implementation

[0011] The technical solutions of 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.

[0012] Example 1: Please see Figure 1 As shown, this embodiment provides a report generation method driven by a tag system, including: Step S10: Receive the report generation request and parse the target analysis dimension in the request. Extract the key statistical indicator set and underlying data entity label set under the target analysis dimension of the government hotline domain. Determine the report generation task execution mode and generate a report generation task execution mode judgment result. The report generation task execution mode judgment result is either a descriptive summary mode or a causal reasoning judgment mode. When the report generation task execution mode judgment result is a causal reasoning judgment mode, generate a structured causal reasoning skeleton with entity-level source tracing anchors. Further, see Figure 2 Step S10 includes: Step S11: Receive the report generation request and parse the target analysis dimension in the report generation request. Extract the key statistical indicators under the target analysis dimension of the government hotline from the government hotline knowledge data management module to obtain the key statistical indicator set. Extract the underlying data entity tags associated with the key statistical indicator set to obtain the underlying data entity tag set. Before the large-scale government hotline model initiates autoregressive text filling, a structured skeleton capable of carrying underlying data pointers is pre-established. This prevents the large-scale government hotline model from discarding the data pointers corresponding to the underlying data entity labels when generating natural language report text in an autoregressive manner. Specifically, the triggering events for initiating the intelligent report generation process include two types: periodic automatic triggering and manual request triggering. Periodic automatic triggering is initiated by the scheduling system automatically launching a report generation request based on a pre-configured report generation cycle in the government hotline domain. Manual request triggering is initiated by users with report subscription permissions submitting a report generation request on the front-end interface. See also Figure 3 This is a schematic diagram of key indicators and entity label extraction provided in the embodiments of this application. Figure 3schematically illustrates the process where a user submits a report generation request carrying target analysis dimensions at a terminal, and inputs the request to the government hotline knowledge data management module. The report generation request carries target analysis dimensions, which are used to define the statistical scope and slicing perspective of the report content, and the target analysis dimensions include one or a combination of an administrative division dimension, a problem category dimension, and a time window dimension. The multi-source heterogeneous data governance module is a data governance component supporting the government hotline knowledge data management module, such as Figure 3 shown in the figure, there is two-way data interaction between the multi-source heterogeneous data governance module and the government hotline knowledge data management module. The multi-source heterogeneous data governance module is responsible for normalizing fields from four types of heterogeneous data sources, namely hotline work orders, handling records, return visit evaluations, and policy documents, into the database and uniformly labeling the fields, and providing a field query interface to the outside world.

[0013] Based on the pre-configured target analysis dimensions in the government hotline field, the field query interface provided by the multi-source heterogeneous data governance module is invoked to extract key statistical indicators for describing the overall situation under the target analysis dimensions. The extracted key statistical indicators are classified according to the administrative division dimension, the problem category dimension, and the time window dimension to form a key statistical indicator set, and each key statistical indicator in the key statistical indicator set carries two attribute fields: a statistical caliber and a statistical period. Based on the key statistical indicator set, the field query interface is invoked again to extract underlying data entity tags associated with the key statistical indicator set, in combination with Figure 3 the extraction scenario shown in the figure, the government hotline knowledge data management module finally uniformly parses and outputs the key statistical indicator set and the underlying data entity tag set, and the data basis of both sets is directly derived from the four types of original data sources: hotline work orders, handling records, return visit evaluations, and policy documents. The underlying data entity tag set includes four types of tags: work order classification tags, handling record tags, return visit evaluation tags, and policy document tags. The four types of tags are respectively labeled by the multi-source heterogeneous data governance module in the links of work order warehousing, handling process node writing, return visit work order recovery, and policy document archiving according to the problem category vocabulary, handling action vocabulary, evaluation dimension vocabulary, and policy theme vocabulary; wherein, a one-to-one mapping relationship is established between the work order classification tags and the underlying original work orders, a one-to-one mapping relationship is established between the handling record tags and the original handling records, a one-to-one mapping relationship is established between the return visit evaluation tags and the return visit evaluation records, and a one-to-one mapping relationship is established between the policy document tags and the policy documents; the handling record tags and the return visit evaluation tags establish an association relationship with the corresponding underlying original work orders through work order identifiers, and the policy document tags are associated with the underlying original work orders of corresponding categories through the theme mapping relationship between the policy theme vocabulary and the problem category vocabulary.

[0014] For example, when the target analysis dimension is composed of the administrative division dimension of a district, the problem category dimension of parking difficulty, and the time window dimension of this month, the key statistical indicator set includes three key statistical indicators: the total number of parking difficulty work orders in a district this month, the month-on-month change rate of parking difficulty work orders in a district this month, and the average completion time of parking difficulty work orders in a district this month. The underlying data entity label set includes multiple underlying original work orders marked by the parking difficulty work order classification label, multiple original processing records marked by the parking difficulty processing record label, multiple follow-up evaluation records marked by the parking difficulty follow-up evaluation label, and multiple policy documents marked by the parking governance policy document label.

[0015] The reason for adopting the approach of extracting the key statistical indicator set and the underlying data entity label set all at once in the initial stage of the intelligent report generation process is that the data scale maintained by the government hotline knowledge data management module continues to expand with business accumulation. Centralizing the extraction action at the initial stage of the process avoids repeated retrieval of the government hotline knowledge data management module in the subsequent skeleton pre-embedding and text filling stages, thereby avoiding interruption of the decoding rhythm of the government hotline's large model due to data retrieval delays during the autoregressive text filling process. The beneficial effect of the one-time extraction approach is that subsequent stages can complete anchor point pre-embedding and connectivity path verification without triggering underlying data queries, reducing the access pressure on the government hotline knowledge data management module in the report generation process. If step S11 is missing, a combination obstacle will arise between the real-time requirements of autoregressive text filling and the latency of underlying data retrieval, and the entity-level traceability anchor points required in step S14 will not be generated.

[0016] Step S12: Determine the report generation task execution mode based on the report generation request corresponding to the set of key statistical indicators, and obtain the report generation task execution mode judgment result; the report generation task execution mode judgment result is a descriptive summary mode or a causal reasoning judgment mode. Specifically, based on the report generation request corresponding to the set of key statistical indicators, task pattern recognition rules are used to determine the execution mode of the report generation task, resulting in a judgment result. Task pattern recognition rules refer to rules that branch and judge the execution mode of the report generation task based on the meta-information of the report generation request. The judgment criteria for task pattern recognition rules include three items: first, whether the report generation request contains causal reasoning appeal keywords; second, the target reader level marked in the report generation request; and third, whether the report generation request requires drill-down queries. Among these, the first and third items are the main determination items for the report generation task execution mode, and the second is an auxiliary verification item. The target reader level does not individually change the judgment result of the report generation task execution mode determined by the main determination items. Causal reasoning appeal keywords refer to a set of words that can imply that the report content needs to explain the causes of problems or demonstrate the governance effects. Causal reasoning appeal keywords include five categories of words: cause analysis, cause assessment, governance effectiveness, mitigation reasons, and reasons for month-on-month decline. These words appear in the report generation request as preset entries in a manually annotated thesaurus, rather than in the report text output by the government hotline big data model during the autoregressive filling stage. The target reader level refers to the hierarchical position of the report recipient within the government organizational structure, as indicated in the report generation request. The target reader level can be categorized into three levels: decision-making, execution, and public. The decision-making level refers to management entities with authority to make overall decisions, allocate resources, or conduct performance evaluations regarding matters related to the government hotline. The execution level refers to business entities responsible for handling government hotline work orders, resolving issues, or implementing rectification measures. The public level refers to recipients who publicly release the report content to the general public. Drill-down query refers to the ability of report recipients to directly access the underlying original work orders and processing records by clicking on concluding statements within the text while reading the report.

[0017] Based on the report generation requests corresponding to the key statistical indicator set, the three judgment criteria of the task pattern recognition rule are checked one by one, and the report generation task execution mode judgment result is output according to the correspondence between the rule antecedent and the rule consequent. When the report generation request requires drill-down query, or when the report generation request contains causal reasoning keywords, the report generation task execution mode judgment result is recorded as causal reasoning judgment mode; when the report generation request does not require drill-down query and does not contain causal reasoning keywords, the report generation task execution mode judgment result is recorded as descriptive summary mode. The target reader level serves as an auxiliary verification item to check whether the report generation task execution mode judgment result matches the report recipient level: when the target reader level is the decision-making level and the report generation task execution mode judgment result is the descriptive summary mode, a mode verification prompt message is generated; when the target reader level is the execution level or the public level and the report generation task execution mode judgment result is the causal reasoning analysis mode, a mode verification prompt message is generated; the mode verification prompt message is used to prompt the user to confirm whether the report generation request needs to be supplemented with causal reasoning request keywords or drill-down query requirements, and the mode verification prompt message does not change the report generation task execution mode judgment result already determined based on the main judgment item.

