Government affair work order intelligent retrieval method based on multi-dimensional feature fusion, adaptive query optimization and intention sampling mining and electronic equipment

Abstract

The invention discloses a government affair work order intelligent retrieval method and electronic equipment based on multi-dimensional feature fusion, adaptive query optimization and intention sampling mining, and the method comprises the steps: constructing a government affair work order multi-dimensional classification system, and constructing a government affair worker one-way quantitative data retrieval system; the system comprises a preorder module, a model intention extraction module, screening and arranging dictionary extraction, multi-dimensional SQL query, ID aggregation, preliminary screening and linear weighting, secondary screening and sorting, result sampling, intention mining, intention enhancement analysis, intention fusion, result fusion and cyclic sampling mining, and result output and analysis. According to the method, high recall ratio and high precision ratio of government affair work order query are realized, and the method has important application value in work order processing in the fields of government hotlines, customer services and the like.

Description

Technical Field

[0001] This invention belongs to the field of computer information retrieval and artificial intelligence technology, specifically relating to a government work order intelligent retrieval method, system and electronic equipment based on multi-dimensional feature fusion, adaptive query optimization and intent sampling mining. Background Technology

[0002] In recent years, with the deepening of the digital transformation of government services, work order processing systems in areas such as government hotlines and customer service have faced the technical challenges of processing massive amounts of data and achieving accurate retrieval. As a key technological support for improving the efficiency of government services, the core of an intelligent government work order query system lies in how to accurately extract user intent from complex natural language queries and achieve efficient and accurate retrieval in a multi-dimensional work order database. Traditional government work order query systems are usually based on fixed-dimensional structured searches, such as filtering by time range, administrative region, or preset categories. While this method is simple to implement, it struggles to handle complex natural language queries and cannot meet the diverse query needs in practical applications.

[0003] Currently, technological developments in the field of government service order inquiries mainly focus on the following aspects: First, traditional retrieval technology based on keyword matching, which performs full-text searches in the service order database using keywords entered by the user; second, query parsing technology based on rule templates, which parses user queries using preset rule templates; and third, classification retrieval technology based on simple machine learning, which automatically categorizes service orders by training classification models. These technologies have improved the automation level of service order inquiries to some extent, but significant shortcomings still exist in practical applications.

[0004] The shortcomings of existing technologies are mainly reflected in the following aspects: First, traditional keyword retrieval methods cannot understand the semantic connotation of the query, resulting in low precision and often returning a large number of irrelevant results. Second, query parsing technology based on rule templates lacks flexibility and struggles to cover complex and varied natural language expressions, making it difficult to guarantee recall. Third, most existing ranking mechanisms are based on simple time-reverse order or keyword frequency sorting, lacking a comprehensive evaluation of the relevance of query results and failing to ensure that the most relevant results are displayed first. Furthermore, when handling multi-dimensional query conditions, existing systems often use fixed-dimensional linear filtering, failing to achieve adaptive fusion and optimization of multi-layer features. Finally, most systems lack intelligent analysis and visualization of query results, failing to provide users with in-depth insight support.

[0005] To overcome these limitations, the industry has begun exploring the application of artificial intelligence technology in the field of government service order queries. For example, some studies have attempted to use deep learning models for query intent recognition and extract semantic features of queries through neural networks. Other studies have tried to introduce vector retrieval technology into service order queries, converting service order content into vector representations through text embedding models to achieve semantic similarity matching. However, these technologies often focus on single-dimensional optimization, failing to achieve deep integration of multi-dimensional features, and still have limitations in query optimization and result ranking. Especially when facing complex government service order query scenarios, how to balance recall and precision, and how to achieve adaptive optimization and ranking of query results, remain pressing technical challenges that need to be addressed.

[0006] Furthermore, existing technologies often lack a systematic integration strategy when dealing with the unique multi-dimensional characteristics of government work orders (such as classification, administrative region, and principal responsibility). The query requirements for government work orders typically involve comprehensive consideration of multiple dimensions, including time, location, issue type, and responsible entity, and existing technologies struggle to effectively coordinate and optimize these dimensions. Simultaneously, government scenarios place high demands on the interpretability and traceability of query results, areas where existing technologies also fall short. Summary of the Invention

[0007] The technical problem solved by this invention is to provide an intelligent retrieval method for government work orders that achieves high recall and high precision based on multi-dimensional feature fusion, adaptive query optimization, and intent sampling mining.

[0008] Technical Solution: To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A method for intelligent retrieval of government work orders based on multi-dimensional feature fusion, adaptive query optimization, and intent sampling mining includes:

[0010] A multi-dimensional classification system for government work orders and a vectorized data retrieval system for government work orders;

[0011] Pre-order module: The pre-order module is used to implement word segmentation, explicit extraction, and vectorized query;

[0012] Model intent extraction: The model intent module is used to process the prompts input by the user to determine the query method and extract the initial user intent.

[0013] Dictionary extraction by filtering: After obtaining the user intent output by the model intent extraction module, extract all relevant content that meets the conditions based on the intent;

[0014] Multidimensional SQL query: Generates SQL statements and outputs the IDs of each work order that meet the requirements from the database based on the generated SQL statements;

[0015] ID aggregation: Employs query result ID aggregation and deduplication mechanisms, and improves recall through the union operation of multiple rounds of query results;

[0016] Preliminary screening and linear weighting: Filtering and sorting the data after ID aggregation;

[0017] Secondary filtering and sorting: The data after the initial filtering and classification are further filtered and sorted, and finally the filtered and sorted data pool is stored in a list in order;

[0018] Result sampling: Sampling is performed from the combined results after the above filtering and sorting, and is divided into three segments: front, middle and back.

[0019] Intent mining: Analyze and mine the content of the sampled results to further abstract structured user intent;

[0020] Intent Enhancement Analysis: Utilizing the logic of the intent extraction module, high-frequency location extraction from the previous mining module is added on the basis of intent extraction;

[0021] Intent fusion: Incorporating the content extracted from intent enhancement analysis into the model's intent extraction module;

[0022] Result fusion and cyclic sampling mining: For the merged and extra content, perform a separate process of dictionary extraction, multidimensional SQL query, ID aggregation, preliminary screening and linear weighting, and secondary screening and sorting.

[0023] Results and Analysis: The results and analysis are presented from four dimensions: basic overview, analysis of main characteristics, summary of key issues, and suggestions for improvement.

[0024] Furthermore, the classification criteria of the multi-dimensional classification system include the system's existing lexical filtering criteria, the system lexical filtering iterative optimization criteria based on the large language model, the work order content classification criteria based on the type of request, the work order classification criteria based on the responsible party, the work order classification criteria based on the type of problem, the work order classification criteria based on the location of the incident, and the work order semantic classification criteria based on vector similarity.

[0025] Furthermore, the government work order vectorized data retrieval system includes: vectorizing each of the three-level classification standards based on the problem type of work orders and their descriptions to generate index files; vectorizing the work order titles and work order addresses of the existing, real-time updated work order data in the system to generate index files; and naming each index file according to "year_month" for subsequent vectorized queries.

[0026] Furthermore, in word segmentation, the jieba word segmentation library and the system-defined dictionary, along with the large language model, are used to segment the user's input query request into nouns, adjectives, place names, and common nouns from the user's input query suggestions.

[0027] Furthermore, in the vectorized query, the semantic similarity retrieval module based on the FAISS vector database performs similarity matching between the user query and the pre-built categorized vector library;

[0028] Three-level classification matching of work order address, work order title, and work order issue based on vector similarity retrieval;

[0029] Based on the index file, the system performs a vectorized query on the user-input query prompts, which is categorized into three levels: work order address, work order title, and work order problem classification. The system then retrieves the top-K similar work order titles, work order addresses, and work order problem classifications from the pre-built FAISS vector database. Finally, the system returns the classification with the highest similarity as a supplementary semantic feature to the query intent.

[0030] Furthermore, the steps for extracting the model intent are as follows:

[0031] Step 1: Check if the user has specified a time range;

[0032] Step 2: Check if the user has specified the type of query they want to make;

[0033] Step 3: Extract the type of query the user wants to make based on the judgment results of Step 2;

[0034] Step 4: Extract direct semantic connectors;

[0035] Step 5: Select the strongest semantic connectors that best meet the requirements from the direct semantic connectors extracted in Step 4;

[0036] Step 6: Related phrase extraction is required;

[0037] Step 7: Determine whether the user's query contains or points to a query category;

[0038] Step 8: Determine whether the user's input "only contains the department name";

[0039] Step 9: Categorize the question type indicated by the user's query;

[0040] Step 10: Further filter the question types;

[0041] Step 11: Does it include determining the responsible party?

[0042] Step 12: Determine whether it only includes the determination of the responsible party;

[0043] Step 13: Extract the responsible parties from the user input.

[0044] Furthermore, the preliminary screening and linear weighting include defining indicator functions, counter definitions, semantic scoring formulas, and classification by the number of associated words.

[0045] Furthermore, the specific steps for intent mining are as follows:

[0046] Step 1: The system extracts the topics the user wants to ask, dividing them into core topics and extended topics;

[0047] Step 2: Extract two layers of keywords, divided into primary keywords and secondary keywords;

[0048] Step 3: Extract combined semantic words from the sample;

[0049] Step 4: Entity recognition;

[0050] Step 5: Let the model extract the filtering strategy, which will be used to guide the subsequent clear description of the intent;

[0051] Step 6: Summarize and evaluate the above analysis.

