Government affair fusion media text generation method based on artificial intelligence

Abstract

The application relates to the technical field of text generation, and specifically discloses a government affair fusion media text generation method based on artificial intelligence, which comprises the following steps: firstly, identifying sentences containing ambiguous expressions from an original policy text; secondly, extracting corresponding sentence structure and constructing a sentence template with variable placeholders; thirdly, combining the administrative discretion benchmark file corresponding to the policy or the general discretion rule to obtain the variable value interval of the ambiguous expression in actual execution and the applicable conditions corresponding to each value; fourthly, discretizing the value interval into multiple discrete value points and converting them into natural language descriptions; fifthly, backfilling the natural language descriptions into the sentence template to generate multiple candidate text sentences; and finally, splicing the candidate text sentences according to the position sequence in the original policy text and fixed sentences to form multiple candidate policy interpretation texts, so that the generated results can adapt to different execution scenarios on the basis of retaining the discretion boundary and conditional semantics of the original text.

Description

Technical Field

[0001] This invention relates to the field of text generation technology, specifically to a method for generating government converged media text based on artificial intelligence. Background Technology

[0002] Vague expressions such as "handled at one's discretion," "appropriately relaxed," and "priority given to those who meet the conditions" are widely found in government policy documents. These expressions belong to normative administrative language and are intended to leave discretionary space for grassroots implementing agencies to adapt to complex and ever-changing real-world scenarios. Existing government document generation methods based on large language models follow the ambiguity resolution logic of general natural language processing when parsing these vague expressions, directly converting them into definitive statements.

[0003] Specifically, when the original policy text states "subsidies may be provided preferentially to groups facing special difficulties," existing methods directly generate an interpretation text stating "all groups facing special difficulties can receive subsidies." This generation method semantically removes the conditional and discretionary nature implied by "may be provided" and "preferentially" in the original text, compressing non-mandatory discretionary space into an absolute execution conclusion. Because the generated interpretation text does not contain obvious errors in terms of sentence fluency and logical integrity, this problem of semantic boundary tampering is difficult to identify during routine review. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based method for generating government converged media texts, thereby solving the aforementioned technical problems.

[0005] The objective of this invention can be achieved through the following technical solutions: The method for generating government converged media text based on artificial intelligence includes the following steps: S1. Input the original policy text to be processed, identify all the positions of sentences containing ambiguous expressions in the original policy text, and associate and store each position of sentence containing ambiguous expressions with the chapter number of the sentence in the original policy text. S2. Call the artificial intelligence model to parse each statement containing fuzzy expressions, extract the fuzzy expression words in the statement and the complete sentence structure of the sentence containing the fuzzy expression words, replace the fuzzy expression words with preset variable placeholders, and generate a sentence template with variable placeholders. S3. Obtain the set of administrative discretion parameters corresponding to the original policy text. The set of administrative discretion parameters includes the variable value range of each ambiguous term in the actual implementation process and the description text of the applicable conditions corresponding to each value in the variable value range. S4. Discretize the variable value range into multiple discrete value points, substitute each discrete value point into a variable placeholder, and generate multiple candidate text statements. Each candidate text statement corresponds to a discrete value point. S5. Concatenate multiple candidate text statements with other statements in the original policy text, excluding those containing ambiguous expressions, to generate multiple candidate interpretation texts. Each candidate interpretation text contains a complete set of policy interpretation content. S6. Output multiple candidate interpretation texts, each candidate interpretation text is accompanied by a description text of the applicable conditions associated with the corresponding discrete value points.

[0006] Preferably, the process of identifying the location of a statement containing an ambiguous expression is as follows: A fuzzy expression vocabulary library is pre-built. Each word entry in the fuzzy expression vocabulary library is accompanied by a semantic category label, which includes three types: condition relaxation, condition tightening, and discretion. Each sentence in the original policy text is traversed, and each sentence is matched with the word entries in the fuzzy expression vocabulary library. When any word entry exists in a sentence, the sentence is marked as containing a fuzzy expression. For the marked sentences, the start character position and end character position of the sentence in the original policy text are extracted. The start character position, end character position, the chapter number of the sentence, and the name of the matched word entry are combined into a sentence record unit. All sentence record units are arranged in ascending order of start character position to generate a sentence position list.

[0007] Preferably, the process of generating sentence templates is as follows: For statements containing fuzzy expressions, the syntactic analysis component in the artificial intelligence model determines the start and end indices of the matched fuzzy expression words in the statement. Using the start index as the split point, the statement is divided into prefix and suffix text segments. Then, using the end index as the split point, the suffix text segment is divided into infix and final suffix text segments. The infix text segment is the fuzzy expression word itself. The prefix text segment, variable placeholders, and final suffix text segment are concatenated in sequence to generate a sentence template with variable placeholders. At the same time, the original word text of the fuzzy expression word replaced by the variable placeholder in the statement is recorded, and a mapping relationship is established between the original word text and the variable placeholder and stored in the mapping table.

[0008] Preferably, the process of obtaining the set of administrative discretion parameters is as follows: Based on the issuing agency name and document number of the original policy text, the administrative discretion benchmark documents associated with the original policy text are retrieved from the government affairs database. Discretionary rule entries corresponding to each vague expression term are extracted from these administrative discretion benchmark documents. Each discretionary rule entry includes the name of the vague expression term, a description of the applicable scenario, the lower and upper limits of the variable value range, and a description of the applicable conditions for each value segment within the variable value range. All extracted discretionary rule entries are aggregated to generate an administrative discretionary parameter set. When no associated administrative discretion benchmark document exists in the government affairs database, a preset general discretionary rule library is invoked. Based on the semantic category label of the vague expression term, the corresponding default discretionary rule entry is matched from the general discretionary rule library, and this default discretionary rule entry is used as a component of the administrative discretionary parameter set.