[0018] For example, when the report generation request is to generate a monthly statistical summary of citywide complaint hotspots for public release, and the request does not require drill-down queries or contain keywords related to causal reasoning, the report generation task execution mode judgment result is recorded as descriptive summary mode. When the report generation request is to submit a statistical summary of citywide complaint hotspots to the decision-making level, but the request does not require drill-down queries or contain keywords related to causal reasoning, the report generation task execution mode judgment result is still recorded as descriptive summary mode, and a mode verification prompt message is generated. When the report generation request is to conduct a causal analysis of the alleviation of parking difficulties in a certain district this month, because the request contains keywords related to causal reasoning, the report generation task execution mode judgment result is recorded as causal reasoning analysis mode. When the report generation request is to provide drill-down query capabilities for parking difficulties in a certain district this month, even if the request does not contain keywords related to causal reasoning, the report generation task execution mode judgment result is still recorded as causal reasoning analysis mode. When the report generation request is for a public analysis of the reasons for the effectiveness of addressing the parking difficulties in a certain district this month, the report generation request contains keywords related to causal reasoning. Therefore, the report generation task execution mode judgment result is still recorded as the causal reasoning analysis mode, and a mode verification prompt is generated to remind the user to confirm whether the report content to be publicly released needs to retain causal reasoning analysis content or requires drill-down queries. This mode verification prompt does not change the judgment result of the causal reasoning analysis mode.

[0019] The method for setting meta-information items in the task pattern recognition rules is as follows: Manual annotation is performed on historical report generation request samples in the government hotline field. These samples consist of randomly selected historical report generation requests from the actual report generation logs of the government hotline field in recent years, with a minimum of 500 samples. Annotators perform binary annotation on each historical report generation request based on whether it requires analytical functions or entity-level traceability. When the historical report text corresponding to a historical report generation request contains causal explanatory sentences, evidentiary supporting sentences, or requires a jump from the report text to the underlying original work order and original processing record, the historical report generation request is classified as a path requiring heavy load. When the historical report text corresponding to a historical report generation request only contains statistical values ​​and connective natural language sentences, and does not require a jump from the report text to the underlying original work order and original processing record, the historical report generation request is classified as a path not requiring heavy load. The statistical analysis identifies common metadata features among historical report generation requests requiring heavy-load paths. These common metadata features are metadata fields whose frequency exceeds a frequency threshold in the sample of historical report generation requests requiring heavy-load paths. The frequency threshold ranges from greater than or equal to 0.7 to less than or equal to 0.9. When statistical results indicate that causal reasoning keywords and drill-down queries directly trigger heavy-load paths, and the target reader level provides auxiliary hints for heavy-load paths, causal reasoning keywords and drill-down queries are identified as primary decision items, and the target reader level is identified as an auxiliary verification item. If the number of primary decision items is too small, some report generation requests requiring entity-level traceability but not containing causal reasoning keywords will be incorrectly classified into descriptive summary mode, losing the opportunity to embed entity-level traceability anchors. If the number of primary decision items is too large, some report generation requests with only statistical publishing needs will be incorrectly classified into causal reasoning analysis mode, causing unnecessary storage and decoding overhead. This embodiment uses causal reasoning keywords and drill-down queries as the main criteria, and target reader level as an auxiliary criterion, so that pattern splitting can simultaneously take into account the needs of causal analysis, entity-level tracing, and report generation throughput.

[0020] For example, when the common meta-information features obtained from manually labeled samples are causal reasoning appeal keywords, target reader level, and drill-down query, the judgment criteria are set to three items. The reason for distinguishing between descriptive summary mode and causal reasoning analysis mode in the report generation task execution mode is that the report output objects in the government hotline field have differentiated traceability granularity requirements: the report generation request corresponding to the descriptive summary mode does not contain causal reasoning appeal keywords and does not require drill-down query, mainly used to output macro-level trend values; the report generation request corresponding to the causal reasoning analysis mode contains causal reasoning appeal keywords or requires drill-down query, requiring specific underlying original work orders and original processing records to support causal reasoning conclusions or drill-down query capabilities. If no mode distinction is made and entity-level traceability anchors are uniformly pre-embedded for all report generation requests, the report corresponding to the descriptive summary mode will carry a large number of entity-level traceability anchors that will not actually be clicked through, resulting in wasted storage and rendering overhead. Furthermore, uniformly enabling the closed-loop verification mechanism of step S20 for all reports will slow down the decoding speed and affect the report generation throughput. The beneficial effect of the pattern differentiation approach is that it follows a lightweight path in descriptive summarization mode and a heavy-load path in causal reasoning mode, with each path matching its corresponding tracing granularity. Without step S12, the execution of steps S13 and S14 would lack branching criteria, forcing all reports to follow the heavy-load path, thus increasing storage and decoding overhead. The placeholder location identification strategy in steps S12 and S21 forms a two-layer throttling effect at both the pattern and location levels: step S12 filters out report generation requests corresponding to the descriptive summarization mode at the request level, while step S21 filters out positions outside dynamic logical placeholders at the vocabulary level, ensuring that the closed-loop verification overhead is precisely controlled within the dynamic logical placeholders that perform causal reasoning.

[0021] Step S13: When the report generation task execution mode judgment result is descriptive summary mode, static filling is performed on the key statistical indicator set to generate a static text skeleton, the summary statistical chart resource identifier corresponding to the key statistical indicator set is obtained, the summary statistical chart resource identifier is anchored to obtain the macro chart traceability anchor point, the macro chart traceability anchor point is pre-embedded in the static text skeleton, and a static text skeleton with macro chart traceability anchor point is obtained. Specifically, a static text skeleton is generated by using a static filling strategy for the set of key statistical indicators. The static filling strategy refers to the processing strategy of sequentially writing the statistical values ​​from the set of key statistical indicators into the filling slots according to a pre-configured filling template. The execution process of the static filling strategy does not involve secondary extraction of the underlying data entity label set, nor does it involve reverse mapping of the underlying original work orders and original processing records. The filling template is a text template composed of alternating natural language sentences and filling slots. A filling slot is a placeholder position reserved in the filling template for the statistical values ​​of the key statistical indicators. Based on the report generation task execution mode judgment result output in step S12, when the report generation task execution mode judgment result is descriptive summary mode, for each key statistical indicator in the set of key statistical indicators obtained in step S11, a pre-configured filling template is retrieved according to the statistical scope and statistical period. The statistical values ​​of the key statistical indicators are written into the corresponding filling slots in the filling templates. All filling templates are then concatenated according to the order of administrative division dimension, problem category dimension, and time window dimension to obtain the static text skeleton. The static text skeleton contains only macro-statistical values ​​and connective natural language sentences, but does not contain causal explanatory sentences or evidence-supporting sentences.

[0022] Based on a static text skeleton, resource identifiers for summary statistical charts corresponding to the key statistical indicator set are extracted. These resource identifiers are resource address strings for the visualization charts pre-generated by the multi-source heterogeneous data governance module for the key statistical indicator set. Each resource identifier uniquely points to a visualization chart, which can be of various types, including line charts, bar charts, and pie charts. A macro-anchor encapsulation algorithm is used to generate macro-chart traceability anchors from the resource identifiers. The algorithm reads the resource address string from the resource identifier and encapsulates it into the internal data structure of the macro-chart traceability anchor. The internal data structure of the macro-chart traceability anchor only includes the resource identifier field and does not contain work order identifiers, processing record identifiers, or encrypted address identifiers. The macro-chart traceability anchors are then written into the metadata fields of the corresponding fill slots in the static text skeleton, resulting in a static text skeleton with macro-chart traceability anchors.

[0023] For example, when the set of key statistical indicators includes two key statistical indicators: the total number of parking difficulty work orders in a certain district this month and the month-on-month change rate of parking difficulty work orders in a certain district this month, the filling template expression for the former is "the total number of parking difficulty work orders in a certain district this month is a fill slot", and the filled static text skeleton fragment is "the total number of parking difficulty work orders in a certain district this month is 800". The filling template expression for the latter is "the month-on-month change rate of parking difficulty work orders in a certain district is a fill slot", and the filled static text skeleton fragment is "the month-on-month change rate of parking difficulty work orders in a certain district is -20%". The metadata field of the filled static text skeleton fragment is pre-embedded with macro chart source anchors pointing to the bar chart of total work order statistics and the line chart of month-on-month change rate statistics.