[0052] Furthermore, the intent fusion process includes constructing intent fusion functions, establishing fusion rules, establishing an overall fusion algorithm, and defining fusion objectives. In the result fusion and cyclic sampling mining, for the content that is merged in, a separate process of dictionary extraction, multi-dimensional SQL query, ID aggregation, preliminary screening and linear weighting, and secondary screening and sorting is performed. The output is inserted into the results before sampling, and then the results after insertion are sampled again until no new intents or new data are added and all results are sampled.

[0053] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0054] The intelligent retrieval method for government work orders based on multi-dimensional feature fusion, adaptive query optimization, and intent sampling mining in this invention achieves high recall and high precision in government work order queries, and has important application value in work order processing in government hotlines, customer service, and other fields.

[0055] (1) High recall rate: Through multiple rounds of SQL queries, ID aggregation and deduplication mechanism, it ensures that all relevant work orders are covered, which significantly improves the recall rate;

[0056] (2) High precision: Through multi-layer feature extraction, intelligent filtering and sorting and linear weighted scoring, the most relevant results are displayed first, which significantly improves the precision.

[0057] (3) Strong adaptability: It supports natural language queries, can understand complex query intents, and adapts to the multi-dimensional query needs of government work orders;

[0058] (4) Good scalability: It adopts a modular architecture design, and each functional module can be independently expanded and updated, which facilitates system maintenance and upgrades;

[0059] (5) Excellent user experience: Provides a visual analysis interface and intelligent analysis reports to help users understand the query results in depth. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the invention;

[0061] Figure 2 It is the decision-making logic of a model with a large time range;

[0062] Figure 3 Extraction steps for models with a large time range;

[0063] Figure 4 It is the process of extracting time-related expression types;

[0064] Figure 5 It is a mapping of a specific time period;

[0065] Figure 6 This is a schematic diagram of word extraction from a large model;

[0066] Figure 7 This is a user-defined word segmentation example;

[0067] Figure 8 It represents the mapping relationship between the order_relation, busi_form, and group_data tables;

[0068] Figure 9 It is the process of explicitly extracting elements from the text;

[0069] Figure 10 It is a three-stage sampling logic diagram;

[0070] Figure 11 This is a logic diagram for sample extraction and intent mining;

[0071] Figure 12 This is a schematic diagram of a specific case. Detailed Implementation

[0072] The present invention will be further illustrated below with reference to specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0073] The present invention provides an intelligent retrieval method for government work orders based on multi-dimensional feature fusion, adaptive query optimization, and intent sampling mining.

[0074] We need to first build a multi-dimensional classification system for government work orders and a vectorized data retrieval system for government work orders.

[0075] 1. The classification criteria of the multi-dimensional classification system include the system's existing lexical filtering criteria, the system's iterative optimization criteria for lexical filtering based on a large language model, the work order content classification criteria based on request type, the work order classification criteria based on the responsible party, the work order classification criteria based on the problem type (three-level classification), the work order classification criteria based on the location of the incident, and the work order semantic classification criteria based on vector similarity. Details are as follows:

[0076] The system's existing lexical filtering standards include: filtering low-importance noun lexical elements, including common words such as "work order," "category," "request," and "problem"; filtering low-importance verb lexical elements, including common action words such as "help," "query," "related," and "involved"; and filtering the full names of the administrative classification areas for work orders built into the system. Taking xx city as an example, this includes "Gaocheng Town," "Guanlin Town," "Wanshi Town," "Environmental Protection Science and Technology Industrial Park," "Economic and Technological Development Zone," "Fangqiao Street," "Hubei Town," "Xushe Town," "Yangxiang Town," "Zhoutie Town," "Zhangzhu Town," "Yixing City," "Xinjian Town," and "Heqiao Town." "Qiting Subdistrict", "Yicheng Subdistrict", "Taihua Town", "Xinzhuang Subdistrict", "Xizhu Town", "Dingshu Town"; the filter system has built-in work order administrative classification area abbreviations, taking xx city as an example, including "Gaocheng", "Guanlin", "Wanshi", "Environmental Science Park", "Economic Development Zone", "Fangqiao", "Hubei", "Xushe", "Yangxiang", "Zhoutie", "Zhangzhu", "Yixing", "Xinjian", "Heqiao", "Qiting", "Yicheng", "Taihua", "Xinzhuang", "Xizhu", "Dingshu".

[0077] The system lexical filtering iterative optimization standard based on the large language model is to extract auxiliary user supplementary sentence patterns and content unrelated to intent based on the sampling content of the matching between the user's query prompts and the system query results. This includes common words similar to "work order", "category", "request", and "problem" in the user's query prompts; and common action words such as "help", "query", "related", and "involved" in the user's query prompts.

[0078] The work order content is classified according to the type of request, including inquiries, requests for help, and suggestions.

[0079] The work order classification standard based on the responsible entity includes, but is not limited to, population entities; industry entities; and organization entities.

[0080] The work order classification standard based on the location of the incident includes, but is not limited to, locations in the following categories: residential and living services; commercial and consumer; cultural, sports, entertainment and tourism; medical and health; educational and scientific research; transportation and infrastructure; government and public institutions; industrial and warehousing; rural; religious and funeral; special purpose venues; financial and business; media and communication; cyberspace; social support and public welfare; and temporary and dynamic venues.

[0081] The work order classification standard based on problem type is a three-tiered system covering public safety, urban and rural construction, housing security, and environmental protection.

[0082] Based on the specific explanation of the three-level classification standard for work orders based on problem type, this serves as the semantic classification standard for work orders based on vector similarity.

[0083] 2. Construct a vectorized data retrieval system for government work orders.

[0084] The government work order vectorized data retrieval system includes: vectorizing each of the three-level classification standards based on problem type and their descriptions to generate index files; and vectorizing the work order titles and addresses of existing, real-time updated work order data to generate index files. Each index file is named according to "year_month" for subsequent vectorized queries.

[0085] Example 1

[0086] This embodiment provides a specific case, and the process is as follows: Figure 1 As shown. In short, this system uses prompt words defined by various criteria established through the multi-dimensional classification system described above, and calls a large model to achieve this. It employs a combination of algorithms and a large model. The specific process is as follows:

[0087] User sends prompt words → User word segmentation → Vectorized query based on established standards → Rule-based responsibility subject and category matching → Large model intent extraction → Database query → Result sampling and large model intent mining → Integration with the intent extracted by the above "Large Model Intent Extraction" → Output query results → Call the large model to analyze the results.

[0088] The overall system workflow includes: 1. Preliminary module; 2. Model intent extraction module; 3. Filtering and sorting dictionary extraction; 4. Multidimensional SQL query; 5. ID aggregation; 6. Preliminary filtering and linear weighting; 7. Secondary filtering and sorting; 8. Result sampling; 9. Intent mining; 10. Intent enhancement analysis; 11. Intent fusion; 12. Result fusion and iterative sampling mining; 13. Result output and analysis. The specific details of the overall workflow are as follows:

[0089] Let U be the text (string) input by the user, D be the work order database containing the table set {busi_form, order_relation, group_data}, and F be the system processing function family.

[0090] The overall system flow can be represented as a function composition:

[0091] ;

[0092] in, It is the intent extraction module; It is the query generation and execution module; It is a filtering and sorting module; It is the results analysis and output module.

[0093] I. Precedence Module

[0094] The main implementations of the preprocessor module include word segmentation, explicit extraction, and vectorized query.

[0095] 1. Word segmentation

[0096] By using the jieba word segmentation library and the system's set dictionary and large language model, the system performs word segmentation on the user's input query request, segmenting nouns, adjectives, place names, and common nouns from the user's input query suggestions, such as... Figure 6 The case shown is illustrated.

[0097] For example, lexical analysis tools can be used to segment and tag user queries, extracting four categories of word units: nouns (n), verbs (v), adjectives (a), and place names (ns).

[0098] Define a set of lexical types The segmentation function is as follows:

[0099]

[0100] in, It is a word segmentation function, each word... Assigned a type .

[0101] 2. Explicit extraction

[0102] Extract the tag set from the table order_relation:

[0103] ;

[0104] Define the extraction function:

[0105] ;

[0106] The explicitly matched category is The main body is Type is the type, T is the jurisdiction, and M is the subject.

[0107] 3. Vectorized query.

[0108] The semantic similarity retrieval module based on the FAISS vector database performs similarity matching between user queries and a pre-built categorized vector library.

[0109] Three-level classification matching of work order address, work order title, and work order issue based on vector similarity retrieval;

[0110] Based on the index file constructed in step S1, a vectorized query is performed on the user-input query suggestions, categorized into three levels: work order address, work order title, and work order issue classification. The top-K similar work order titles, work order addresses, and work order issue classifications are retrieved from the pre-built FAISS vector database.

[0111] Return the most similar categorization as a supplementary semantic feature to the query intent.

[0112] Let the vectorization function be... Map the user query text q to a vector v, FAISS index. The similarity query is as follows:

[0113] ;

[0114] Where sim is the cosine similarity, k is the number of returned results, v is a vector element in the vector library V, x is the content of each column of the work order data table, and busi_form is the work order data table.

[0115] The preceding modules are the two processing steps in the above processing chain: user sends prompt words → "segment user words → vectorized query based on classification criteria".