[0009] Preferably, the process of obtaining multiple discrete value points is as follows: For each discretionary rule entry in the administrative discretionary parameter set, obtain the lower limit and upper limit values ​​of the variable value range in the discretionary rule entry. Based on the semantic category label of the fuzzy expression term corresponding to the discretionary rule entry, select the corresponding discretization step size value from the discretization strategy library. Starting from the lower limit value, generate a numerical sequence by increasing the discretization step size value until the maximum value in the numerical sequence does not exceed the upper limit value. Take each value in the numerical sequence as a discrete value point. When there is no discretization step size value corresponding to the semantic category label in the discretization strategy library, take the lower limit value and upper limit value of the variable value range as two discrete value points respectively.

[0010] Preferably, the process of generating candidate text statements is as follows: An artificial intelligence model is invoked to convert discrete value points into natural language descriptive text. The AI ​​model selects the corresponding conversion template from a pre-set text conversion template library based on the semantic category label of the fuzzy expression words containing the discrete value points. The conversion template includes format rules for converting numerical values ​​to text and quantifier suffix text. The natural language descriptive text is filled into the variable placeholder positions to generate the filled text statement. The original vocabulary text of the fuzzy expression words replaced by the variable placeholders is extracted from the mapping table. The consistency between the filled text statement and the context of the original vocabulary text is checked. When the subject of the filled text statement is inconsistent with the subject of the original vocabulary text, the subject text is extracted from the original vocabulary text and inserted into the beginning of the filled text statement to generate candidate text statements.

[0011] Preferably, the process of concatenating the statements is as follows: Obtain all statements in the original policy text that are not marked as containing ambiguous expressions. Arrange all statements that are not marked as containing ambiguous expressions in the order of their appearance in the original policy text to generate a fixed text statement sequence. For each candidate text statement, obtain the original position number of the statement containing ambiguous expressions in the original policy text corresponding to the candidate text statement. Insert the candidate text statement into the fixed text statement sequence at the position corresponding to the original position number, and replace the statement containing ambiguous expressions corresponding to the original position number to generate a complete candidate text statement sequence. Concatenate each statement in the candidate text statement sequence in order, inserting newline characters between statements to generate candidate interpretation text.

[0012] Preferably, the process of outputting candidate interpretation text is as follows: Multiple candidate interpretation texts are sorted according to the numerical values ​​of their corresponding discrete points to generate a candidate interpretation text list. A unique identifier is generated for each candidate interpretation text in the list. The unique identifier is associated with the storage path of the candidate interpretation text and written into an index file. A corresponding applicable condition description text is attached to each candidate interpretation text in the index file. The applicable condition description text is written directly into the entry position of the index file corresponding to the unique identifier in plain text form. The index file and all candidate interpretation texts are packaged and output as a converged media data package.

[0013] The beneficial effects of this invention compared to the prior art are as follows: This invention identifies ambiguous statements in the original policy text, extracts the sentence structure of sentences containing ambiguous terms, and replaces them with variable placeholders, thus preserving the original text's sentence framework and discretionary semantic structure. The variable value range in the administrative discretion parameter set is discretized into multiple discrete value points. Each discrete value point is substituted with a variable placeholder to generate candidate text statements, allowing the same ambiguous statement to generate multiple semantically mutually exclusive candidate interpretation texts, each conforming to the discretionary range boundaries. These candidate text statements are then concatenated with fixed statements in the original policy text to output multiple candidate interpretation texts. Each candidate interpretation text includes a corresponding description of applicable conditions, enabling operators in different implementation scenarios to select the matching interpretation version based on actual conditions. The entire solution fully preserves the discretionary range information corresponding to the ambiguous statements during the generation process, avoiding the semantic tampering problem of converging discretionary statements into absolute conclusions. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart illustrating the AI-based method for generating government converged media text according to the present invention.

[0016] Figure 2This is a flowchart illustrating the process of generating candidate text statements according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figures 1-2 As shown, this invention is an artificial intelligence-based method for generating government affairs converged media text, comprising the following steps: Step 1: Identify statements containing ambiguous expressions; Specifically, after the original policy text to be processed is read into memory, text normalization is first performed, replacing full-width spaces, consecutive line breaks, and duplicate characters with a single delimiter, while retaining the original character sequences of title numbers, item numbers, and item numbers. Then, sentences are segmented according to periods, semicolons, colons, and line breaks to generate a sequence of sentences with sequential numbers. During the segmentation process, the starting and ending character positions of each sentence in the original policy text are accumulated. Subsequently, a pre-built fuzzy expression vocabulary is called to match each sentence one by one. The fuzzy expression vocabulary is stored in the form of a vocabulary record table. Each record contains at least a vocabulary name, vocabulary length, semantic category label, and matching priority. The semantic category label is limited to condition relaxation, condition tightening, and discretionary categories. For example, "appropriately relaxed" and "preferential support" are marked as condition relaxation, "strict control" and "strict restriction" are marked as condition tightening, and "handled at one's discretion" and "can be adjusted" are marked as discretionary.

[0019] For each individual statement, a string scan is performed from left to right. When the character pointed to by the scan pointer matches the first character of any term, a substring of the same length is extracted and compared. If the substrings are completely identical, it is recorded as a hit term. When multiple hit terms exist at the same position, the term with the longer word length and higher matching priority is selected as the final result. The chapter number to which the hit statement belongs is extracted. The chapter number is obtained by identifying the first, second, or third level number in the nearest parent title line, and is written into the same statement record unit along with the statement's start character position, end character position, and hit term name. After all statements have been scanned, all statement record units are arranged in ascending order of their start character positions to generate a statement position list for direct use in subsequent sentence template construction steps.