[0024] The reason for using a combination of static fill strategy and macro chart source anchors in descriptive summary mode is that the report generation request corresponding to descriptive summary mode does not contain keywords for causal reasoning and does not require drill-down queries. This type of report generation request is mainly used to output the overall trend of macro statistical values ​​rather than the details of a single low-level original work order. The visualization chart pointed to by the macro chart source anchor is sufficient to meet the visualization jump requirements of the report recipient corresponding to descriptive summary mode. The beneficial effect of the static fill strategy and macro chart source anchor processing method is that it retains the visualization jump entry to the summary statistical chart resource identifier without introducing the storage and rendering overhead brought by entity-level source anchors, so that the report output in descriptive summary mode has both generation speed and visualization jump capability. When the report generation task execution mode judgment result is descriptive summary mode, the static text skeleton with macro chart source anchors is output as the descriptive report text stream to the report rendering end, and the closed-loop verification mechanism of step S20 is not executed; when the report generation task execution mode judgment result is causal reasoning judgment mode, step S14 is executed and then step S20 is entered. If step S13 is missing, the report generation request corresponding to the descriptive summary mode will have to follow the heavy-load path corresponding to step S14, resulting in wasted storage and rendering overhead. Steps S13 and S14 together constitute two parallel branches downstream of step S10. The two parallel branches do not intersect, ensuring that step S10, when responding to different report generation requests, not only guarantees the clear direction of the skeleton generation path but also avoids logical overlap between branches.

[0025] Step S14: When the report generation task execution mode judgment result is the causal reasoning analysis mode, a dynamic logical placeholder insertion strategy is adopted for the key statistical indicator set and the underlying data entity label set to generate a structured causal reasoning skeleton with dynamic logical placeholders. The structured causal reasoning skeleton is composed of three types of placeholder positions arranged alternately: conclusive statement position, cause explanation position, and evidence support position. Each type of placeholder position is represented by a dynamic logical placeholder. Based on the underlying data entity label set, an encrypted addressing generation algorithm is used to generate entity-level traceability anchors for each dynamic logical placeholder in the structured causal reasoning skeleton, and the entity-level traceability anchors are written into the metadata field of the dynamic logical placeholders to obtain a structured causal reasoning skeleton with entity-level traceability anchors. The entity-level traceability anchors include work order identifiers, processing record identifiers, and encrypted addressing identifiers.

[0026] Specifically, see Figure 4 This is a schematic diagram of the structured causal reasoning skeleton provided in the embodiments of this application. Figure 4 The diagram illustrates the composition and relationship between the structured causal reasoning framework and entity-level source tracing anchors. The dynamic logic placeholder insertion strategy refers to the processing strategy of inserting dynamic logic placeholders into three types of placeholder positions—conclusion statement positions, causal explanation positions, and evidence support positions—according to a pre-configured causal reasoning framework template. The conclusion statement position is the placeholder that carries the conclusive expression of the governance situation in the report; the causal explanation position is the placeholder that carries the explanatory expression of the causes behind the conclusion statement in the report; and the evidence support position is the placeholder that carries the guiding expression of the evidence materials cited in the causal explanation in the report. These three types of placeholder positions are arranged alternately in the structured causal reasoning framework in the order of conclusion statement position, causal explanation position, and evidence support position, combined with… Figure 4 As shown, each type of placeholder is represented by a dynamic logical placeholder. During the autoregressive text filling stage of the government hotline big data model, the dynamic logical placeholder will be filled with the corresponding candidate natural language conclusion words. Based on the report generation task execution mode judgment result output in step S12, when the report generation task execution mode judgment result is the causal reasoning judgment mode, a conclusive statement position is generated for each key statistical indicator according to the correspondence between the key statistical indicator set and the underlying data entity label set. Then, based on the work order classification label and processing record label associated with the key statistical indicator, a cause explanation position paired with the conclusive statement position is generated. Finally, based on the work order classification label and follow-up evaluation label associated with the key statistical indicator, an evidence support position paired with the cause explanation position is generated. The three types of placeholders are then assembled according to the pre-configured causal reasoning skeleton template to obtain a structured causal reasoning skeleton with dynamic logical placeholders.

[0027] Based on the underlying data entity tag set, an entity-level tracing anchor point is generated for each dynamic logical placeholder in the structured causal reasoning skeleton using an encrypted addressing generation algorithm. The execution process of the encrypted addressing generation algorithm is as follows: The work order category tag and processing record tag corresponding to the current dynamic logical placeholder are read from the underlying data entity tag set; a unique work order identifier pointing to the underlying original work order is extracted from the work order category tag, and a unique processing record identifier pointing to the original processing record is extracted from the processing record tag; a one-way hash operation is performed on the work order identifier and processing record identifier to obtain a digest string; the digest string is concatenated with a random salt value and then processed by a key pair encryption algorithm to obtain the encrypted addressing identifier. The encrypted addressing identifier is used to establish a binding with the click event handler of the corresponding rendering element in the report UI interface during the subsequent rendering stage. The encrypted addressing identifier itself does not expose the plaintext of the work order identifier and processing record identifier. See also... Figure 4 , Figure 4 The document clearly illustrates the specific information contained in the entity-level traceability anchor point, namely, writing the work order identifier, processing record identifier, and encrypted addressing identifier into the internal data structure of the entity-level traceability anchor point, and... Figure 4 The association between dynamic logical placeholders and entity-level traceability anchors illustrates the writing of entity-level traceability anchors into the metadata fields of dynamic logical placeholders, resulting in a structured causal reasoning skeleton with entity-level traceability anchors. The one-way hash operation and key pair encryption algorithm involved in the encrypted addressing generation algorithm are well-known computational tools in the field. The one-way hash operation has irreversible and collision-resistant characteristics, while the key pair encryption algorithm uses a public key on the encryption side and a private key on the decryption side. In the entity-level traceability anchor, the work order identifier is assigned by the multi-source heterogeneous data governance module when the work order is entered into the database, and the processing record identifier is assigned by the multi-source heterogeneous data governance module when the processing flow node is written; both are globally unique identifiers. The random salt value is obtained by concatenating the request identifier of the report generation request with the timestamp of the report generation time to obtain the original salt value string. After performing a one-way hash operation on the original salt value string, the first fixed-length bytes are truncated as the random salt value.

[0028] The method for setting the byte length of the random salt value is as follows: During the calibration phase, random salt values ​​with different byte lengths are generated, and the collision rate of the encrypted addressing identifier is statistically analyzed. The calibration phase refers to the offline phase in which parameters are pre-determined before the official launch of the business system. The samples used in the calibration phase are randomly selected from the historical work orders and historical processing records in the government hotline field. The calibration phase does not involve writing data to the official operating environment. The collision rate refers to the proportion of duplicate values ​​in the encrypted addressing identifiers obtained after performing the encrypted addressing generation algorithm on the same number of work order identifiers and processing record identifiers. The acceptable engineering range refers to the upper limit range of the collision rate set before the business goes live based on the requirements of the downstream report UI interface for the accuracy of click-through. The smallest byte length corresponding to the collision rate falling within the acceptable engineering range is selected as the byte length of the random salt value. When the random salt value is too small, the collision rate of the encrypted addressing identifier increases, and different underlying original work orders may generate the same encrypted addressing identifier, causing the click navigation of the corresponding rendered element in the report UI to point to the wrong underlying original work order. When the random salt value is too large, the byte length of the encrypted addressing identifier increases, and the storage overhead of the rendered element increases accordingly. For example, when the minimum byte length corresponding to the collision rate falling within the acceptable range during the calibration phase is 16 bytes, the random salt value byte length is set to 16 bytes.

[0029] For example, when the key statistical indicator corresponding to the conclusive statement position is a 20% month-on-month decrease in parking difficulty work orders in a certain district, the cause explanation position paired with the conclusive statement position is used to express the reason for the month-on-month decrease in parking difficulty work orders, and the evidence support position paired with the cause explanation position is used to support the follow-up evaluation and original processing record guidance for the month-on-month decrease in parking difficulty work orders; the entity-level traceability anchor points corresponding to the three dynamic logical placeholders are respectively bound to the work order identifiers of multiple underlying original work orders corresponding to the parking difficulty work order classification label, the processing record identifiers of multiple original processing records corresponding to the parking difficulty processing record label, and the encrypted addressing identifiers corresponding to the work order identifiers and processing record identifiers.