[0116] Segmenting user vocabulary involves separating general nouns, place names (excluding administrative regions input by the system), verbs (excluding words used by the user to complete sentences that are irrelevant to their intent, such as "help" and "search" in "help me search for work orders related to xxx"), and adjectives from the user's input (e.g., "help me search for work orders related to xxx"). This step can be accomplished using a large model. The results are as follows... Figure 7 The image shows an example of user-defined word segmentation.

[0117] After word segmentation is completed, the text needs to be explicitly extracted based on the dictionary's responsible subject and category.

[0118] The database contains two tables. The first table, `relation`, records the subject and category extracted by the model after processing each work order. The second table, `cate_group_data`, records all first-level, second-level, and third-level categories and their subordinate relationships. Both tables are derived from the `busi_form` table.

[0119] The first table, `order_relation`, records the responsible party and problem type classification extracted from each work order by the large model. The main fields in the table structure are `order_relation(order_id, context, type)`. The `order_id` column is the work order ID, `context` is the tag content (the extracted subject or the assigned category), and `type` is the type. `type` has two tags: `T` and `M`. If it's `T`, it indicates the assigned category, meaning this row is tagged with the category of the work order with the current `order_id`. If it's `M`, it indicates the subject, meaning this row is tagged with the subject of the work order with the current `order_id`.

[0120] The second table is the `group_data` table, with key fields including `group_data(id, data_group, data_val, name, type)`. Each record corresponds to a unique category. Focusing on a single record, `id` is the unique code for this category; `data_group` indicates the category's hierarchy (is it a first-level, second-level, or third-level category?); `data_val` indicates the category to which this category belongs (the first category of a first-level category is 01; taking problem type categories as an example, there are 13 categories, with the thirteenth category being 13; the first second-level category under a first-level category is 0101, the second is 0102, and the third is 0103; the first second-level category under a first-level category corresponds to a third-level category, again taking the first category as an example, the first third-level category is 010101, the second is 010102, and so on); `name` is the name of this category. The `type` column indicates its type, similar to `M` and `T` in the first table, used to indicate whether it's a problem type category or a responsibility entity category.

[0121] like Figure 8 The diagram illustrates the mapping relationship between the order_relation, busi_form, and group_data tables.

[0122] With these two tables, each time a user interacts, the system extracts the content of type T and type M from the `order_relation` table, dividing it into two groups: one for type T and one for type M. Then, it iterates through these two groups, checking each entry to see if the user's input explicitly contains one or more data points from either group. If so, it extracts them—a form of "explicit inclusion" mentioned several times in this document. Another type of explicit inclusion is semantically indicated. The determination of whether it's explicitly indicated first involves a preliminary judgment using vector queries (which perform probability comparisons), followed by further semantic judgment and extraction of vector query results by the larger model.

[0123] like Figure 9 The image shows the process of explicitly extracting elements from the text.

[0124] Therefore, after the explicit text extraction described above, the next step is the vector similarity query process, which focuses on describing the structure and function of the `group_data` table and the original `busi_form` table. The table structure of `group_data` has already been explained above, so we won't go into detail here. Below is the `busi_form` table. The main structure of the `busi_form` table is: `busi_form(id, create_date, form_type, content_text, title, case_goal, case_address, complete_dept, busi_type_id, busi_type_child_id, busi_type_grandson_id)`. The `id` field clearly indicates the work order ID; `create_date` specifies the date and time the work order was created; `form_type` indicates the work order type, used to record user requests, with only six categories: "Consultation," "Suggestion," "Help," "Complaint," "Report," and "Other"; `content_text` records the specific content of each work order; `title` extracts a title for the work order content; `case_goal` indicates the purpose of the work order, which needs to be recorded if the work order has a specific purpose, such as requesting a certain department to expedite a certain task; `case_address` is the specific address mentioned in the work order content, which may be just a community name or a specific building and household within a community; `complete_dept` indicates the relevant unit, such as the Public Security Bureau, Housing and Urban-Rural Development Bureau, Market Supervision Administration, or a specific administrative district, such as the Economic Development Zone; the following `busi_type` values ​​represent the three levels of management: Level 1 → Level 2 → Level 3. Each column of this table needs to be vectorized and stored as a vectorized file (which can be encapsulated using FAISS, bin files, or pkl files) for vectorized queries. This generates a path for fast, multi-dimensional information retrieval based on content. Although there is noise, there is always usable content. Even if there is no usable content, it has no impact because the large model will filter it in the model intent extraction module below. See the model intent extraction module for details.

[0125] II. Model Intent Extraction Module

[0126] Let the extracted intent structure be a tuple:

[0127] ;

[0128] in, It is a set of demand types;

[0129] It is a set of words that directly relate to semantics;

[0130] It is a set of strong semantic conjunctions;

[0131] It is a set of words (or phrases) that must be associated (each element is a word or phrase);

[0132] It is a binary judgment (corresponding to steps 7, 8, 11, and 12);

[0133] It is a set of reasoning paths, where each element is a triple (first level, second level, third level);

[0134] It is the set of categories filtered after vector query;

[0135] It is a set of words representing the responsible parties.

[0136] The extraction process can be viewed as a series of functions:

[0137] .

[0138] The model intent extraction module is the first part of the multi-layer feature extraction mechanism for user query intent. This first part can be divided into two components: query method determination and user intent extraction. For ease of description, we will first provide a linear description of the mechanism's process, and then provide a more detailed description with illustrations in the following sections.

[0139] User intent extraction is the core of this system. Query methods are divided into three categories: subject-based query, category-based query, and subject-plus-category-based query. Category-based query mainly refers to the aforementioned classification standards for responsible entities and problem types.

[0140] Extract multi-dimensional query intent elements, including time range elements, work order type elements, geographical region elements, related word set (strong semantic related words + direct semantic related words), core semantic related word extraction, etc.

[0141] 1) Time range element extraction: Use a large model to identify and extract the time range covered in the query suggestions entered by the user; extract the time interval involved in the query and support the conversion of natural language time expression.

[0142] Time-range feature extraction is achieved through a large model. The decision logic for time-range extraction in the large model is as follows: Figure 2 As shown, the process includes user input, extraction of relevant words, classification processing, merging of time ranges, and output in a standard format. The classification processing of time is specifically divided into absolute time (direct conversion), relative time (equivalent to calculation based on the current date), and special time periods (applying mapping rules).

[0143] For a given time range, the large model extraction steps are as follows: Figure 3 As shown, this includes user input prompts, identifying time keywords, determining time types (including special time periods, relative time ranges, and absolute time ranges), and extracting specific intervals. For example... Figure 4 The diagram shows the process of extracting time expression types. Absolute time ranges are further divided into specific years, specific dates, and fixed time periods; relative time ranges are further divided into those based on current time, those based on historical time, and those based on future time; and special time periods are further divided into holiday periods, seasonal periods, and business-specific periods.

[0144] like Figure 5 The diagram shows a mapping for specific time periods, including the specific time period, the mapped time period, and the intent mapping. For example, the mapping of special time periods such as National Day, New Year's Day, and Dragon Boat Festival to corresponding content. The mapping process is as follows: the corresponding special time period is mapped to a specific time period, and the specific time period is then extracted based on the user's intent.

[0145] 2) Work order type element extraction: explicit work order type specification check and type inference: use a large model to identify and extract the work order types covered in the query suggestions entered by the user.

[0146] Based on the work order content classification standard according to the type of request, a large model is used to identify whether the query prompts entered by the user explicitly specify the work order type; based on the explicit identification results, the business type to which the user's intent belongs is inferred from the prompts entered by the user, including consultation, complaint, report, suggestion, request for help and others.

[0147] 3) Related Word Set: Extract the core noun keywords for the query (not just nouns; could this be changed to "semantically related related words," as it sometimes includes verbs and adjectives) and their weight classifications, distinguishing between strong semantic related words and direct semantic related words. Specifically, this includes:

[0148] We use a large model to extract direct semantic related words from the query suggestions entered by the user; (the system's implementation method is to put the direct words first and the direct words last, but this order is not very important for the sake of our description)

[0149] Use a large model to extract strong semantic related words from the query suggestions entered by the user;

[0150] For example, direct semantic association words include specific livelihood-related issues, specific entity objects (villages, communities, abandoned townships), and geographical area elements (addresses) for identification and extraction. Based on a pre-set administrative division dictionary, it matches city, district / county, and township-level geographical entities; time (summer and winter vacations, etc.); and specific address categories.

[0151] 4) Extraction of core semantic related words (groups): Extract core semantic related words (groups) from the user's input query. These related words are the exact matching words that must be included in the search.

[0152] The model intent extraction module focuses on the user-system interaction process. Traditional search engines involve users sending query requests to the system, such as "Help me find the work order for xxx". The standards established by the multi-dimensional classification system for government work orders categorize and classify content in this format sent by users.

[0153] The first part, described below, outlines thirteen steps.

[0154] Intent extraction based solely on "user input":

[0155] After completing the preliminary stage (preliminary module), the model will process the prompts entered by the user, generally in the format "Help me find work orders related to xxx", processing only this one sentence. However, the system does not require the format to be "Help me find work orders related to xxx", and the model will process the user's input text according to the following steps.

[0156] Step 1: Check if the user has specified a time range.

[0157] For example: "Please help me find the work orders related to xxxx this year."

[0158] Step 2: Check if the user has specified the type of query they want to make.

[0159] For example: - "Help me find work orders related to xxx for consultation and complaint types" - ["Y"]

[0160] - "Help me find the work orders related to xxx" - ["N"]

[0161] Step 3: Extract the type of query the user wants to make based on the results of Step 2.