[0020] For example, the original policy text is set as the Temporary Assistance Measures for Urban and Rural Residents in Difficulty in a certain city. Article 4 of Chapter 2 states that eligible applicants may receive a discretionary increase in temporary assistance when encountering sudden difficulties. Article 7 of Chapter 3 states that the application period may be appropriately extended for families with special difficulties. After reading the policy text, the original title structure is retained, and Chapter 1 (General Provisions), Chapter 2 (Assistance Standards), and Chapter 3 (Application Procedures) are identified as chapter titles. The main text is then segmented using periods, semicolons, and line breaks. The resulting statements are: Statement 1: To standardize the distribution of temporary assistance; Statement 2: Eligible applicants may receive a discretionary increase in temporary assistance when encountering sudden difficulties; Statement 3: Applicants should submit their applications within 30 days of the occurrence of difficulties; Statement 4: The application period may be appropriately extended for families with special difficulties. A fuzzy vocabulary database pre-stores entries such as "discretionary increase," "appropriate extension," "priority guarantee," and "strict review," with "discretionary increase" marked as "discretionary discretion" and "appropriate extension" marked as "relaxed conditions." During sentence-by-sentence scanning, the accuracy of the matched phrases is adjusted accordingly in sentence 2, while the accuracy of matched phrases is relaxed appropriately in sentence 4. Position information is then recorded. For example, sentence 2 has a starting character position of 128 and an ending character position of 149 in the original full text, belongs to chapter 2, item 4, and has a matched phrase name adjusted accordingly; sentence 4 has a starting character position of 210 and an ending character position of 226, belongs to chapter 3, item 7, and has a matched phrase name relaxed appropriately. Two sentence recording units are then created. The first recording unit contains 128, 149, chapter 2, item 4, and adjusted accuracy; the second recording unit contains 210, 226, chapter 3, item 7, and relaxed appropriately. These are then sorted by starting character position to obtain a list of sentence positions. Subsequent steps only process these two marked sentences; the remaining sentences are retained as fixed text.

[0021] Step 2: Generate sentence templates; Specifically, after sequentially reading each statement containing ambiguous expressions from the statement position list, the original statement text, the name of the hit term, the starting character position, the ending character position, and the chapter number are combined as a record to be parsed and sent to the artificial intelligence model. The artificial intelligence model first performs word segmentation and syntactic analysis on the statement. The word segmentation result outputs a sequence of word elements in the original character order. Each word element is accompanied by a word element start index, word element end index, and part-of-speech tag. The syntactic analysis result gives the subject-predicate, verb-object, modifier-head, and adverbial-head relationships between each word element, thereby determining the actual boundary of the hit term in the statement.

[0022] When a fuzzy expression consists of two or more consecutive tokens, the smallest starting index of the consecutive tokens is used as the word's starting index, and the largest ending index is used as the word's ending index. When there are two or more fuzzy expressions in the same sentence, they are processed separately according to the ascending order of the starting indices, and an independent variable placeholder is assigned to each fuzzy expression. The variable placeholders are generated using a fixed character format, specifically written as variable1, variable2, variable3, etc. They are not numbered repeatedly within the same sentence. Different sentences can be further expanded by combining the sentence sequence number to the form of sentence5 variable1, to ensure unique location during subsequent backfilling.

[0023] After obtaining the starting index of the vocabulary, the sentence is segmented into prefix and suffix text segments according to this index. The prefix text segment is the continuous text from the first character of the sentence to the character before the starting index of the vocabulary, and the suffix text segment is the continuous text from the starting index of the vocabulary to the end of the sentence. A second segmentation is then performed based on the relative position of the ending index of the vocabulary within the suffix text segment, resulting in infix and final suffix text segments. The infix text segment is the original fuzzy expression vocabulary itself, and the final suffix text segment is the continuous text remaining after the fuzzy expression vocabulary ends. Subsequently, the prefix text segment, variable placeholders, and final suffix text segment are concatenated in their original order to form a sentence template. For example, "For eligible applicants, the subsidy ratio can be appropriately increased" can generate "For eligible applicants, the subsidy ratio can be increased by variable 1".

[0024] To avoid losing the original sentence structure, the subject, predicate, object, limiting elements, and quantity expressions in the sentence are preserved before splicing. The word order, punctuation position, and chapter affiliation information of non-ambiguous expressions are not changed. After the template is generated, the original vocabulary text, variable placeholders, sentence number, vocabulary start index, vocabulary end index, semantic category label, and corresponding sentence template number are written into a mapping table. The mapping table is stored in a row-by-row manner, with each row corresponding to one variable placeholder. This is used to look up the replaced original vocabulary text based on the variable placeholders later, and to maintain the one-to-one correspondence between the template, vocabulary, and sentence position during the discrete value backfilling stage.

[0025] For example, following statement 2 above, the original sentence is: "When eligible applicants encounter sudden difficulties, the temporary relief standard may be increased at the discretion of the authorities." After the statement is fed into the AI ​​model, the word segmentation results are first output, sequentially identified as: eligible, applicant, in, encounter, sudden difficulties, when, may, increase at the discretion of the authorities, temporary relief standard. Syntactic analysis further identifies "applicant" as the core subject, "increase at the discretion of the authorities" as a predicate phrase, and "temporary relief standard" as the object. The model locates the boundary of "increase at the discretion of the authorities" in the word sequence, assuming its starting index is 18 and its ending index is 22. Based on this index, the statement is segmented: the prefix text segment is "When eligible applicants encounter sudden difficulties, may," and the suffix text segment is "increase the temporary relief standard at the discretion of the authorities." Further segmentation is done using the ending index position, resulting in the infix text segment "increase at the discretion of the authorities," and finally, the suffix text segment is "temporary relief standard." The default variable placeholder generation rule stipulates that the first variable in the same sentence is written as Variable 1. Therefore, the prefix text fragment, Variable 1, and the final suffix text fragment are sequentially concatenated to form a sentence template. The content is: "Eligible applicants can receive temporary assistance under Variable 1 when encountering sudden difficulties." A record is written into the mapping table, including the variable placeholder Variable 1, the original vocabulary text (adjusted as needed), the sentence number 2, the vocabulary start index 18, the vocabulary end index 22, the semantic category label (discretionary class), and the template number MB0002. Looking at sentence 4, "The application period can be appropriately relaxed for families with special difficulties," the model identifies the start index for appropriate relaxation as 10 and the end index as 14. After replacing these with Variable 1, the template "The application period can be extended for families with special difficulties under Variable 1" is generated, and another record is written into the mapping table. The result is that the subject, object, limiting conditions, and syntactic framework of the original sentence are all preserved; only vague words are replaced with fillable positions. Subsequent changes only require filling in different expressions into the variable positions.