[0030] The processing method combining dynamic logical placeholder insertion and encrypted addressing generation algorithm is adopted because the government hotline large model generates natural language words character by character using continuous semantic features during the autoregressive text filling process. If entity-level traceability anchors are not pre-embedded before the autoregressive text filling begins, the correspondence between continuous semantic features and discrete underlying original work order entities will be overwritten during the autoregressive text filling process. The beneficial effect of the dynamic logical placeholder insertion strategy and encrypted addressing generation algorithm is that the candidate natural language conclusion words always coexist with the entity-level traceability anchors during the autoregressive text filling process, establishing a correspondence path carrying addressing information between continuous semantic features and discrete underlying original work order entities; at the same time, the encrypted addressing identifier does not expose the plaintext of the work order identifier and the processing record identifier, and the entity information of the underlying original work order and original processing record will not be leaked from the metadata field of the dynamic logical placeholder. If step S14 is missing, the closed-loop verification mechanism in step S20 will lose the entity-level traceability anchors used for verification, and the bidirectional closed-loop verification corresponding to step S20 cannot be established. The entity-level tracing anchor point generated in step S14 is used as the starting point for the connected path search in steps S21 to S23, and also as the addressing source for the rendering element click event handler binding in step S24.

[0031] In the intelligent analysis report generation scenario of the government hotline field, decision-makers often need to click on the macro statistical conclusions in the report to penetrate and view the specific work order details and resolution cases. However, conventional automatic report generation methods treat the government hotline large model as a black box decoder. In the autoregressive text filling process of the unidirectional output mapping path, the data pointers corresponding to the underlying data entity labels are discarded and overwritten. This leads to an irreversible topological break between the continuous semantic features corresponding to the text generation space and the discrete entity relationship graph corresponding to the data source space. As a result, when responding to drill-down requests, it often jumps to a blank page, which greatly weakens the credibility of the output results of the artificial intelligence analysis platform. To overcome this dilemma, this embodiment proposes to pre-establish a structured skeleton that can carry the underlying data pointers in step S10 before the government hotline large model starts autoregressive text filling. At the beginning of the process, the key statistical indicator set and its associated underlying data entity label set under the target analysis dimension are extracted at once, thereby avoiding the interruption of the decoding rhythm of the large model due to repeated retrieval of underlying data in subsequent generation. Subsequently, based on the task pattern recognition rules, the presence of causal reasoning keywords and the requirement for drill-down queries are used as the primary judgment criteria, while the target reader level is used as an auxiliary verification criterion. The report generation request is intelligently diverted into either a descriptive summary mode or a causal reasoning analysis mode. When the target reader level does not match the judgment result of the report generation task execution mode determined by the primary judgment criteria, a pattern verification prompt is generated to allow the user to confirm whether the report generation request needs to supplement causal reasoning keywords or drill-down query requirements. The pattern verification prompt does not change the judgment result of the report generation task execution mode already determined by the primary judgment criteria. This not only provides branch criteria for the skeleton generation path but also takes into account the throughput of report generation and the differentiated traceability granularity requirements on a global scale.

[0032] After initial data extraction and pattern segmentation, differentiated skeleton pre-embedding strategies are adopted for different task modes. When the report generation request is determined to be in descriptive summary mode, a lightweight path is executed, using a static filling strategy to directly write statistical values ​​into the fill slots, and pre-embedding macro-level chart traceability anchors pointing to the visualization charts in the metadata field, thus meeting the visualization jump requirements of the macro situation without increasing entity-level traceability overhead. When the report generation request is determined to be in causal reasoning and judgment mode, a heavy-load path is adopted, using a dynamic logical placeholder insertion strategy to generate a structured causal reasoning skeleton composed of alternating conclusive statement positions, causal explanation positions, and evidence support positions. More importantly, through an encrypted addressing generation algorithm, the extracted work order identifier and processing record identifier are compressed into encrypted addressing identifiers and written into the metadata field of the aforementioned dynamic logical placeholders, thereby forming entity-level traceability anchors. This pre-embedded mechanism not only ensures that the large model always has a scaffold to carry the underlying data pointers during subsequent decoding, but also protects the security of the original entity information by using one-way hashing and random salt encryption technology. While reporting the drilling-down traceability capability to the front-end UI interface, it also prevents unauthorized leakage of the underlying data.

[0033] Once the structured causal reasoning skeleton with entity-level source anchors is built and fed into the government hotline's large model, step S20 initiates a controlled text decoding mechanism based on bidirectional constraints of the knowledge graph during the autoregressive text filling stage. Faced with continuously generated natural language vocabulary, instead of blindly verifying all output content, a placeholder position recognition strategy is adopted. By comparing the lexical pointer with the pre-embedded start and end markers in the skeleton, the current filling position is judged in real time. Once it is identified that the current position is inside a dynamic logical placeholder, a closed-loop verification mechanism is immediately triggered, temporarily storing the generated candidate natural language conclusion vocabulary in the verification buffer; if it is outside, the connective natural language sentence structure is directly allowed to the report text stream, thus accurately converging the closed-loop verification overhead at the core position of causal judgment, effectively ensuring the overall decoding throughput. For candidate vocabulary temporarily stored in the verification buffer, a named entity recognition algorithm is then executed to extract semantic entities, and matching graph nodes are found in the knowledge graph maintained by the government hotline knowledge data management module through the graph node retrieval interface. This reverse mapping process successfully bridges the gap between the continuous semantic feature space and the discrete entity knowledge graph, establishing clear comparable endpoints for subsequent connectivity path verification.

[0034] After completing the reverse mapping, substantive causal correlation verification and conditional branching processing are immediately implemented, fundamentally breaking the black-box illusion of conventional large-scale models. Based on the work order identifiers pre-embedded in the entity-level source anchors, the underlying original work order entity node is located as the endpoint. A knowledge graph path search algorithm constrained by a preset hop count threshold is used to explore the connection path between the graph node corresponding to the candidate word and the underlying original work order entity node in the knowledge graph. The finite hop count constraint prevents the infinite spread of the search depth and strictly filters out pseudo-connections that may be caused by weakly associated nodes. Finally, conditional branching processing is performed on the candidate natural language conclusion words based on the connection path search results: if no real connection path exists, the output of the word is rejected and a constraint signal is fed back to the government hotline large-scale model, forcing it to resample until an expression with real data support is found; if multiple resampling attempts are unsuccessful, a default placeholder is filled in the original position and the result is submitted for manual review. Conversely, if the path verification passes, the candidate natural language conclusion word is incorporated into the report text stream, and the encrypted addressing identifier is simultaneously read from the source anchor and written into the click event handler of the corresponding rendered element. Through the aforementioned progressive two-way comparison closed loop, not only was the persistent problem of fabricating causal relationships in the government hotline's large model driven by fluency corrected, but also a visualized and verifiable two-way penetrating traceability channel was constructed for the decision-making level in the final report output.

[0035] Step S20: Input the structured causal reasoning skeleton with entity-level traceability anchors into the government hotline big model for autoregressive text filling to generate candidate natural language conclusion words; map the candidate natural language conclusion words inversely to a set of graph nodes in the knowledge graph; search for the connection path between the set of graph nodes and the underlying original work order entity node corresponding to the entity-level traceability anchor; perform conditional branch processing on the candidate natural language conclusion words according to the connection path search results to achieve bidirectional traceability between the candidate natural language conclusion words and the underlying original work order.

[0036] Further, step S20 includes: Step S21: Input the structured causal reasoning skeleton with entity-level source tracing anchors into the government hotline big model for autoregressive text filling. Use a placeholder position recognition strategy to identify the current filling position in real time and obtain the position recognition result. When the position recognition result is that the current filling position is inside the dynamic logical placeholder, trigger the closed-loop verification mechanism and temporarily store the candidate natural language conclusion words generated by the government hotline big model in the verification buffer. When the position recognition result is that the current filling position is outside the dynamic logical placeholder, directly output the generated filling words to the report text stream. Step S20 is used to ensure that when the government hotline big model performs autoregressive text filling on the structured causal reasoning skeleton with entity-level traceability anchors generated in step S14, the candidate natural language conclusion words obtained after filling and the underlying original work order uniquely pointed to by the entity-level traceability anchors have a connected path in the knowledge graph maintained by the government hotline knowledge data management module, so that the final rendering product maintains bidirectional traceability between the candidate natural language conclusion words and the underlying original work order. Specifically, autoregressive text filling refers to the decoding process of the government hotline big model generating natural language text word by word in a left-to-right order. The placeholder position recognition strategy refers to the strategy of real-time judgment of the boundary relationship between the current filling position of the government hotline big model and the dynamic logical placeholder; in the pre-embedding stage of step S14, each dynamic logical placeholder is assigned a start mark and an end mark. The start mark and end mark are embedded in the structured causal reasoning skeleton in the form of special words and retained in the internal word pointer trajectory, and do not participate in the natural language output of the government hotline big model. The validation buffer is a memory area reserved for temporarily storing candidate natural language conclusion words. The validation buffer is connected to the output stream of the government hotline large-scale model via a gating switch. The gating switch is a switching component that switches the direction of the government hotline large-scale model's output stream based on the position recognition result. The gating switch has two working states: a temporary storage path and a direct path. The opening and closing of the gating switch is driven by the position recognition result output by the placeholder position recognition strategy. The report text stream is a data stream that continuously outputs confirmed words to the report rendering end.