[0162] If the second step determines "yes" (i.e., ["Y"]), then the user-specified request type is extracted. For example, as shown in the case in the second step, "consultation" and "report" will be extracted. If the second step determines "no" (i.e., ["N"]), then all request types are packed into "(consultation, help, suggestion, complaint, report, other)" to avoid omissions.

[0163] Step 4: Extract direct semantic links, including business keywords, villages / communities, abandoned townships (former names), residential areas, hospitals, schools, scenic spots, public institutions, banks, commercial establishments, companies, holidays and special periods, specific household addresses, current phenomena, and public demands. For example:

[0164] - "May Day Tourism Peak" → ["May Day", "Tourism", "Peak", "Tourists", "Travel"].

[0165] - "Business Environment" → ["Business", "Environment", "Enterprises", "Services", "Approvals"].

[0166] Step 5: From the direct semantic connectives extracted in Step 4, select the strongest semantic connectives that best meet the requirements. Each strong semantic connective must be strongly related to the question, directly covering or addressing the question's intent; each strong connective must be unambiguous when interpreted individually and its meaning must be clear within the relevant context; each strong connective must be independent of each other, with no semantic overlap, covering different aspects of the question; each strong connective must be semantically rich, capable of independent information retrieval, directly addressing the question's answer; multiple strong connectives can be combined to enhance the query. For example:

[0167] Question: "Please help me check all noise-related requests for help and complaints since 2024."

[0168] Input: ["noise", "noise", "disturbing residents", "construction", "life"]

[0169] Output: ["noise", "noise", "loud", "noisy"].

[0170] Step 6: Extraction of Core Semantic Connectives (Groups) (hereinafter referred to as core connectives): Extract core semantic connectives from the user's input query. These connectives are exact matching words that must be included in the search. (Because this step is the core of the entire system's filtering, it involves adding and piecing together the query results bit by bit. After the semantic mining module described later is added, the system's understanding of this intent feature becomes more intelligent and comprehensive. You can refer to the implementation of the subsequent mining module for further understanding, such as:)

[0171] - User input: Query work orders related to road damage. Must include the semantic subject: road surface.

[0172] Extracted output: [["Road surface", "damaged"], ["Road", "damaged"], ["Road surface", "broken"], ["Road surface", "bad"], ["Road surface", "broken"], ["Road surface", "bad"], ["Road surface", "potholes"], ["Road surface", "pits"], ["Road surface", "collapse"], ['Road', 'collapse'], ...]

[0173] - User input: Querying work orders for out-of-town medical insurance this year

[0174] The semantic subject must be included: cross-regional medical insurance

[0175] Extracted output: [["Out-of-town", "Medical Insurance"], ["Out-of-town", "Medical Insurance"], ["Out-of-town", "Medical Treatment"], ["Out-of-town", "Medical Treatment"], ...]

[0176] (The next steps, seven, eight, nine, and ten, are about classifying and categorizing the problems.)

[0177] Step 7: Determine if the user's query contains or points to a query category. This step also allows the large model to make its own judgment. Write prompt words so that the large model can understand the intermediate classification method based on the example, and output ["Y"] or ["N"] as in step 2. An example of passing it to the model is as follows:

[0178] Input: "Stray dog ​​related work order" → ["N"] (Only stray dogs are specified, no problem type is specified)

[0179] Input: "Request for assistance regarding highway construction issues" → ["Y"] (Construction issues are present)

[0180] Step 8: Step 7 determines whether the user's input includes the category / region; step 8 determines whether the user's input "only contains the category / region." This judgment is also crucial, determining whether the user is querying a route for a specific category / region. Again, this is left to the model to make the determination. An example of passing this to the model is as follows:

[0181] - Input: "Stray dog ​​related work order" → ["N"] (Only stray dogs with a specific target, no problem type)

[0182] - Input: "Public sewer related work order" → ["Y"] (A clear category name "public sewer" has been defined, but no specific actual object has been identified)

[0183] (Step 7 is an auxiliary judgment for step 8. If only one of the two modules exists, the judgment will be inaccurate. However, if both modules exist together, practice has proven that the model's self-retrieval ability can accurately judge both of them. This is also a form of organic combination.)

[0184] Step 9: Infer the category of the question type the user query is pointing to. Here, regardless of the judgments made in the first two steps—whether a category is included or only a specific category is specified for the query—the larger model helps analyze which question type is being queried. For example, input the following example into the model, allowing the model to think for itself and ensuring its initiative. Input the detailed category table from the model (the three-level categories are explained, indicating which category belongs to this question type) and specific examples, allowing the model to match them automatically. An example of inputting into the model is as follows:

[0185] Query: "Please help me check the business environment work order"

[0186] Output: [['Comprehensive Economic Sector', 'Comprehensive Approval', 'Other Administrative Approvals'], ['Comprehensive Economic Sector', 'Vertical-Managed Industries', 'Financial Supervision'], ['Comprehensive Economic Sector', 'Government Regulation', 'Investment Promotion']]

[0187] Query: "Please help me check the work orders for drainage system issues during the flood season."

[0188] Output: [['Urban and Rural Construction', 'Emergency Repair', 'Other Emergency Repair'], ['Agriculture, Forestry, Water Resources and Soil', 'Water Conservancy', 'Flood Control and Drought Relief'], ...]

[0189] The output is in Python list format, which facilitates subsequent processing. Taking the work order "Please help me find the work order for flood season drainage system issues" as an example, the output shows two problem type classification paths, each path being "Level 1 Classification → Level 2 Classification → Level 3 Classification". In this example, the model infers two classification paths because sometimes after a user calls 12345 to explain their needs, the operator may misjudge the situation, not knowing which type of problem the user described. The user might randomly select a category. Therefore, this invention aims to prevent this from happening in the query system, allowing the model to infer the classification automatically.

[0190] Step 10: Further filter the problem type: To prevent confusion about the problem category, based on the vectorized files described in the previous steps, vector similarity retrieval is sometimes accurate and sometimes inaccurate. Therefore, the model filters again to determine which content is directly or indirectly related, achieving a more precise judgment. The model learns and filters based on the input examples and makes its own judgment. An example of input to the model is as follows: (Content is the original work order content; input is the vectorized retrieval result; output is the model's filtered result):

[0191] User input: "Help me find work orders related to stray dogs"

[0192] Output must match: [

[0194] "Dog Management: Inquiries, reports, and complaints regarding dog licensing, unlicensed dog ownership, walking dogs off-leash, stray dog ​​management, and large dog bites / injuries."

[0195] "Assistance to Homeless People: Matters Concerning the Assistance of Homeless and Begging Individuals"

[0196] Animal carcass removal: Reports and complaints regarding the failure to promptly remove animal carcasses from public places.

[0197] Animal disease prevention: Inquiries and reports regarding animal disease outbreaks, etc.

[0198] "Veterinary Management: This includes inquiries, reports, complaints, and suggestions related to veterinarians and pet hospitals." ]

[0200] (Steps 11, 12, and 13 involve processing the responsible party. Besides the rule-based matching mentioned earlier, these steps form the basis for the "responsible party query" method. The logic for this judgment is similar to that for determining the categorization, first checking if the user input contains the relevant information.)

[0201] Step 11: Determine if a responsible party is included. By calling the large model, the model decides whether the user's input contains a responsible party; if so, it outputs "Y"; otherwise, it outputs "N". Similarly, input examples are given to the large model to allow it to learn the classification method. An example is shown below:

[0202] Enter: "Please help me find work orders related to the Zisha teapot industry in XX town this year."

[0203] Output: ["Y"]

[0204] Enter: "Please help me find all work orders related to logistics and transportation issues in the Economic and Technological Development Zone this year."

[0205] Output: ["N"]

[0206] Step 12: Determine whether only the responsible party is included. This is the basis for deciding whether to proceed with the subject-based routing. This step also outputs either "Y" or "N". If the result is "Y", the subject extracted in step 13 below will be added to the core related word group for content-based judgment in the subsequent intent mining stage. The judgment method is also to input examples into the large model, allowing the model to learn the analysis method.

[0207] Input: "Help me find elevator-related work orders" → Output: ["Y"]

[0208] Input: "Help me find work orders related to elevator malfunctions" → Output: ["N"]

[0209] Step 13: Extract the responsible party (core object) from the user input. Whether the result of this extraction step is adopted is determined by the two steps mentioned above. The model may independently infer this part of the content for output. Whether or not inference is made is affected by the two steps mentioned above, but the result is independent of the two steps mentioned above. Sometimes the results of the two steps above will affect the inference in this step. An example of input to the model is as follows:

[0210] Enter: "Please help me find work orders related to a burst sewer pipe."

[0211] Output: ["Sewer", "Sewage Pipe", "Pipe", "Drainage Pipe", "Manhole Cover", "Main Pipe", "Sewage Pipe", "Drainage Network", "Sewage Pipe", ...].

[0212] III. Dictionary Extraction

[0213] After completing the thirteen steps of intent extraction described above, the initial extraction of user intent has been achieved. Having obtained this structured user intent, various related contents that meet the criteria can be extracted in conjunction with the intent. The method already implemented in the system uses the following:

[0214] Based on our understanding of query methods, we categorize queries by dimensions: time frame, type of request, semantic related words, type of issue, and type of responsible party. These can be further divided into two types: one based on existing standard matching; and the other based on semantic extraction, such as the aforementioned direct related words, strong related words, and core related phrases, which are abstracted from the semantic meaning according to requirements, requiring not only literal interpretation but also associative thinking. Therefore, we conducted the following queries from these two perspectives.