[0026] Step 3: Obtain the set of administrative discretion parameters; Specifically, after completing the sentence template and mapping table construction, the issuing agency name and document number fields are extracted from the first page, title area, and signature area of ​​the original policy text. The issuing agency name is obtained by identifying the issuing agency's signature line and the agency name preceding the title. The document number is obtained by matching a continuous character segment composed of the agency's abbreviation, year identifier, and sequence number. The combination of the issuing agency name and document number is then written into the search request as the search key. The government database pre-stores a policy document index table and an administrative discretion benchmark document index table. The policy document index table records the policy title, issuing agency name, document number, subject category, and effective status. The administrative discretion benchmark document index table records the benchmark document number, associated policy document number, applicable item name, rule entry position, and version status.

[0027] The system performs a precise match in the administrative discretion benchmark document index table based on the search key. When the document numbers are completely identical, the corresponding administrative discretion benchmark document is located directly. When the document number is missing or there are format differences, a second match is performed based on the issuing agency name and policy subject classification. The document with a normal effective status and the release time closest to the original policy text is selected from the matching results as the associated document.

[0028] After obtaining the relevant documents, the document text is segmented into rule fragments by clause, clause, and item. Then, by scanning each fuzzy term in the mapping table, discretionary rule entries containing the term or its synonym are extracted. Each discretionary rule entry must include at least the term name, applicable scenario description text, lower limit value, upper limit value, and value segment condition text. The applicable scenario description text is defined as continuous text in the rule entry used to limit the applicable object, type of matter, and execution situation, such as the applicant being a low-income family or the illegal act being minor. The variable value range is defined as the adjustable value boundary explicitly given in the discretionary rule entry; the value can be expressed as a proportion, amount, period, or grade code. The applicable condition description text refers to the conditional explanation content bound to the specific discretionary value corresponding to a certain fuzzy term, used to indicate under what factual circumstances the value can be adopted. This text is not a repetition of the value itself, but rather a textual limitation on the applicable object, behavioral circumstances, eligibility status, supporting materials, risk level, or special matters. For example, when the variable range corresponding to a certain vague expression is 10 to 30, the applicable condition description text can be written as: the applicant's family income is lower than the local minimum living standard and they have submitted proof of hardship, or the illegal behavior lasted for a short period and they have proactively rectified it. When generating candidate interpretation texts, each discrete value point must be associated with a corresponding applicable condition description text, enabling operators to determine the usability of the candidate text based on the actual scenario. In other words, the applicable condition description text serves to map the discretionary values ​​to the actual applicable scenarios item by item. For items that only record the grading standards without directly specifying the total range, the minimum grading value is used as the lower limit value, the maximum grading value as the upper limit value, and the numerical segment condition text is established according to the original order of the grading values.

[0029] After completing all word matching, the extracted discretionary rule entries are aggregated according to the mapping relationship between fuzzy expression word names and variable placeholders to generate an administrative discretionary parameter set. When there is no associated administrative discretionary benchmark document in the government database, a preset general discretionary rule library is read. The general discretionary rule library stores default discretionary rule entries with semantic category tags as the main index. The condition relaxation category corresponds to the relaxation range from low to high and the applicable conditions, the condition tightening category corresponds to the tightening range from light to heavy and the applicable conditions, and the discretionary category corresponds to multiple discrete discretionary levels and applicable scenario descriptions. Then, the corresponding default discretionary rule entries are extracted according to the semantic category tags of the fuzzy expression words, and the word name, variable placeholder, and statement number fields are added. After the default discretionary rule entries and the extracted exclusive rule entries adopt the same record structure, they together form the administrative discretionary parameter set.

[0030] For example, taking the aforementioned civil affairs policy as an example, the original policy text title area shows the issuing agency as a certain municipal civil affairs bureau, and the document number as Minfa

[2025] No. 18. After extracting these two fields, the associated administrative discretion benchmark documents are retrieved from the government database. A policy record exists in the policy document index table, titled "Temporary Assistance Measures for Urban and Rural Residents in Difficulty in a Certain City," with the issuing agency being a certain municipal civil affairs bureau and the document number as Minfa

[2025] No. 18. An associated record exists in the administrative discretion benchmark document index table, with benchmark document number Caiji