[0037] A structured causal reasoning skeleton with entity-level source tracing anchors is input into the government hotline big data model. The model generates candidate natural language conclusion terms one by one according to an autoregressive text filling method. When each candidate natural language conclusion term is generated, a placeholder position recognition strategy is used to identify the current filling position in real time to obtain the position recognition result. The placeholder position recognition strategy is implemented as follows: read the current word pointer of the government hotline big data model, and compare the current word pointer with the number of start markers and the number of end markers that have been passed. When the number of start markers is greater than the number of end markers, the position recognition result is recorded as inside the dynamic logical placeholder; when the two numbers are equal, the position recognition result is recorded as outside the dynamic logical placeholder. When the position recognition result is inside the dynamic logical placeholder, a closed-loop verification mechanism is triggered, and the candidate natural language conclusion terms generated by the government hotline big data model and the entity-level source tracing anchors in the corresponding dynamic logical placeholder metadata fields are synchronously and temporarily stored in the verification buffer. When the position recognition result is outside the dynamic logical placeholder, the generated filling terms are directly output to the report text stream.

[0038] The reason for adopting a placeholder position recognition strategy combined with gating switches and verification buffers is that the autoregressive text filling of the government hotline large model progresses continuously in the form of a continuous word stream. Candidate natural language conclusion words and connective natural language sentences appear alternately in the word stream. Only candidate natural language conclusion words carrying causal inference conclusions need to undergo closed-loop verification in steps S22 to S24. If a closed-loop verification mechanism is uniformly enabled for all filling words, the filling words corresponding to connective natural language sentences will be repeatedly sent to the reverse mapping and connectivity path search stages. The latency of the verification stage is added to each filling word, which slows down the overall decoding speed. The beneficial effect of the differentiated flow processing method based on position recognition results is that it accurately converges the closed-loop verification overhead to the internal position of the dynamic logical placeholder carrying the causal inference conclusion, ensuring both the traceability of causal conclusions and the decoding throughput of connective natural language sentences.

[0039] For example, when a cause explanation position in the structured causal reasoning skeleton is defined by a start marker and an end marker, the multiple candidate natural language conclusion words generated by the government hotline big model after the start marker in the autoregressive text filling stage are all located inside the cause explanation position and are temporarily stored one by one in the verification buffer. When the government hotline big model continues to fill until the end marker, the position recognition result switches to outside the dynamic logical placeholder, and the subsequently generated connective natural language sentences no longer trigger the closed-loop verification mechanism and are directly output to the report text stream. If step S21 is missing, the government hotline big model cannot distinguish between the positions inside and outside the dynamic logical placeholder, and the reverse mapping in step S22 and the connectivity path verification in step S23 will lose their trigger criteria.

[0040] Step S22: Semantic entity extraction is performed on the candidate natural language conclusion words in the verification buffer to obtain a set of semantic entities. The graph nodes that match the set of semantic entities are searched in the knowledge graph maintained by the government hotline knowledge data management module to obtain the set of graph nodes corresponding to the candidate natural language conclusion words, thus completing the reverse mapping of the candidate natural language conclusion words to the knowledge graph nodes. Specifically, semantic entity extraction refers to the process of identifying the corresponding terms representing entity categories such as people, events, time, place, and attributes from candidate natural language conclusion vocabulary. Semantic entity extraction employs the well-known named entity recognition algorithm, which learns the semantic features of various entities based on labeled training corpora and outputs an entity category label for each lexical unit in the input text during the inference phase. Each semantic entity in the semantic entity set consists of two attribute fields: an entity term and an entity category label. The knowledge graph is a semantic network maintained by the government hotline knowledge data management module, with underlying original work order entity nodes, original processing record entity nodes, follow-up evaluation entity nodes, and policy document entity nodes as node types, and business relationships as edges. The graph node retrieval interface is an interface provided by the government hotline knowledge data management module. The input parameter is the semantic entity set, and the return result is a set of graph nodes that match the semantic entity set. Based on the candidate natural language conclusion words temporarily stored in the verification buffer in step S21, a named entity recognition algorithm is used to extract semantic entities from the candidate natural language conclusion words to obtain a semantic entity set. For each semantic entity in the semantic entity set, node matching is performed in the knowledge graph through the graph node retrieval interface. The node matching determination method is as follows: the string similarity between the entity term of the semantic entity and the named attribute field of the node in the knowledge graph is calculated, and the category consistency between the entity category label of the semantic entity and the type attribute field of the node in the knowledge graph is compared. When the string similarity is greater than or equal to the preset string similarity threshold and the category consistency comparison result is the same category, the node is included in the graph node set corresponding to the candidate natural language conclusion word. The string similarity calculation adopts the text similarity calculation method known in the art, including but not limited to the edit distance comparison method.

[0041] The preset string similarity threshold is set as follows: Recall and precision are statistically analyzed for different values ​​of the string similarity threshold on manually aligned samples in the field of government hotlines. The manually aligned samples are alignment pairs pre-established between historical candidate natural language conclusion words and historical graph nodes by annotators with knowledge of government hotline business. Recall is the ratio of the number of candidate samples correctly recalled as graph node sets to the total number of correct alignment pairs. Precision is the ratio of the number of correct alignment pairs among the candidate samples recalled as graph node sets to the total number of candidate samples recalled as graph node sets. The string similarity value corresponding to the peak value of the product of recall and precision is selected as the preset string similarity threshold. If the preset string similarity threshold is too small, graph nodes that are similar in form to semantic entity terms but semantically unrelated will be mistakenly included in the graph node set, causing false connected paths to be generated in the connectivity path verification step S23. If the preset string similarity threshold is too large, legitimate corresponding graph nodes will be excluded from the graph node set due to subtle string differences, and legitimate candidate natural language conclusion words will be incorrectly rejected. For example, when the peak value of the product of recall and precision obtained from manually aligned samples corresponds to a string similarity value of 0.8, the preset string similarity threshold is set to 0.8.

[0042] The reason for adopting a processing method that combines semantic entity extraction with graph node retrieval is that the candidate natural language conclusion vocabulary belongs to a continuous semantic feature space, while the underlying original work order entity nodes belong to a discrete knowledge graph node space. There are no directly comparable semantic coordinates between the two. It is necessary to go through two steps, entity extraction and node retrieval, to establish a reverse mapping path from continuous semantic features to discrete graph nodes, providing comparable endpoints from the candidate natural language conclusion vocabulary side for the connectivity path verification in step S23. For example, when the candidate natural language conclusion vocabulary is "parking lot expansion", the semantic entity set obtained after semantic entity extraction includes two semantic entities: the entity term "parking lot" and the entity category label "facility class", and the entity term "expansion" and the entity category label "event class". By performing node matching in the knowledge graph through the graph node retrieval interface, the resulting graph node set includes graph nodes of the facility class with the named attribute field "parking lot in a certain area" and graph nodes of the event class with the named attribute field "parking lot expansion project in a certain area". If step S22 is missing, there will be no common semantic comparison endpoint between the candidate natural language conclusion words and the underlying original work order entity nodes, and the connectivity path verification in step S23 will not be able to start. The graph node set obtained in step S22 and the underlying original work order entity node uniquely pointed to by the entity-level tracing anchor in step S14 together constitute the two ends of the connectivity path verification, and form a homologous mapping relationship with the underlying data entity label set provided in step S11 at the data end, ensuring that the endpoint node of the reverse mapping and the endpoint node pointed to by the entity-level tracing anchor in the forward embedding are located in the same knowledge graph coordinate system.