[0215] - Query all work orders for each secondary category of the intended extraction path: specify the time range, work order type, and secondary category name.

[0216] - Query all work orders for each path and level-three classification for the intended extraction: specify the time range, work order type, and level-three classification name.

[0217] - Query all possible third-level management work orders for which you intend to extract: specify the time range, work order type, and third-level management name.

[0218] - Query work orders containing all core related terms extracted by intent: specify time range, work order type, and related terms.

[0219] -Query work orders containing all direct semantic conjunctions and strong semantic conjunctions for intent extraction: specify time range, work order type, and work orders with both types of semantic conjunctions.

[0220] The above five SQL queries are executed in parallel using a thread pool. The collected data forms five pools filled with work order IDs, serving as the basis for the subsequent linear weighted filtering and sorting. In addition to preparing these filtering materials, a complete SQL query for the intended purpose also needs to be prepared, as detailed in the multidimensional SQL query section below.

[0221] IV. Multidimensional SQL Queries.

[0222] Multidimensional SQL queries are a crucial preliminary step in the process of getting user intent to search results, and include two parts: data input and SQL generation.

[0223] 1. Data Input

[0224] The data is queried using the following inputs: a time range list (including start and end times), a work order request type list (request types determined by the above intent), administrative region, direct semantic related words, strong semantic related words, core semantic related phrases, a list of core responsible subject words explicitly contained in literal text based on dictionary matching, a list of categorization words explicitly contained in literal text based on dictionary matching, a list of categorization paths inferred by the large model, a list of possible categorizations filtered by the large model, the large model's judgment result on whether the user specified a request type, the large model's judgment result on whether the user input contained the categorization of the problem type, and the large model's judgment result on whether the user input only contained the categorization of the problem type.

[0225] 2. SQL generation

[0226] Based on the multi-dimensional features extracted from the large model, multi-round SQL query statements are dynamically generated.

[0227] Based on structured query intent representation, SQL query statements are dynamically generated in rounds 1-n according to the complexity of feature combination. (n is the number of judgment nodes extracted from the intent as routes, with the extracted specific intent content as a factor, combined with the number of intent elements, and arranged in a fixed or dynamic manner.)

[0228] Numerical values ​​(time range, type of claim, administrative region, conjunctions, core semantic conjunctions, jurisdiction, subject; where time range, type of claim, and administrative region are fixed, while conjunctions, jurisdiction, subject, and core semantic conjunctions are dynamic). Includes (but is not exhaustive):

[0229] Basic query generation: Generate basic queries based on time range, work order type, administrative region, set of related words, and set of core semantic related words;

[0230] Extended query generation: When a set of subject terms exists, generate an extended query that includes subject term matches;

[0231] Category query generation: When category exists, generate a query based on multi-level category.

[0232] Semantic similarity query generation: Generate semantic similarity queries based on similarity suggestions obtained from vector retrieval.

[0233] Each round of query statements is constructed using parameterized SQL, supporting dynamic condition combinations and secure execution.

[0234] The SQL generation process is divided into two parts: fixed and variable. The fixed part includes time range conditions, administrative region conditions, and work order request type conditions. The variable part includes responsible entity conditions, jurisdiction conditions, and related keyword conditions. The query process is as follows: based on the time, administrative region, and request type (the result explicitly determines whether to use all or user-specified options based on the input request type), multiple SQL queries are generated as follows: SQL queries containing only the responsible entity (e.g., explicitly matching the responsible entity), SQL queries containing only the jurisdiction, SQL queries containing both the responsible entity and the jurisdiction, SQL queries containing related keywords, SQL queries containing core related keywords, and SQL queries containing the responsible entity + jurisdiction + related keywords + core related keywords. Each query generates an independent SQL query, retrieving only the work order ID.

[0235] Parallel query generation: Based on intent I, generate n types of SQL queries:

[0236]

[0237] in, This indicates the generation strategy.

[0238] Each type of query returns a set of work order IDs:

[0239]

[0240] Execute is the execution function that generates SQL and submits it to the database.

[0241] VI. ID Aggregation

[0242] A query result ID aggregation and deduplication mechanism is adopted, and the recall rate is improved by performing a union operation on the query results from multiple rounds.

[0243] Multiple rounds of SQL queries are executed in parallel. After obtaining the query results from each round, the unique identifier (id field) of each work order is extracted to form an ID set. A union operation is performed on the ID sets from the multiple rounds of queries to merge all related work order IDs. This step ultimately outputs a deduplicated work order ID pool. Based on the merged work order ID pool, a final query statement is generated to obtain complete work order data, ensuring high recall.

[0244] After obtaining the work order IDs from these parts, they are all placed into a pool (concatenated into a tuple covering all intents), ensuring that each ID is unique. Then, an SQL query is used to retrieve all work order data corresponding to the IDs in this pool at once, which will be used as the data to be filtered.

[0245] This concludes the description of the multidimensional SQL query module. After all the content is retrieved, it is filtered and weighted according to the screening and sorting factors.

[0246]

[0247] Here, IDPool represents the ID pool, which means the union of the sets of IDs. Represents a set of IDs.

[0248] Final candidate work order ID pool:

[0249] .

[0250] VI. Preliminary Screening and Linear Weighting

[0251] The initial screening and linear weighting process includes defining indicator functions, counter definitions, semantic scoring formulas, and classification by the number of related words.

[0252] 1. Define indicator functions: For work order d and word w, define indicator functions:

[0253] ;

[0254] 2. Define counters: These include direct conjunction counts, strong conjunction counts, general noun lexical counts, place name noun lexical counts, and adjective lexical counts. Specific representations are as follows:

[0255] Directly related word count: ;

[0256] Strongly related words count: ;

[0257] General noun lexical count: ;

[0258] Place name noun word count: ;

[0259] Adjective lexical count: .

[0260] 3. Semantic scoring formula:

[0261] The semantic score includes the score for gold words, the score for silver words, the score for strong semantic connectors, the score for subject words, the score for categorized words, the score for noun lexical elements, the score for verb lexical elements, the score for adjective lexical elements, and the score for direct semantic connectors.

[0262] Define i as the index of the query results, r i For the current query results, the function (ω,r) represents whether word ω appears in the content of result r (1 if it appears, 0 otherwise), then the key score is... The calculation formula is expressed as follows:

[0263]

[0264] Among them, a 1~a5 is a first-class weight, b1~b4 are second-class weights, ρ1~ρ4 are the intercept vertices, added based on the existence result, a 1~ a5, b1~b4, and ρ1~ρ4 satisfy the following conditions:

[0265] ; ;

[0266] ; ; .

[0267] 4. Categorize by the number of related words

[0268] set up Then work order d is assigned to container:

[0269]

[0270] in, 'm' is the container, and 'm' represents the quantity of the conjunction.

[0271] The filtering and classification process involves further parsing of the intent (and also performing a second check on the intent), which involves checking each work order retrieved according to the ID after the above ID aggregation. The function takes the result pool retrieved from the ID pool, the user intent extracted by the model's intent extraction module, the complete administrative region name (such as Economic and Technological Development Zone, xx town, etc.), the abbreviation of the administrative region (Economic Development Zone, xx), and the word segments segmented from the user input (the word segmentation results from the previous module) as input, and performs filtering and sorting.

[0272] Step 1: Check the results pool to ensure that each result is unique and there are no duplicates.

[0273] Step 2: Define containers (lists). Containers include: silver word container, work order container with five or more related words, work order container with five related words, work order container with four related words, work order container with three related words, work order container with two related words, work order container with one related word, work order container with no related words, and impurity container without subject words (used for filtering when extracting explicitly contained subject words and the intent is to determine subject only).

[0274] (Here we begin iterating through each work order in the query result pool)

[0275] Step 3: Set a base score for the weighted scoring and define the word class counter and value.

[0276] [Basic Score]: Score container: 0 points (empty upon initialization); Basic score for adjective morphemes; Basic score for general conjunctions: 1 point; Basic score for place name morphemes: 5 points; Basic score for common noun morphemes: 7 points.

[0277] [Counters] (work order content includes counts), including gold word counters; silver word counters; responsible subject word counters; category word counters; direct semantic link word counters; strong semantic link word counters; adjective word counters; place name word counters; and general noun word counters.

[0278] Step 4: Ensure that the semantic association words extracted from user input do not contain administrative region names.

[0279] The implementation involves iterating through each related word (direct related word, strong related word, core related word phrase) to ensure that it is not in the full list of administrative regions or the list of abbreviated administrative regions.

[0280] Step 5: Check the length of the list of strongly related words in the intent.

[0281] If the list length is not 0, and if the length is 1, then the first word is designated as a "silver word." If the length is not 1, then the first and second words are added to the silver word container. The number of words in the container is the value of the silver word counter. (Duplicate words are removed to ensure each word is unique.)

[0282] Query method determination: Subject query check (step six)

[0283] Step 6: Routing of Main Query Method

[0284] If it is not a single-category query (the result of the single-category judgment is N), and the responsible subject is explicitly included, or the intent to include only the subject is checked as Y, the query method will be determined to be a subject query.