[2025] No. 07, associated policy document number Minfa

[2025] No. 18, and the applicable matter name being "Temporary Assistance Standard Discretionary Benchmark." After reading the main text of this benchmark document, the articles are segmented. In Article 6, the following content is found: If an applicant encounters general hardship and their family income declines significantly in the short term, the temporary assistance standard can be increased by 10 to 20% above the basic standard; if an applicant encounters a major emergency and provides valid proof of medical expenses, the temporary assistance standard can be increased by 21 to 30% above the basic standard. This allows us to extract the discretionary rule entries corresponding to "discretionary increase." The term is "discretionary increase," the applicable scenario description is "adjustment of temporary assistance standards," the lower limit of the range is 10, and the upper limit is 30. The condition text for the numerical segment is split into two parts: one part (10-20) corresponds to general hardship and a significant short-term decline in family income, and the other part (21-30) corresponds to a major emergency and the provision of valid proof of medical burden. Further searching for the rule entries corresponding to "appropriate relaxation," we find in rule 9 that for families with special difficulties who cannot apply on time due to hospitalization, natural disasters, or the absence of a guardian, the application period can be extended by 5 to 15 days. Here, we can extract the term "appropriate relaxation," the applicable scenario description "application period adjustment," the lower limit of the range (5), the upper limit of the range (15), and the condition text for the numerical segment: 5-10 corresponds to hospitalization or short-term movement restrictions, and 11-15 corresponds to natural disasters or the absence of a guardian leading to the inability to apply on time. Aggregating these two rule entries forms the set of administrative discretionary parameters corresponding to this policy text. If no discretionary benchmark file is found in the database, for example, if a newly issued trial notice from a county has not yet established a dedicated benchmark file, then default rules are read from the general discretionary rule library. For example, for terms like "appropriate relaxation," the library can pre-set default entries as a time-limited relaxation range of 3 to 10 days, where 3 to 5 days corresponds to general procedural obstruction, and 6 to 10 days corresponds to significant objective obstacles, thereby supplementing the parameter set.

[0031] Taking the rule above, which allows for discretionary increases in temporary relief standards, as an example, the discrete value range is 10 to 30. However, each value within this range cannot be used independently of the actual situation; therefore, a corresponding descriptive text of the applicable conditions is required. Assuming that discretization yields five discrete value points: 10, 15, 20, 25, and 30, the descriptive text of the applicable conditions must clearly state under what circumstances each value point can be applied. For example, 10 corresponds to a decrease in the applicant's family income but without affecting basic living standards, and proof of income has been submitted; 15 corresponds to an increase in family expenditure pressure due to short-term unemployment, and the street office has completed a home visit verification; 20 corresponds to an applicant experiencing a sudden illness, with a significant increase in medical expenses and submission of hospitalization invoices; 25 corresponds to an applicant experiencing a major accident causing temporary loss of income for the main breadwinner of the family; and 30 corresponds to an applicant simultaneously experiencing a major accident, continuous treatment, and no other stable source of income for the family. The descriptive text of the applicable conditions here is not a rewriting of the values ​​10, 15, 20, 25, and 30, nor is it simply written as low, medium, and high levels; rather, it maps each value to a specific actual implementation scenario. When generating candidate interpretation texts, if a candidate text states "increase by 25", then the same candidate text will be accompanied by a statement explaining that the applicant's main breadwinner has temporarily lost their source of income due to a major accident. Only after seeing this statement can the staff determine whether the version can be used for the current applicant.

[0032] Step 4: Obtain multiple discrete value points and generate corresponding candidate text statements for each discrete value point; Specifically, after obtaining the set of administrative discretion parameters, the corresponding discretion rule entries are read one by one according to the variable placeholders recorded in the mapping table. For each discretion rule entry, the lower and upper limits of the variable value range are first parsed, and the data type of the value range is identified. The data type is determined based on the measurement expression in the original text of the discretion rule entry, including percentage values, monetary values, time limit values, and grade values. Among them, the percentage values ​​are stored in percentage or decimal form, the monetary values ​​are uniformly converted to the same currency unit, the time limit values ​​are uniformly converted to days or months, and the grade values ​​are represented by a predefined sequential code.

[0033] The discretization strategy library pre-stores step size records according to semantic category labels. Each step size record includes at least a semantic category label, applicable data type, step size value, and rounding method. Condition relaxation categories correspond to step size records that increase from lower to higher values; condition tightening categories correspond to step size records that increase from lighter to heavier values; and discretionary categories correspond to step size records that retain the middle range. After reading the semantic category label of the discretionary rule entry, the library searches for step size records with the same category and data type. For example, if a rule's range is 10 to 30 and the data type is a proportional value, and a step size value of 5 is found, then 10 is used as the starting point, and subsequently 15, 20, 25, and 30 are generated sequentially until the new value exceeds the upper limit. The generated values ​​are then written sequentially into the discrete value point list. For monetary and time-limited values, a unit standardization check is performed after the discrete value points are generated to ensure that all value points under the same rule entry use the same unit. If there is no matching step size record in the discretization strategy library, the lower limit value and the upper limit value are directly written into the discrete value point list so that each discretion rule entry corresponds to at least 2 discrete value points.

[0034] After the discrete value points are generated, they are bound to the applicable condition description text corresponding to the numerical segment containing those values, forming candidate value records. Each candidate value record includes at least a variable placeholder, the discrete value, the numerical unit, a semantic category label, the applicable condition description text, and a sentence template number. Subsequently, an artificial intelligence model is invoked to perform natural language conversion on the discrete value points in the candidate value records. The text conversion template library pre-stores conversion templates corresponding to different semantic categories and data types. Each conversion template includes the numerical input location, numerical writing format, quantifier suffix text, and semantic connectors. For example, a percentage value can be converted to "increase by 20%", "relax by 15%", or "control within 10%"; a time limit value can be converted to "extend by 3 days" or "shorten by 5 days"; and a monetary value can be converted to "subsidy of 2000 yuan" or "reduction of 500 yuan". After reading the semantic category labels, the artificial intelligence model selects the corresponding conversion template, and then converts the discrete value points into natural language description text according to the writing format in the template. If the original vague expression words contain degree adverbs or directional words, the directional semantics are retained, so that the text generated by the relaxed class values ​​and the text generated by the tightened class values ​​are consistent in semantic direction.