[0043] Step S23: Read the work order identifier from the entity-level traceability anchor point, locate the underlying original work order entity node that the work order identifier uniquely points to in the knowledge graph maintained by the government hotline knowledge data management module based on the work order identifier, and use the knowledge graph path search algorithm to search for the connection path between the graph node set and the underlying original work order entity node in the knowledge graph maintained by the government hotline knowledge data management module to obtain the connection path search results. Specifically, a knowledge graph path search algorithm refers to an algorithm that searches for a sequence of paths connecting two endpoint nodes within a knowledge graph. Knowledge graph path search algorithms are implemented using either breadth-first search or depth-first search methods, both well-known in the field. Breadth-first search expands nodes layer by layer from the nearest to the farthest starting node, while depth-first search recursively moves forward along a single path until the endpoint node is reached or the maximum number of search layers is reached. The preset hop count threshold refers to the maximum number of edges allowed during the search process of the knowledge graph path search algorithm. Based on the graph node set obtained in step S22, and based on the bottom-level original work order entity node uniquely pointed to by the work order identifier in the entity-level traceability anchor point pre-embedded in step S14, the work order identifier is read from the entity-level traceability anchor point, and the bottom-level original work order entity node is located in the knowledge graph according to the work order identifier; a knowledge graph path search algorithm is adopted, taking each graph node in the graph node set as the starting node and the bottom-level original work order entity node as the ending node, and searching for the path sequence connecting the starting node and the ending node. During the search process, the number of search layers is constrained by a preset hop count threshold. When the number of search layers reaches the preset hop count threshold and the bottom-level original work order entity node is not reached, the current path branch is abandoned; the search is performed one by one on each graph node in the graph node set to obtain the connected path search results. When at least one graph node in the graph node set can be connected to the underlying original work order entity node under the preset hop count threshold constraint, the connection path search result is recorded as a connected path exists, and all nodes and edges traversed by the connected path are recorded as path details; when none of the graph nodes in the graph node set can be connected to the underlying original work order entity node under the preset hop count threshold constraint, the connection path search result is recorded as a connected path does not exist.

[0044] The preset hop count threshold ranges from 2 to 5. The method for setting the preset hop count threshold is as follows: Coverage statistics are performed on manually annotated samples of historical intelligent analysis reports in the government hotline field. These manually annotated samples are sets of aligned pairs annotated by annotators with knowledge of government hotline business, highlighting reasonable causal relationships between historical candidate natural language conclusion words and underlying original work orders. Coverage refers to the ratio of the number of reasonable causal relationships that can successfully find connected paths under different hop count thresholds to the total number of reasonable causal relationships when searching for reasonable causal relationships in the manually annotated samples using a knowledge graph path search algorithm. The minimum hop count corresponding to a coverage rate reaching the preset ratio is selected as the preset hop count threshold. The preset ratio ranges from 0.9 to 0.98. If the preset hop count threshold is set too low, the actual path length corresponding to the causal relationship between entities in the knowledge graph exceeds the preset hop count threshold. Reasonable causal relationships may be misjudged as non-existent paths because the paths have not been fully searched, and correct candidate natural language conclusion terms are incorrectly rejected. If the preset hop count threshold is set too high, graph nodes with only a weak association with candidate natural language conclusion terms may also form a connection path with the underlying original work order entity node through a long link, and incorrect candidate natural language conclusion terms may be misjudged as having legitimate tracing basis. For example, when the minimum hop count corresponding to a coverage rate of 0.95 obtained from manually labeled samples is 3, the preset hop count threshold is set to 3.

[0045] The reason for using a knowledge graph path search algorithm constrained by a preset hop count threshold is that causal relationships between entities in a knowledge graph can usually be established directly with a finite number of hops. Limited hop count search strikes a balance between the accuracy of path search and computational cost. Without hop count constraints, the knowledge graph path search algorithm would expand indefinitely throughout the entire knowledge graph, and the search time would increase exponentially with the size of the knowledge graph nodes. Long links between weakly related nodes would render the causal constraints ineffective. For example, when the set of graph nodes corresponding to candidate natural language conclusion terms includes the facility-type node "Parking Lot in a Certain Area" and the event-type node "Parking Lot Expansion Project in a Certain Area," under the constraint of a preset hop count threshold of 3, the knowledge graph path search algorithm finds a two-hop connected path from the "Parking Lot in a Certain Area" node through the "Parking Lot Expansion Project in a Certain Area" node to the underlying original work order entity node. The connected path search result is recorded as a connected path, and all nodes and edges traversed by the two-hop path are recorded as path details. Without step S23, there would be no knowledge graph-level correlation verification between the candidate natural language conclusion words and the underlying original work order entity nodes. Even if step S22 has reverse-mapped the candidate natural language conclusion words to the graph node set, it would be impossible to determine whether there is a reasonable causal relationship between the semantic entities corresponding to the graph node set and the underlying original work order entity nodes. Step S23 uses the connected path search results as the input for the subsequent conditional branch processing in step S24, and together with the graph node set obtained in step S22 and the entity-level tracing anchor points pre-embedded in step S14, it forms a two-way comparison closed loop between the generation end and the graph end.

[0046] Step S24: Perform conditional branching on the candidate natural language conclusion words temporarily stored in the verification buffer based on the connectivity path search results; when the connectivity path search results show no connectivity path, reject the output of the candidate natural language conclusion words to the report text stream, feed back a constraint signal to the government hotline big model, and force the government hotline big model to resample the candidate natural language conclusion words according to the constraint signal and return to step S22 for a new round of semantic entity extraction and reverse mapping; when the cumulative number of resampling times reaches the upper limit of the number of resampling times and the connectivity path search results still show no connectivity path, fill the default placeholder mark in the dynamic logic placeholder position and mark it as pending manual review; when the connectivity path search results show that a connectivity path exists, output the candidate natural language conclusion words to the report text stream, read the encrypted addressing identifier from the entity-level traceability anchor point, write the encrypted addressing identifier into the click event handler corresponding to the candidate natural language conclusion words, and complete the click jump binding configuration.

[0047] See Figure 5 This is a schematic diagram of a bidirectional traceability scenario provided in an embodiment of this application. Specifically, conditional branching refers to processing the temporarily stored data based on the connected path search results output in step S23. Figure 5The verification buffer shows the processing steps for different subsequent actions performed on candidate natural language conclusion words. Constraint signals are feedback signals recognizable by the government hotline big data model, used to notify the model that the current candidate natural language conclusion word failed the connectivity path verification and require resampling. The implementation of constraint signals includes, but is not limited to, setting the output probability of the current candidate natural language conclusion word to zero during the decoding bundle search of the government hotline big data model, forcing the model to sample the second-highest probability candidate natural language conclusion word. Default placeholders are pre-defined placeholder characters that are filled into the dynamic logic placeholder position when the cumulative resampling count reaches the upper limit and the connectivity path search result still shows no connectivity path, with an additional annotation indicating manual review. Click event handlers are event callback structures corresponding to the rendered elements of the reporting UI interface, executing pre-registered jump logic when the rendered element is clicked. Based on the connectivity path search results output in step S23, conditional branch processing is performed on the candidate natural language conclusion words temporarily stored in the verification buffer in step S21. When the search result for a connected path does not show a connected path, a rejection and resampling branch is executed: the output of candidate natural language conclusion words to the report text stream is rejected, and a constraint signal is fed back to the government hotline big model. The government hotline big model resamples candidate natural language conclusion words based on the constraint signal. The resampled candidate natural language conclusion words return to step S22 for a new round of semantic entity extraction and reverse mapping. After each round of resampling, the resampling count counter is incremented by 1. When the cumulative resampling count reaches the upper limit of the resampling count and the search result for a connected path still does not show a connected path, a default placeholder is filled in the dynamic logic placeholder position and marked as awaiting manual review, so that the report generation process does not fall into an infinite loop due to the deadlock of a single dynamic logic placeholder. When the search result for a connected path shows a connected path exists, an output and click jump branch is executed: combined with Figure 5 It can be seen that this branch will output candidate natural language conclusion words to the report text stream, read the encrypted addressing identifier from the entity-level traceability anchor point pre-embedded in step S14, write the encrypted addressing identifier into the click event handler corresponding to the candidate natural language conclusion words, and complete the click jump binding configuration; when the rendered element on the report UI interface is clicked, the click event handler initiates a retrieval request for the underlying original work order and original processing record to the government hotline knowledge data management module based on the encrypted addressing identifier, realizing the visual traceability path from the report text to the underlying original work order.