[0285] The system will merge the explicitly extracted responsible parties and the responsible parties extracted from the intent (including alternative names for the responsible parties, such as "stray dog" or "stray hound") into a "subject pool," which is a list. Then, it will iterate through this list, checking if it contains any of the three word groups (general nouns, place names, and verbs). If it does, it will be added to the key word counter. If the work order content does not contain the key word, the work order will be added to a container of impurities that does not contain the subject word, awaiting final removal from the work order pool.

[0286] Step 7: Counting other parts of speech

[0287] If a literal match is found to be a relevant term, the relevant term counter will be incremented.

[0288] Iterate through the list of direct semantic conjunctions. If a direct semantic conjunction is matched in the content, the number of direct semantic conjunctions will be incremented.

[0289] Iterate through the list of strong semantic connectors. If a strong semantic connector is matched in the content, the number of strong semantic connectors will be accumulated.

[0290] Iterate through the list of general noun morphemes. If a general noun morpheme is matched in the content, the count of the general noun morpheme will be incremented.

[0291] Iterate through the list of place name noun morphemes. If a place name noun morpheme is matched in the content, the count of the place name noun morpheme will be incremented.

[0292] Iterate through the list of adjective morphemes, and increment the count of adjective morphemes if an adjective morpheme is matched in the content.

[0293] Step 8: Calculate the semantic content score according to word class. For example:

[0294] .

[0295] After counting is complete, a new "semantic score" attribute is added to the corresponding work order in the work order pool (each work order in the work order pool has all its attributes recorded, which is a Python dictionary). The value of this attribute is the value calculated by the above formula.

[0296] Step 9: Add the work order to the corresponding number of containers according to the number of direct semantic conjunctions. For example, if the direct semantic conjunction counter reaches 6, a work order container with more than five conjunctions will be added. If it reaches 5, a work order container with 5 conjunctions will be added.

[0297] (The traversal is now complete)

[0298] Step 10: If step 6 determines that the query is a subject query, remove the members (excluding work orders corresponding to the subject) from the subject query impurity list from each container.

[0299] Step 11: Return the container containing 5 or more, 5, 4... a total of 7 related words to the system.

[0300] VII. Secondary Filtering and Sorting (Route Rules)

[0301] Define the routing function:

[0302] ;

[0303] For work order d, define the retention conditions:

[0304] If it is a cluster query: d must satisfy... It also contains core related keywords.

[0305] If it is a main query: d must contain the main term and have a score in the top 20%.

[0306] For mixed queries: d must contain at least one related term or belong to a category set.

[0307] Formalization:

[0308] ;

[0309] Here, Keep is a route filtering function that returns a boolean value indicating whether to retain the work order.

[0310] After the initial screening and classification, the system needs to further filter the data in the seven pools. The system will perform seven rounds of sorting on the data, with each round taking the user's intent, the result list (a specific bowl), and various dictionaries extracted from the screening dictionary extraction module as input. The screening process is as follows:

[0311] The system will first sort the work orders in each pool in descending order according to the weighted scores mentioned above.

[0312] (The following process involves iterating through each bowl of results, checking the attributes of each work order, starting with the id and content_text.)

[0313] Centralized query method routing

[0314] First, the filtering module checks whether the judgment of "only local query" in the user intent extracted by the model is Y. If it is Y, it will be judged as query method routing.

[0315] If the answer is Y, indicating a category query, the model will check if it belongs to the third-level category the user is looking for in the path inferred by the model. It will then check all IDs under the found third-level category to see if it belongs. If it does, it will be retained. If it doesn't belong, it will further check whether it's in the second-level category pool or the potential category pool. If it's in the potential category pool, it will check if the work order content contains core semantic related phrases. If it does, it will be retained; otherwise, it will be added to the list to be filtered. If it's in the second-level category pool, it will similarly check if it contains core semantic related phrases; if it does, it will be retained; otherwise, it will be added to the list to be filtered.

[0316] If the value is N, it first checks if the work order content contains core semantic linking words. If it does, it retains the work order; otherwise, it checks if there are other semantic linking words, if it's in the third-level categorization list, or if it's in the possible categorization list. It first checks if there are semantic linking words or noun phrases; if so, it checks if the work order is in the first 1 / 5 of the list. If so, it retains the work order; otherwise, it discards it. If the work order is in the third-level or possible categorization ID pool, it checks if it contains core semantic linking words; if so, it retains the work order; otherwise, it discards it.

[0317] (Every work order in every bowl goes through this process)

[0318] Finally, the seven data pools after filtering are put into a list in order.

[0319] VIII. Result Sampling

[0320] The system samples from the assembled result after the above filtering, dividing the sampling into three segments: beginning, middle, and end. The model's context window limit is fixed, so the first consideration is the length of the sampled content and whether the model can handle it. The content received by the model consists of: system prompt word length and user prompt word length. To ensure system stability (token count and character count are inconsistent), redundancy is added. Current available length = total context window limit - user prompt word length - redundancy. After calculating the available window limit, the assembled result needs to be sampled, dividing the window limit into three sliding windows for the beginning, middle, and end. Figure 10 As shown.

[0321] Set up a list of filtered work orders Sort by score in descending order. Define a length function.

[0322] The number of characters in the work order content.

[0323] Given length constraints:

[0324] Total Limitation The system prompts the length. User-suggested length Redundancy Initial Intent Length .

[0325] Available length:

[0326] .

[0327] Will Divide the sample into three equal parts (leaving a buffer zone) to obtain the constraints for each segment. .

[0328] Sampling function Return subsequence:

[0329] ;

[0330] Among them, Head, Middle, and Tail are work orders selected from the corresponding paragraphs until the length limit is reached. Indicates the head sampling window limit. Indicates the sampling window limit in the middle. This indicates the sampling window limit at the tail end.

[0331] 8. Intent Discovery

[0332] After sampling, the model performs intent mining (or information extraction) on the sampled data. The basic process of intent mining is as follows: Figure 11 As shown, the system analyzes and mines the sampled results by calling a large model, and further abstracts the structured user intent.

[0333] Let the sampling work order set be... The mining function is:

[0334]

[0335] in: =(Core topic, Extended topics); =(Primary keywords, Secondary keywords)

[0336] =(direct semantic phrases, indirect semantic phrases); =(High-frequency locations, stakeholders, typical cases); = (Including conditions, Excluding conditions, Validation checks).

[0337] The specific steps are as follows:

[0338] Step 1: The system extracts the topics the user wants to ask, dividing them into core topics and extended topics.

[0339] Step 2: Extract two layers of keywords, divided into primary keywords and secondary keywords.

[0340] The main keywords are the highest priority keywords, which directly describe the core issue, including words describing the problem phenomenon, words describing the impact and consequences, and words describing the solutions.

[0341] Words describing the problem phenomena include the phenomena themselves, such as "fault" or "congestion"; the characteristics of the problem, such as "frequent" or "sudden"; and the specific form of the problem, such as "dropping" or "outage".

[0342] Words describing impacts and consequences include impacts on personal safety, such as "safety hazard"; impacts on life / work, such as "difficulty in traveling"; economic or property losses, such as "property loss"; and psychological / experiential impacts, such as "fright" or "poor experience".

[0343] The terms describing the solutions include the solutions expected by users, such as "repair" and "unclog"; the measures that relevant departments should take, such as "investigate" and "rectify"; and management / supervisory measures, such as "strengthen inspections".

[0344] Secondary keywords are used to supplement information, providing context and background. These include terms related to time or space, relevant parties, and documents / procedures.

[0345] The extracted time- and space-related terms include time dimensions (when it occurs, frequency, duration) and spatial dimensions (location, regional characteristics).

[0346] Relevant parties include responsible parties such as the property management company and the market supervision administration; affected parties such as homeowners and tourists; and third parties such as the maintenance company.

[0347] Terms or entities related to certificates and procedures include official documents such as "business license"; approval processes such as "application" and "approval"; and standards such as "national standards".

[0348] Step 3: Extract combined semantic words from the sample.

[0349] Combinatorial semantic words are divided into direct semantic word groups and indirect semantic word groups. Direct patterns are semantic word groups that the model directly extracts from the samples and can directly match the description of the problem. For example, if I input "help me query work orders related to typhoon days", it extracts the following: [["typhoon","cause","damage"], ["heavy rain","leads to","flooding"], ["typhoon day","crops","damaged"], ["strong wind","damaged","facilities"], ["flood control and drought relief","repair","waterway"]].

[0350] Indirect semantic phrases are used to represent more complex semantic relationships. The model extracts the cause and effect of the problem by understanding the contextual logic of each work order in the sample, capturing scattered but semantically related keywords within the content. For example, inputting "Help me find work orders related to typhoon weather," it extracts: [["Too hot weather", "Fields", "Crops", "Drought"], ["Waterway is damaged", "Hope", "Repair as soon as possible"], ["Farmland", "Damaged", "No one is handling it"], ["Poor drainage", "Typhoon", "Impact", "Travel"]]

[0351] Step 4: Entity Recognition. This step focuses on key aspects of the sample. Sometimes, when a user queries for a work order categorized by a specific issue type, they actually mean work orders within their intended scope, rather than the system's rigid issue type classification. After all, users don't care about the system's design. Therefore, to address this, besides various subqueries and overall queries—which are subject to various limitations and standard matching—these steps target semantically related words. This step further extracts entities as "keywords." For example, if a location frequently experiences fires in a certain neighborhood or flooding in another area, specific entity matching is required. This focuses on three dimensions:

[0352] - High-frequency locations: Extracted results are sorted by frequency of occurrence. Each location includes: name, number of occurrences, and a description of the problem characteristics. This helps identify "high-incidence areas" in the sampled content. The actual results, including those outside the sample, may differ, but the entities extracted from the sampled data can still be used as a reference. For example, if I input "help me query work orders related to typhoon days", it extracts the following: [{"name":"xx town","count":"1 time (the only clearly defined location in the sample)","description":"the specific town in the sample, possibly a high-frequency area affected by typhoons"},{"name":"xxx village","count":"1 time (the only clearly defined village name in the sample)","description":"located in Heqiao Town, a concentrated area of ​​farmland and waterway problems"}).