[0035] After natural language processing (NLP) is completed, a sentence template with variable placeholders is retrieved based on the sentence template number. The natural language description text is then filled into the positions of the variable placeholders to obtain the filled text statement. After the filled text statement is generated, the original vocabulary text replaced by the variable placeholder and its original sentence are retrieved from the mapping table. Contextual consistency is then checked against the filled text statement. Contextual consistency check compares the correspondence between the filled text statement and the original sentence in terms of subject, predicate direction, and object. The subject is determined by the agent word or topic word in the aforementioned syntactic analysis results. If the subject in the original sentence is the applicant, the competent authority, or the violator, and the subject is missing in the filled text statement due to variable substitution, the subject text is extracted from the original sentence and inserted at the beginning of the filled text statement. If the filled text statement already contains the same subject text as the original sentence, it is retained and not inserted again. For sentences containing qualifiers, time-limited phrases, or conditional clauses, the original text in the template is retained without deleting or modifying non-variable parts, thus generating candidate text statements consistent with the original policy context. For each generated candidate text statement, the candidate text statement, its corresponding discrete value point, the corresponding applicable condition description text, variable placeholders, and sentence template number are written into the candidate statement record table. These records are then used to fill in multiple candidate interpretation texts in the full text according to their original positions.

[0036] For example, we continue using the clause "adjustments to increase the temporary relief standard." The administrative discretion parameter set already contains a lower limit of 10 and an upper limit of 30, with the semantic category label being "discretionary" and the data type being "proportional value." The discretization strategy library pre-sets a step size of 5 for the discretionary class proportional value, so a numerical sequence is generated starting from 10, resulting in 10, 15, 20, 25, and 30. Each value is then bound to the corresponding applicable condition description text, forming 5 candidate value records. Next, we enter the natural language conversion stage. The text conversion template library contains a template for the discretionary class proportional value, with the format rule being "increase the basic standard by a numerical value plus a percentage." Therefore, 10 is converted to an increase of 10% on the basic standard, and 15 is converted to an increase of 15% on the basic standard. These natural language descriptions are then filled into the previously generated sentence template "Eligible applicants encountering sudden difficulties, variable 1: temporary relief standard," resulting in 5 filled text statements. The corresponding content is as follows: Eligible applicants facing sudden difficulties can receive a 10% increase in temporary relief from the basic standard, and eligible applicants facing sudden difficulties can receive a 15% increase in temporary relief from the basic standard. Since directly inputting this directly results in slightly awkward semantics, contextual consistency checks and word order corrections are necessary. The model checks that the subject of the original sentence is "eligible applicants," and since the subject is not lost after inputting, no subject needs to be added. The template is then adjusted based on the object "temporary relief standard," correcting it to "eligible applicants facing sudden difficulties can receive a 10% increase in temporary relief," and "eligible applicants facing sudden difficulties can receive a 15% increase in temporary relief." Next, regarding the clause on appropriately extending the application period, the range is 5 to 15 days. The step size for the extended period value in the discretization strategy library is set to 5, generating three discrete value points: 5, 10, and 15. The text conversion template is written as "application period extension plus value plus day," thus yielding application period extensions of 5 days, 10 days, and 15 days. After filling in the template with the variable 1 for the application period for families with special difficulties, the word order is adjusted to create options such as extending the application period by 5 days, 10 days, and 15 days for families with special difficulties. Each candidate text statement is bound to an applicable condition description; for example, an extension of 5 days corresponds to short-term hospitalization, an extension of 10 days corresponds to continuous hospitalization or significant restriction of movement, and an extension of 15 days corresponds to natural disasters or lack of guardianship that objectively prevents timely application. In this way, S4 does not simply enumerate numbers, but divides the discretionary range into multiple candidate execution versions that can be directly implemented in the text, and prepares for the next step of splicing back into the full text.

[0037] Step 5: Concatenate the statements to generate candidate interpretation text; Specifically, after the candidate text statements are generated, all statement records are read from the statement sequence corresponding to the original policy text. Statements not marked as containing ambiguous expressions are identified based on the statement position list. The statement number, original position number, chapter number, and statement text of these statements are extracted and written into a fixed text statement table. The original position number refers to the sequence number of the statement in the original policy text arranged in order of appearance, such as sentence 1, sentence 2, and sentence 3, used to maintain structural consistency between the candidate interpretation text and the original policy text.

[0038] Then, all records in the fixed text statement table are read from smallest to largest according to their original position numbers to generate a fixed text statement sequence, while preserving the original position vacancies of the marked statements in the sequence. The candidate text statement record table pre-stores the original statement number, variable placeholder number, sentence template number, discrete value point, and candidate text statement body corresponding to each candidate text statement. Therefore, a one-to-one correspondence can be established between the candidate text statements and the marked statements in the original policy text based on the original statement numbers.