[0048] The method for setting the upper limit of the resampling count is as follows: Success rate and latency statistics are performed on offline playback samples of historical intelligent analysis reports in the government hotline domain for different values ​​of the upper limit of the resampling count. The offline playback samples consist of a sample set composed of historical report generation requests and corresponding historical candidate natural language conclusion terms. The success rate refers to the ratio of the number of candidate natural language conclusion terms that pass the connectivity path verification after a limited number of resampling rounds under different upper limits of the resampling count to the total number of candidate natural language conclusion terms. Latency refers to the average time taken to complete one round of autoregressive text filling and closed-loop verification mechanism coordination. The upper limit of the resampling count is selected at the inflection point where the success rate no longer increases. If the upper limit of the resampling count is too small, some candidate natural language conclusion terms are filled with default placeholders before sufficient resampling, increasing the number of placeholders requiring manual review in the report. If the upper limit of the resampling count is too large, the overall latency of the report generation process increases. For example, when the upper limit of the resampling count corresponding to the success rate inflection point obtained from the offline playback sample statistics is 3 times, the upper limit of the resampling count is set to 3 times.

[0049] The approach of combining a rejection and feedback resampling branch with an output and click-to-jump branch is adopted because the rationality of candidate natural language conclusion terms generated during the autoregressive text imputation process in the government hotline model is uncertain. If all candidate natural language conclusion terms are output uniformly, those lacking causal connectivity will enter the report text stream, leading to fabricated causal relationships in the report content. Conversely, if all candidate natural language conclusion terms are rejected uniformly, reasonable ones will also be excluded from the report, preventing the autoregressive text imputation from converging. The beneficial effect of conditional branching is that it both blocks fabricated causal relationships from entering the report text stream and preserves the output path for reasonable candidate natural language conclusion terms. The encrypted addressing identifier itself does not expose the plaintext of the work order identifier and the processing record identifier, ensuring that the underlying original work order and processing record entity information is only decrypted in the backend retrieval stage. Figure 5 For example, when the government hotline big model generates a traffic violation fine reduction at a certain cause explanation position, and there is no connected path between the traffic violation fine reduction and the underlying original work order entity node under the constraint of a preset hop count threshold of 3, step S24 rejects the output of the traffic violation fine reduction and feeds back the constraint signal; the government hotline big model resamples based on the constraint signal to obtain newly built parking spaces in a certain area, and the connected path search results obtained by steps S22 and S23 are recorded as having a connected path, such as... Figure 5 The diagram illustrates the connection paths established between entity nodes within the knowledge graph. In step S24, the newly created parking space in a certain area is output to the report text stream, and the encrypted addressing identifier is written to the click event handler of the corresponding rendered element. Figure 5This also demonstrates how the click event handler, based on encrypted addressing identifiers, successfully establishes a closed-loop mapping process between the underlying original work orders and original processing records. Without step S24, the connectivity search results output in step S23 cannot be transformed into real-time correction of the government hotline's large-scale model decoding behavior and click-based redirection of report UI rendering elements, rendering the report's credibility and traceability impossible. Step S24 transforms the connectivity search results into dual control signals for both the generation and rendering ends, which, together with the candidate natural language conclusion terms temporarily stored in the verification buffer in step S21 and the encrypted addressing identifiers pre-embedded in step S14, constitute a complete closed-loop exit from causal verification to rendering binding.

[0050] When the government hotline's large-scale model receives a structured causal reasoning skeleton with entity-level source anchors and initiates autoregressive text filling in its causal reasoning analysis mode, it needs to fill in specific causal explanation positions with words that can explain the overall situation to assist decision-makers in making scientific governance judgments. However, conventional automatic report generation methods often treat the large-scale model as an absolutely reliable generator, assuming that every candidate word it outputs corresponds to a real causal relationship. In reality, the large-scale model relies solely on word-by-word sampling based on probability distributions at the lexical level during the autoregressive decoding stage, completely unaware of the actual processing procedures recorded in the underlying original work orders. This decoding mechanism, driven solely by language fluency, is highly prone to generating seemingly fluent and natural content that is actually causally unrelated to the underlying original work orders. For example, when explaining the reasons for the easing of the "parking difficulty" problem, the large-scale model might fabricate the conclusion of "reduction or exemption of traffic violation fines," a conclusion that cannot be found in the underlying processing records at all. Since conventional methods do not establish a reverse comparison path between the generation end and the underlying knowledge graph, relying solely on contextual probability cannot identify such fabricated conclusions. In addition, conventional post-event rule review is separated from the autoregressive generation process, which not only fails to correct deviations in a timely manner during the decoding stage, but also results in the final rendering end being unable to obtain the binding information that can be clicked and redirected.

[0051] To overcome this technical bottleneck, this embodiment introduces a two-way comparison closed loop between the generation end and the knowledge graph end in the autoregressive text filling process. This mechanism does not perform post-verification only after all text has been generated; instead, it initiates real-time verification of the existence of a connection path between each candidate natural language conclusion word and the underlying original work order the moment the word is generated. The entire controlled decoding link proceeds progressively through four levels: location recognition trigger, reverse mapping, connection path verification, and conditional branch processing. First, based on the placeholder location recognition strategy, newly generated candidate words are accurately diverted to the verification buffer or directly imported into the report text stream. Then, using a named entity recognition algorithm, candidate words falling into the buffer are reverse-mapped into a set of nodes in the knowledge graph. Next, a path search algorithm constrained by a preset hop count threshold is used to explore the actual connection path between the graph node set and the underlying original work order entity nodes. Finally, based on the objective results of the path verification, the candidate word is either rejected and resampled, or output and bound to a click-through conditional branch. This interconnected two-way closed-loop design reshapes the decoding logic and traceability capabilities of the government hotline large-scale model. On the one hand, relying on runtime location identification and diversion processing, the computational overhead of closed-loop verification is precisely limited to the internal position of dynamic logic placeholders. In conventional large-scale model outputs, connective phrases and conclusive words are intertwined and mixed in probability distributions, often facing the dilemma of either blindly reviewing the entire system or completely abandoning the review process. However, this method, through clever real-time diversion, successfully embeds the verification mechanism seamlessly into the autoregressive decoding loop, without interrupting the generation rhythm of the large-scale model, and shifting the causal constraint checkpoint from post-event review to immediate feedback during generation, ensuring that the credibility of each conclusive word is firmly locked at the moment of generation. On the other hand, the conditional branch switching implemented based on the connectivity path verification results endows the entire generation process with the dual attributes of self-correction and traceability binding. Delving into its underlying mechanism, when a candidate word is detected to lack support from a real path, the rejection and resampling branch imposes a constraint signal on the large model, forcing it to proactively avoid those erroneous but high-probability word regions in the next round of sampling, thus achieving effective correction based on the real relationships of the knowledge graph. Conversely, when a candidate word passes path verification, the output and click-to-jump branch automatically writes the pre-embedded encrypted addressing identifier into the event handler of the corresponding rendering element. The collaborative operation of these two branches successfully transforms the large model from an unmanageable black-box decoder into a reliable generator strictly constrained by the knowledge graph and exposing penetrating traceability capabilities. Ultimately, this ensures that the decision-making level can safely and accurately reach the real scene of the underlying original work orders and processing records based on every causal conclusion in the report.

[0052] Example 2: This embodiment, based on Embodiment 1, provides a tag-based report generation system, such as... Figure 6 As shown, it includes: The skeleton generation module is used to receive report generation requests and parse them to obtain a set of key statistical indicators and a set of underlying data entity labels. It determines the report generation task execution mode and generates a report generation task execution mode judgment result. The report generation task execution mode judgment result is either a descriptive summary mode or a causal reasoning judgment mode. When the report generation task execution mode judgment result is a causal reasoning judgment mode, it generates a structured causal reasoning skeleton with entity-level source tracing anchors. Text Filling Module: Input the structured causal reasoning skeleton with entity-level source anchors into the government hotline big model for autoregressive text filling to generate candidate natural language conclusion words; and back-map the candidate natural language conclusion words to a set of graph nodes in the knowledge graph. Closed-loop verification module: used to search for the connection path between the graph node set and the underlying original work order entity node corresponding to the entity-level traceability anchor point. Based on the connection path search results, conditional branch processing is performed on the candidate natural language conclusion words to realize bidirectional traceability between the candidate natural language conclusion words and the underlying original work order.

[0053] Furthermore, in the skeleton generation module, the method for obtaining the set of key statistical indicators and the set of underlying data entity labels includes: The target analysis dimension in the parsing report generation request is analyzed. Key statistical indicators under the target analysis dimension of the government hotline are extracted from the government hotline knowledge data management module to obtain a set of key statistical indicators. The underlying data entity tags corresponding to the set of key statistical indicators are then extracted to obtain a set of underlying data entity tags.

[0054] The method for generating the task execution mode from the judgment report includes: Based on the report generation request corresponding to the set of key statistical indicators, the task pattern recognition rules are used to determine the execution mode of the report generation task, and the result of the report generation task execution mode determination is obtained. The judgment criteria of the task pattern recognition rules include three items: the first item is whether the report generation request contains keywords for causal reasoning; the second item is the target reader level marked in the report generation request; and the third item is whether the report generation request requires drill-down query. Among them, the first and third items are the main determination items of the report generation task execution mode, and the second item is the auxiliary verification item of the report generation task execution mode.