[0353] The model identifies stakeholders and extracts frequently mentioned responsible parties or relevant organizations from the collected work order content samples. The model then sorts these by importance or frequency based on the prompts. For example, if I input "help me find work orders related to typhoons," it will extract the following: ["citizens", "villagers", "water resources department", "emergency management bureau", "village committee"].

[0354] Typical Case: Extracting summaries of frequently occurring or typical work orders from the work order sample that closely match user intent. This involves describing recurring specific problem scenarios in the sample; these are not abstract problems, but concrete examples used to understand the "typical manifestations" of the problems collected in the sample. For example, if I input "Help me find work orders related to typhoons," it might extract: ["Damage to crops caused by damage to farmland irrigation facilities after a typhoon", "Damaged irrigation ditches not repaired in time affecting agricultural production", "Heavy rainfall causing poor drainage and waterlogging in farmland"].

[0355] Step 5: Let the model extract filtering strategies to guide subsequent clear descriptions of intent. The above steps all involve the model extracting key information for later retrieval, whether words or phrases—these are hard data. However, further extraction is needed for some cases, especially edge cases. For example, consider the tricky scenario: "A citizen said they were walking on a bridge and felt very cold from the strong wind." If the model misinterpreted "strong wind" as a typhoon in previous steps, what do we do? This is where intent mining becomes crucial. We not only need to provide the model with retrieval materials, but also require the model to extract the intent using natural language and more structured matching words, ensuring that subsequent queries are both based on evidence and have a compelling basis. This module's extraction can be termed filtering strategies from the following dimensions: First, a structured inclusion strategy, also described using words:

[0356] (Using the example of "help me find work orders related to typhoon days")

[0357] - Includes conditions: divided into three levels, the following is only a reference standard for the model to refer to;

[0358] Words that must be included (for reference, not the only criterion): direct semantic words that appear frequently in the sample;

[0359] [Typhoon, Heavy Rain, Strong Winds, Disasters, Flood Control and Drought Relief]

[0360] Keywords that should be included: auxiliary keywords to improve relevance; ["repair", "waterway", "farmland", "drainage", "crops", "flooding", "emergency"]

[0361] Semantic matching words: One or more words extracted from the samples that can summarize the intent contained in typical, high-frequency work orders and other samples. [["Typhoon", "Damage"], ["Rainstorm", "Flooding"], ["Waterway", "Repair"], ["Crops", "Drought"], ["Flood Control and Drought Relief", "Water Conservancy"]]

[0362] That concludes the extraction of the included strategies. The next step is to exclude the included strategies.

[0363] Exclusion criteria: After extracting the features of the work orders that can be inferred to be retained, it is necessary to further extract the features of the content to be filtered out. However, "noise" represents distortion. This part of the concept cannot be extracted in a structured and universal way like the intent features mentioned above. At this time, it is more about natural language description. Here, we also start from three dimensions: (taking "help me find work orders related to typhoon days" as an example).

[0364] The characteristics of noise work orders are described in natural language: ["Only mentions hot weather but no keywords such as typhoon or rainstorm", "Only reflects high temperature and drought but no extreme weather association", "Only inquires about weather forecasts but no actual disaster situation"].

[0365] Noise topics were extracted and described in natural language: ["Traffic congestion during typhoons (no facility damage)", "Power outage inquiries during typhoons (no repair requests)", "Complaints about tour cancellations after typhoons"].

[0366] Borderline cases prone to misjudgment are summarized in natural language: ["Irrigation demand solely due to high temperatures (no typhoon association), "Routine water conservancy maintenance did not mention severe weather"].

[0367] After verifying and extracting the aforementioned noise features, further verification steps need to be designed based on the outlined profile, described in natural language; ["Check if the content_text contains core keywords such as typhoon, rainstorm, disaster", "Confirm if it involves disaster-affected objects such as water conservancy facilities, farmland, and drainage", "Verify if it falls within the time range of July 1 to October 31, 2025", "Confirm that the form_type is a request for help, complaint, or report rather than an inquiry"]

[0368] At this point, the noise has been basically outlined, and the boundaries of the intent are relatively clear.

[0369] Step 6: Summarize and evaluate the above analysis from four dimensions;

[0370] Scope definition: Allow the model to clearly define the boundary range of the sampled query in natural language, including what is included and what is not included, and the limitations of time, space, and theme dimensions; for example: "involving emergency requests such as damage to water conservancy facilities, farmland disasters, and poor drainage caused by extreme weather such as typhoons, rainstorms, and strong winds, focusing on the theme of flood control and drought relief."

[0371] Common types: The model extracts regular characteristics, typical manifestations, and high-frequency combination relationships of problems from the samples, such as: "After a typhoon, damage to farmland irrigation facilities (such as irrigation ditches) leads to crop damage, and citizens seek repair assistance; drainage system failure causes the risk of farmland flooding or drought."

[0372] Core issues: The model extracts fundamental problems, underlying causes, or systemic defects from the work orders collected in the sample, and the pain points that need to be addressed first. For example: "Insufficient disaster resistance capacity of rural water conservancy facilities, delayed emergency repair response, and lack of protection against agricultural losses."

[0373] Data quality assessment: This section evaluates the limitations of the model on the sample data, identifies data issues requiring attention, and outlines considerations for data selection. For example: "The sample contains only one record; more data is needed to validate high-frequency locations and semantic patterns; some expressions, such as 'it's too hot,' need to be considered in context to determine if they are indirectly affected by the typhoon, avoiding misclassification as independent high-temperature issues."

[0374] At this point, the sample intent mining module has completed its mining process.

[0375] 10. Intent Enhancement Analysis

[0376] This section uses the same logic and prompts as the model's intent extraction module, but extracts fewer prompts. It removes all Y / N judgments from the intent extraction module, except for the final intent judgment regarding the responsible party. Then, it adds high-frequency location extraction from the previous mining module. Note that it doesn't extract any administrative districts, only specific neighborhoods, hospitals, schools, etc., such as Experimental Middle School, XX University, XX Garden, XX Village, etc. To save context and for robustness, it doesn't extract specific building or resident information. Other information, such as administrative affiliation, is extracted as usual. The output format is the same as the model's intent extraction module, with the addition of the high-frequency location module and the removal of some judgment module content. The input here is the intent-mined information above, described, with the final judgment strategy placed first.

[0377] XI. Intentional Fusion

[0378] The content extracted by the above enhanced analysis module is added to the model intent extraction module. Here, the high-frequency words after the above enhanced analysis are added to the "strong semantic association words". If the judgment of both subjects is Y, each subject is added to the core semantic association word group.

[0379] Compared to the initial intent extraction, if there are additional classification paths or information in the extracted related terms and related modules, these classification paths and information will be incorporated into the intent extracted by the model's intent module. The specific implementation process is as follows:

[0380] 1. Basic Definition

[0381] set up:

[0382] The intent results (dictionary) of the basic factor extraction module;

[0383] The intent result (dictionary) of the intent enhancement analysis module is the part with high-frequency locations after the basic intent is extracted and mined by the sampling analysis model;

[0384] To enhance the results of the analysis module;

[0385] key_words is the variable name for words that are directly semantically related;

[0386] strong_key_words is the variable name for strong semantic keywords;

[0387] essential_keywords is the variable name for core semantic related word groups;

[0388] Categories are a list of categorization paths inferred by the model;

[0389] possible_categorizations_info is a description of the three-level categorizations selected by the model after vector similarity retrieval, along with their descriptions.

[0390] 2. Construct the intent fusion function

[0391] Define the intent fusion function This function merges the intent fusion results into the base intent:

[0392]

[0393] in, This is the updated result of the underlying intent.

[0394] 3. Merging Rules (Step-by-Step Mapping)

[0395] 3.1 Fusion of Direct Semantic Connectives

[0396]

[0397] That is, from Add key_words to Remove duplicates from key_words

[0398] 3.2. High-frequency location fusion

[0399] For each location ,

[0400] ;

[0401] .

[0402] 3.3. Fusion of strong semantic conjunctions

[0403] .

[0404] 3.4. Integration of core semantic conjunctions

[0405] .

[0406] 3.5. Categorization and Integration

[0407] .

[0408] 3.6. Potential Information Integration by Category

[0409] .

[0410] 3.7. Explicitly Included Object Merging

[0411] Conditional judgment:

[0412]

[0413] If satisfied, then for each object :

[0414] .

[0415] 4. Overall Fusion Algorithm

[0416]

[0417] in: It is the intent extraction function; It's a system prompt message; It is a user suggestion word; It is the set of model parameters.

[0418] User prompts The construction method is as follows:

[0419]

[0420] in: Original user input; The intention is to enhance the analysis results; This is the result of basic intent extraction.

[0421] 5. Integration Objectives

[0422] The objective function for intent fusion is:

[0423]

[0424] in, It is a measure of the semantic similarity between intent and user input.