[0039] For a given candidate text statement, first read its corresponding original position number, then write the candidate text statement into the empty space at the same position in the fixed text statement sequence. Replace the ambiguous statement corresponding to the original statement number with this candidate text statement, thus obtaining a complete candidate text statement sequence. If there are more than two statements containing ambiguous expressions in the original policy text, then read the candidate text statements corresponding to each position from the candidate text statement record table, and write them one by one into the corresponding position in the fixed text statement sequence according to their original position numbers. This ensures that each replaced position is filled with a candidate text statement that has completed discrete value backfilling, while the unreplaced positions retain the fixed statement content from the original policy text. After all positions are written, output the complete candidate text statement sequence in the order of the original position numbers, then concatenate the statements in the sequence into continuous text, inserting line breaks between adjacent statements, and maintaining the original line break level between chapter titles and the main text, generating a candidate interpretation text. By repeating the same replacement and concatenation process for the remaining candidate text statements of the same vague expression, multiple candidate interpretation texts can be generated. Each candidate interpretation text contains complete policy interpretation content, and except for the replaced positions, the order of the remaining statements, chapter affiliation, and fixed text content remain consistent with the original policy text. The policy interpretation content is a complete text composed of the fixed statement content in the original policy text and the candidate text statements generated for the vague expression. Specifically, the original policy text is first segmented into statements, retaining statements without vague expressions as fixed content. Then, sentence templates are extracted from statements containing vague expressions, and multiple candidate text statements are generated based on the discrete value points in the administrative discretion parameter set. Subsequently, according to the original position number of each statement in the original policy text, the candidate text statements are backfilled into the corresponding vague statement positions, replacing the original vague expression statements with the candidate text statements. These are then re-concatenated with the remaining unreplaced fixed statements in their original order to obtain a complete text. This complete text is called a policy interpretation because the original vague expressions such as "handle at your discretion," "appropriately relax," and "prioritize support" have been replaced with explicit statements with specific discretionary values. At the same time, the original policy text's chapter structure, context, and unambiguous statements have been retained, thus forming a complete interpretation text that can be read and used directly.

[0040] Step Six: Output multiple candidate interpretation texts; Specifically, after generating all candidate interpretation texts, the candidate statement record table corresponding to each candidate interpretation text is read first. The discrete value points associated with each candidate interpretation text, their units, applicable condition descriptions, and the main text of the candidate interpretation text are extracted from this table. The candidate interpretation texts are then written as independent text files to a preset storage directory. The sorting of candidate interpretation texts is based on the discrete value points. Multiple candidate interpretation texts corresponding to the same fuzzy expression are arranged in ascending order of discrete value points. If the discrete value points contain unit information, unit consistency is achieved before comparing the numerical values. When there are more than two fuzzy expressions in the original policy text, the sorting key for the candidate interpretation texts is written sequentially according to the original position order of the discrete value points corresponding to each fuzzy expression. If the preceding values ​​are the same, the following values ​​are compared, thus forming a stable list of candidate interpretation texts.

[0041] After the candidate interpretation text list is generated, a unique identifier is generated for each candidate interpretation text in the list. The unique identifier is generated by concatenating fixed fields in a fixed order and must include at least the original policy text number, the candidate interpretation text sequence number, and the discrete value point sequence number, ensuring that each candidate interpretation text in the same batch of output files has a unique correspondence. Subsequently, an index file is created, stored in a record-by-record structure. Each record corresponds to one candidate interpretation text, and the record content must include at least the unique identifier, the candidate interpretation text storage path, the discrete value point value, the unit of the value, the applicable condition description text, and the original statement number. The storage path is defined as the relative position string of the candidate interpretation text file in the output directory, for example, interpretation text 3 in the text set directory.

[0042] The applicable condition description text is written directly into the record entry corresponding to the unique identifier in plain text form, without image or external link processing. This allows the reader to directly obtain the applicable scenario description corresponding to the candidate interpretation text when parsing the index file. After all candidate interpretation texts are written, the index file and each candidate interpretation text file are grouped according to a unified directory structure. The directory contains the index file, text files, and batch information files. This directory is then encapsulated into a converged media data package. The header of the converged media data package contains the batch number and generation time for subsequent distribution, retrieval, and retrieval. In this way, each candidate interpretation text not only retains the complete text but also allows direct location of the corresponding discrete value point and applicable condition description text in the index file using the unique identifier.

[0043] This invention addresses the problem of how ambiguous expressions in government policy texts are easily erroneously rewritten as definitive conclusions during automatic generation. The focus is not on directly eliminating flexible language from the original text, but on preserving the discretionary boundaries behind this language and transforming them into a controllable, expandable, and refillable generated structure. Expressions like "handle with discretion," "appropriately relax," and "prioritize support" in the original policy text inherently serve to reserve adjustment space for specific implementation scenarios. Directly rewriting these expressions using a general model often compresses conditional expressions into absolute results, resulting in a fluent but distorted version of the original policy intent. To avoid this deviation, the truly discretionary statements in the entire text are separated from fixed statements. Only these key positions are subject to controlled processing, while the rest of the content retains the original text order and semantic framework. This limits the scope of erroneous rewriting to the smallest possible unit. Next, the sentences containing the ambiguous expressions are extracted into sentence templates with variables. Essentially, this transforms the previously uncontrollable natural language rewriting process into a process of replacing local variables, preserving the subject, the object of the action, the limiting conditions, and the context of the chapter, thus avoiding semantic drift caused by rewriting entire sentences from the source. Subsequently, administrative discretion benchmark documents or general discretionary rule bases corresponding to the policy are introduced. This prevents the model from guessing the meaning of ambiguous words based on language habits, instead establishing a correspondence between the ambiguous expressions and the value ranges and applicable conditions in actual implementation. This step connects the normative expressions in the policy text with external discretionary bases, giving the generated content a traceable rule source. Finally, the continuous interval is split into multiple discrete value points to transform the abstract discretionary space into candidate versions that can be generated item by item. Since continuous intervals cannot be directly used for text output, discretization allows each value point to correspond to a clear statement and a description of applicable conditions. The discrete values ​​are then converted into natural language that fits the original sentence context and backfilled into the template, resulting in multiple candidate sentences with different semantic boundaries but all within the original discretionary range. Finally, the full text is backfilled according to the original position, forming multiple complete interpretation texts. The essential principle behind this approach is to break down the original one-off, black-box text generation into several consecutive steps: fuzzy sentence recognition, sentence parameterization, rule mapping, interval discretization, and controlled backfilling. Structural constraints and rule constraints are used to limit the generation results, ensuring that each output text is merely an expansion of the original discretionary space, rather than an unauthorized rewriting of the policy meaning. Therefore, while maintaining the original text framework and discretionary attributes, it can provide corresponding versions of interpretation content for different execution scenarios.