[0055] When the report generation task execution mode is determined to be descriptive summary mode, static text skeleton is generated by static filling of the key statistical indicator set. The resource identifier of the summary statistical chart corresponding to the key statistical indicator set is obtained. The summary statistical chart resource identifier is anchored to obtain the macro chart traceability anchor point. The macro chart traceability anchor point is pre-embedded in the static text skeleton to obtain the static text skeleton with the macro chart traceability anchor point.

[0056] When the report generation task execution mode judgment result is the causal reasoning judgment mode, a dynamic logical placeholder insertion strategy is adopted for the key statistical indicator set and the underlying data entity label set to generate a structured causal reasoning skeleton with dynamic logical placeholders. The structured causal reasoning skeleton is composed of three types of placeholder positions arranged alternately: conclusive statement position, cause explanation position, and evidence support position. Each type of placeholder position is represented by a dynamic logical placeholder.

[0057] The method for generating the structured causal reasoning skeleton with entity-level source tracing anchors includes: Based on the underlying data entity tag set, an entity-level tracing anchor is generated for each dynamic logical placeholder in the structured causal reasoning skeleton using an encrypted addressing generation algorithm. The entity-level tracing anchor is then written into the metadata field of the dynamic logical placeholder, resulting in a structured causal reasoning skeleton with entity-level tracing anchors. The entity-level tracing anchors include work order identifiers, processing record identifiers, and encrypted addressing identifiers.

[0058] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0059] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0060] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A report generation method driven by a tag system, characterized in that, The method includes: The system receives and parses a report generation request to obtain a set of key statistical indicators and a set of underlying data entity labels. It then determines the report generation task execution mode and generates a report generation task execution mode determination result, which can be either a descriptive summary mode or a causal reasoning analysis mode. When the report generation task execution mode judgment result is descriptive summary mode, static filling is performed on the key statistical indicator set to generate a static text skeleton, the summary statistical chart resource identifier corresponding to the key statistical indicator set is obtained, the summary statistical chart resource identifier is anchored to obtain the macro chart traceability anchor point, the macro chart traceability anchor point is pre-embedded in the static text skeleton, and a static text skeleton with macro chart traceability anchor point is obtained. When the report generation task execution mode judgment result is the causal reasoning judgment mode, a dynamic logical placeholder insertion strategy is adopted for the key statistical indicator set and the underlying data entity label set to generate a structured causal reasoning skeleton with dynamic logical placeholders. The structured causal reasoning skeleton is composed of three types of placeholder positions arranged alternately: conclusive statement position, cause explanation position, and evidence support position. Each type of placeholder position is represented by a dynamic logical placeholder. Based on the underlying data entity label set, an encrypted addressing generation algorithm is used to generate entity-level traceability anchors for each dynamic logical placeholder in the structured causal reasoning skeleton. The entity-level traceability anchors are written into the metadata field of the dynamic logical placeholders to obtain a structured causal reasoning skeleton with entity-level traceability anchors. The entity-level traceability anchors include work order identifiers, processing record identifiers, and encrypted addressing identifiers. The structured causal reasoning skeleton with entity-level traceability anchors is input into the government hotline big model for autoregressive text filling to generate candidate natural language conclusion words. The candidate natural language conclusion words are then mapped inversely to a set of graph nodes in a knowledge graph. The connection path between the set of graph nodes and the underlying original work order entity nodes corresponding to the entity-level traceability anchors is searched. Based on the search results of the connection path, conditional branching is performed on the candidate natural language conclusion words to achieve bidirectional traceability between the candidate natural language conclusion words and the underlying original work orders. The conditional branching method includes: when the connected path search result indicates that no connected path exists, rejecting the output of candidate natural language conclusion words to the report text stream, feeding back a constraint signal to the government hotline big model, and forcing the government hotline big model to resample candidate natural language conclusion words according to the constraint signal; when the connected path search result indicates that a connected path exists, outputting candidate natural language conclusion words to the report text stream, reading the encrypted addressing identifier from the entity-level tracing anchor point, and writing the encrypted addressing identifier into the click event handler corresponding to the candidate natural language conclusion word.

2. The report generation method driven by a tag system according to claim 1, characterized in that, The methods for obtaining the set of key statistical indicators and the set of underlying data entity labels include: The target analysis dimension in the parsing report generation request is analyzed. Key statistical indicators under the target analysis dimension of the government hotline are extracted from the government hotline knowledge data management module to obtain a set of key statistical indicators. The underlying data entity tags corresponding to the set of key statistical indicators are then extracted to obtain a set of underlying data entity tags.

3. The report generation method driven by a tag system according to claim 1, characterized in that, The method for generating the task execution mode from the judgment report includes: Based on the report generation request corresponding to the set of key statistical indicators, the task pattern recognition rules are used to determine the execution mode of the report generation task, and the result of the report generation task execution mode determination is obtained. The judgment criteria of the task pattern recognition rules include three items: the first item is whether the report generation request contains keywords for causal reasoning; the second item is the target reader level marked in the report generation request; and the third item is whether the report generation request requires drill-down query. Among them, the first and third items are the main determination items of the report generation task execution mode, and the second item is the auxiliary verification item of the report generation task execution mode.

4. The report generation method driven by a tag system according to claim 1, characterized in that, The underlying data entity tag set includes four categories: work order classification tags, processing record tags, follow-up evaluation tags, and policy document tags. A one-to-one mapping relationship is established between work order classification tags and underlying original work orders, a one-to-one mapping relationship is established between processing record tags and original processing records, a one-to-one mapping relationship is established between follow-up evaluation tags and follow-up evaluation records, and a one-to-one mapping relationship is established between policy document tags and policy documents.

5. The report generation method driven by a tag system according to claim 4, characterized in that, The execution process of the encrypted addressing generation algorithm is as follows: read the work order category label and processing record label corresponding to the current dynamic logical placeholder from the underlying data entity label set; extract the work order identifier that uniquely points to the underlying original work order corresponding to the work order category label, and extract the processing record identifier that uniquely points to the original processing record corresponding to the processing record label; perform a one-way hash operation on the work order identifier and the processing record identifier to obtain a digest string, and concatenate the digest string with a random salt value and process it through the key pair encryption algorithm to obtain the encrypted addressing identifier.

6. The report generation method driven by a tag system according to claim 1, characterized in that, The autoregressive text filling method includes: A structured causal reasoning skeleton with entity-level source tracing anchors is input into the government hotline big data model for autoregressive text filling. A placeholder position recognition strategy is used to identify the current filling position in real time to obtain the position recognition result. When the position recognition result is that the current filling position is inside the dynamic logical placeholder, a closed-loop verification mechanism is triggered, and the candidate natural language conclusion words generated by the government hotline big data model are temporarily stored in the verification buffer. When the position recognition result is that the current filling position is outside the dynamic logical placeholder, the generated filling words are directly output to the report text stream.

7. The report generation method driven by a tag system according to claim 6, characterized in that, The reverse mapping method includes: Semantic entity extraction is performed on the candidate natural language conclusion words in the verification buffer to obtain a set of semantic entities. Then, the graph nodes that match the set of semantic entities are searched in the knowledge graph maintained by the government hotline knowledge data management module to obtain the set of graph nodes corresponding to the candidate natural language conclusion words.

8. A tag-based report generation system, used to implement the tag-based report generation method according to any one of claims 1-7, characterized in that, The system includes: The skeleton generation module is used to receive report generation requests and parse them to obtain a set of key statistical indicators and a set of underlying data entity labels. It determines the report generation task execution mode and generates a report generation task execution mode judgment result. The report generation task execution mode judgment result is either a descriptive summary mode or a causal reasoning judgment mode. When the report generation task execution mode judgment result is a causal reasoning judgment mode, it generates a structured causal reasoning skeleton with entity-level source tracing anchors. Text Filling Module: Input the structured causal reasoning skeleton with entity-level source anchors into the government hotline big model for autoregressive text filling to generate candidate natural language conclusion words; and back-map the candidate natural language conclusion words to a set of graph nodes in the knowledge graph. Closed-loop verification module: used to search for the connection path between the graph node set and the underlying original work order entity node corresponding to the entity-level traceability anchor point. Based on the connection path search results, conditional branch processing is performed on the candidate natural language conclusion words to realize bidirectional traceability between the candidate natural language conclusion words and the underlying original work order.

Images