[0425] This process ensures: completeness (intent features extracted from multiple modules are integrated); deduplication (avoiding duplicates); consistency (maintaining semantic consistency and not deviating from the original intent); and enhancement (enhancing query capabilities through fusion).

[0426] Conditional convergence: Let the result of the t-th iteration be... The convergence condition of the algorithm is:

[0427]

[0428] At this point, the algorithm terminates and outputs the result. .

[0429] This mathematical expression formally describes the logic, constraints, and objectives of the intention fusion process.

[0430] 12. Results Fusion and Cyclic Sampling Mining

[0431] For the content that is merged in, perform multiple rounds of dictionary extraction, multidimensional SQL query, ID aggregation, preliminary screening and linear weighting, and secondary screening and sorting separately; then sample the results before insertion and sample the results after insertion until no new data is added and all results are sampled.

[0432] Let the result of round t be ,but:

[0433]

[0434] until = (convergence).

[0435] XIII. Results Output and Analysis

[0436] The results and analysis are presented from four dimensions: basic overview, analysis of main features, summary of key issues, and work suggestions. Specifically, the model is based on the content of the work order.

[0437] Example 2

[0438] This embodiment provides an intelligent retrieval system for government work orders based on multi-dimensional feature fusion, adaptive query optimization, and intent sampling mining, to implement the method of Embodiment 1, including:

[0439] Pre-order module: The pre-order module is used to implement word segmentation, explicit extraction, and vectorized query;

[0440] Model intent extraction module: processes the prompts input by the user to determine the query method and extract the initial user intent;

[0441] Filtering and dictionary extraction module: After obtaining the user intent output by the model intent extraction module, it extracts all relevant content that meets the conditions based on the intent;

[0442] Multidimensional SQL query module: Generates SQL statements and outputs the IDs of each work order that meet the requirements from the database based on the generated SQL statements;

[0443] ID aggregation module: Employs query result ID aggregation and deduplication mechanism, and improves recall through union operation of multiple rounds of query results;

[0444] Preliminary filtering and linear weighting module: Filters and sorts the data after ID aggregation;

[0445] Secondary filtering and sorting module: Further filtering and sorting of the data after the initial filtering and classification, and finally storing the filtered and sorted data pool into a list in order;

[0446] Result sampling module: Samples are taken from the combined results after the above filtering and sorting, and the sampling is divided into three segments: front, middle and back.

[0447] Intent mining module: Analyzes and mines the sampled results to further abstract structured user intent;

[0448] Intent Enhancement Analysis Module: Utilizing the logic of the intent extraction module, it adds the extraction of high-frequency locations from the previous mining module, building upon the intent extraction process.

[0449] Intent Fusion Module: Adds the content extracted from intent enhancement analysis to the model's intent extraction module;

[0450] The results fusion and cyclic sampling mining module: For the merged and extra content, it performs a separate process of dictionary extraction, multidimensional SQL query, ID aggregation, preliminary screening and linear weighting, and secondary screening and sorting.

[0451] The results output and analysis module provides results and analysis from four dimensions: basic overview, analysis of main features, summary of key issues, and work suggestions.

[0452] Example 3

[0453] An electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method of Embodiment 1 and the system of Embodiment 2.

[0454] Example 4

[0455] A computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the method steps of Embodiment 1 and the system of Embodiment 2.

[0456] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A government affair work order intelligent retrieval method based on multi-dimensional feature fusion, adaptive query optimization and intention sampling mining, characterized in that, include: A multi-dimensional classification system for government work orders and a vectorized data retrieval system for government work orders; Pre-order module: The pre-order module is used to implement word segmentation, explicit extraction, and vectorized query; Model intent extraction: The model intent module is used to process the prompts input by the user to determine the query method and extract the initial user intent. Dictionary extraction by filtering: After obtaining the user intent output by the model intent extraction module, extract all relevant content that meets the conditions based on the intent; Multidimensional SQL query: Generates SQL statements and outputs the IDs of each work order that meet the requirements from the database based on the generated SQL statements; ID aggregation: Employs query result ID aggregation and deduplication mechanisms, and improves recall through the union operation of multiple rounds of query results; Preliminary screening and linear weighting: Filtering and sorting the data after ID aggregation; Secondary filtering and sorting: The data after the initial filtering and classification are further filtered and sorted, and finally the filtered and sorted data pool is stored in a list in order; Result sampling: Sampling is performed from the combined results after the above filtering and sorting, and is divided into three segments: front, middle and back. Intent mining: Analyze and mine the content of the sampled results to further abstract structured user intent; Intent Enhancement Analysis: Utilizing the logic of the intent extraction module, high-frequency location extraction from the previous mining module is added on the basis of intent extraction; Intent fusion: Incorporating the content extracted from intent enhancement analysis into the model's intent extraction module; Result fusion and cyclic sampling mining: For the merged and extra content, perform a separate process of dictionary extraction, multidimensional SQL query, ID aggregation, preliminary screening and linear weighting, and secondary screening and sorting. Results and Analysis: The results and analysis are presented from four dimensions: basic overview, analysis of main characteristics, summary of key issues, and suggestions for improvement. 2.The government affair work order intelligent retrieval method based on multi-dimension feature fusion, adaptive query optimization and intent sampling mining according to claim 1, characterized in that: The classification criteria of the multi-dimensional classification system include the system's existing lexical filtering criteria, the system lexical filtering iterative optimization criteria based on the large language model, the work order content classification criteria based on the type of request, the work order classification criteria based on the responsible party, the work order classification criteria based on the type of problem, the work order classification criteria based on the location of the incident, and the work order semantic classification criteria based on vector similarity. 3.The government affair work order intelligent retrieval method based on multi-dimension feature fusion, adaptive query optimization and intent sampling mining according to claim 1, characterized in that: The government work order vectorized data retrieval system includes: vectorizing each of the three-level classification standards based on the problem type of work orders and their descriptions to generate index files; vectorizing the work order titles and work order addresses of the existing, real-time updated work order data in the system to generate index files; and naming each index file according to "year_month" for subsequent vectorized queries. 4.The government affair work order intelligent retrieval method based on multi-dimension feature fusion, adaptive query optimization and intent sampling mining according to claim 1, characterized in that: In word segmentation, the jieba word segmentation library and the system-defined dictionary, along with the large language model, are used to segment the user's input query request into nouns, adjectives, place names, and common nouns from the query suggestions entered by the user. 5.The government affair work order intelligent retrieval method based on multi-dimension feature fusion, adaptive query optimization and intent sampling mining according to claim 1, characterized in that: In vectorized queries, the semantic similarity retrieval module based on the FAISS vector database performs similarity matching between user queries and pre-built categorized vector libraries. Three-level classification matching of work order address, work order title, and work order issue based on vector similarity retrieval; Based on the index file, the system performs a vectorized query on the user-input query prompts, which is categorized into three levels: work order address, work order title, and work order problem classification. The system then retrieves the top-K similar work order titles, work order addresses, and work order problem classifications from the pre-built FAISS vector database. Finally, the system returns the classification with the highest similarity as a supplementary semantic feature to the query intent. 6.The government affair work order intelligent retrieval method based on multi-dimension feature fusion, adaptive query optimization and intent sampling mining according to claim 1, characterized in that: The steps for extracting model intent are as follows: Step 1: Check if the user has specified a time range; Step 2: Check if the user has specified the type of query they want to make; Step 3: Extract the type of query the user wants to make based on the judgment results of Step 2; Step 4: Extract direct semantic connectors; Step 5: Select the strongest semantic connectors that best meet the requirements from the direct semantic connectors extracted in Step 4; Step 6: Related phrase extraction is required; Step 7: Determine whether the user's query contains or points to a query category; Step 8: Determine whether the user's input "only contains the department name"; Step 9: Categorize the question type indicated by the user's query; Step 10: Further filter the question types; Step 11: Does it include determining the responsible party? Step 12: Determine whether it only includes the determination of the responsible party; Step 13: Extract the responsible parties from the user input. 7.The government affair work order intelligent retrieval method based on multi-dimension feature fusion, adaptive query optimization and intent sampling mining according to claim 1, characterized in that: The initial screening and linear weighting process includes defining indicator functions, counter definitions, semantic scoring formulas, and classification by the number of related words. 8.The government affair work order intelligent retrieval method based on multi-dimension feature fusion, adaptive query optimization and intent sampling mining according to claim 1, characterized in that: The specific steps for intent mining are as follows: Step 1: The system extracts the topics the user wants to ask, dividing them into core topics and extended topics; Step 2: Extract two layers of keywords, divided into primary keywords and secondary keywords; Step 3: Extract combined semantic words from the sample; Step 4: Entity recognition; Step 5: Let the model extract the filtering strategy, which will be used to guide the subsequent clear description of the intent; Step 6: Summarize and evaluate the above analysis.

9. The intelligent retrieval method for government work orders based on multi-dimensional feature fusion, adaptive query optimization, and intent sampling mining as described in claim 1, characterized in that: In intent fusion, the process includes constructing intent fusion functions, establishing fusion rules, establishing an overall fusion algorithm, and fusion objectives. In result fusion and cyclic sampling mining, for the content that is merged in multiple ways, a separate process of dictionary extraction, multidimensional SQL query, ID aggregation, preliminary screening and linear weighting, and secondary screening and sorting is performed. The output is inserted into the result before sampling, and then the result after insertion is sampled again until no new data is added and all results have been sampled.

10. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in claims 1 to 9.

Images