[0044] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A method for generating government converged media text based on artificial intelligence, characterized in that: Includes the following steps: S1. Input the original policy text to be processed, identify all the positions of sentences containing ambiguous expressions in the original policy text, and associate and store each position of sentence containing ambiguous expressions with the chapter number of the sentence in the original policy text. S2. Call the artificial intelligence model to parse each statement containing fuzzy expressions, extract the fuzzy expression words in the statement and the complete sentence structure of the sentence containing the fuzzy expression words, replace the fuzzy expression words with preset variable placeholders, and generate a sentence template with variable placeholders. S3. Obtain the set of administrative discretion parameters corresponding to the original policy text. The set of administrative discretion parameters includes the variable value range of each vague expression in the actual implementation process and the description text of the applicable conditions corresponding to each value in the variable value range. S4. Discretize the variable value range into multiple discrete value points, substitute each discrete value point into a variable placeholder, and generate multiple candidate text statements. Each candidate text statement corresponds to a discrete value point. S5. Concatenate multiple candidate text statements with other statements in the original policy text, excluding those containing ambiguous expressions, to generate multiple candidate interpretation texts. Each candidate interpretation text contains a complete set of policy interpretation content. S6. Output multiple candidate interpretation texts, each candidate interpretation text is accompanied by a description text of the applicable conditions associated with the corresponding discrete value points.

2. The method for generating government converged media text based on artificial intelligence according to claim 1, characterized in that, The process of identifying the location of a statement containing ambiguous expressions is as follows: A fuzzy expression vocabulary library is pre-built. Each word entry in the fuzzy expression vocabulary library is accompanied by a semantic category label, which includes three types: condition relaxation, condition tightening, and discretion. Each sentence in the original policy text is traversed, and each sentence is matched with the word entries in the fuzzy expression vocabulary library. When any word entry exists in a sentence, the sentence is marked as containing a fuzzy expression. For the marked sentences, the start character position and end character position of the sentence in the original policy text are extracted. The start character position, end character position, the chapter number of the sentence, and the name of the matched word entry are combined into a sentence record unit. All sentence record units are arranged in ascending order of start character position to generate a sentence position list.

3. The method for generating government converged media text based on artificial intelligence according to claim 1, characterized in that, The process of generating sentence templates is as follows: For statements containing fuzzy expressions, the syntactic analysis component in the artificial intelligence model determines the start and end indices of the matched fuzzy expression words in the statement. Using the start index as the split point, the statement is divided into prefix and suffix text segments. Then, using the end index as the split point, the suffix text segment is divided into infix and final suffix text segments. The infix text segment is the fuzzy expression word itself. The prefix text segment, variable placeholders, and final suffix text segment are concatenated in sequence to generate a sentence template with variable placeholders. At the same time, the original word text of the fuzzy expression word replaced by the variable placeholder in the statement is recorded, and a mapping relationship is established between the original word text and the variable placeholder and stored in the mapping table.

4. The method for generating government converged media text based on artificial intelligence according to claim 1, characterized in that, The process of obtaining the set of administrative discretion parameters is as follows: Based on the issuing agency name and document number of the original policy text, the administrative discretion benchmark documents associated with the original policy text are retrieved from the government affairs database. Discretionary rule entries corresponding to each vague expression term are extracted from these administrative discretion benchmark documents. Each discretionary rule entry includes the name of the vague expression term, a description of the applicable scenario, the lower and upper limits of the variable value range, and a description of the applicable conditions for each value segment within the variable value range. All extracted discretionary rule entries are aggregated to generate an administrative discretionary parameter set. When no associated administrative discretion benchmark document exists in the government affairs database, a preset general discretionary rule library is invoked. Based on the semantic category label of the vague expression term, the corresponding default discretionary rule entry is matched from the general discretionary rule library, and this default discretionary rule entry is used as a component of the administrative discretionary parameter set.

5. The method for generating government converged media text based on artificial intelligence according to claim 1, characterized in that, The process of obtaining multiple discrete value points is as follows: For each discretionary rule entry in the administrative discretionary parameter set, obtain the lower limit and upper limit values ​​of the variable value range in the discretionary rule entry. Based on the semantic category label of the fuzzy expression term corresponding to the discretionary rule entry, select the corresponding discretization step size value from the discretization strategy library. Starting from the lower limit value, generate a numerical sequence by increasing the discretization step size value until the maximum value in the numerical sequence does not exceed the upper limit value. Take each value in the numerical sequence as a discrete value point. When there is no discretization step size value corresponding to the semantic category label in the discretization strategy library, take the lower limit value and upper limit value of the variable value range as two discrete value points respectively.

6. The method for generating government converged media text based on artificial intelligence according to claim 1, characterized in that, The process of generating candidate text statements is as follows: An artificial intelligence model is invoked to convert discrete value points into natural language descriptive text. The AI ​​model selects the corresponding conversion template from a pre-set text conversion template library based on the semantic category label of the fuzzy expression words containing the discrete value points. The conversion template includes format rules for converting numerical values ​​to text and quantifier suffix text. The natural language descriptive text is filled into the variable placeholder positions to generate the filled text statement. The original vocabulary text of the fuzzy expression words replaced by the variable placeholders is extracted from the mapping table. The consistency between the filled text statement and the context of the original vocabulary text is checked. When the subject of the filled text statement is inconsistent with the subject of the original vocabulary text, the subject text is extracted from the original vocabulary text and inserted into the beginning of the filled text statement to generate candidate text statements.

7. The method for generating government converged media text based on artificial intelligence according to claim 1, characterized in that, The process of concatenating sentences is as follows: Obtain all statements in the original policy text that are not marked as containing ambiguous expressions. Arrange all statements that are not marked as containing ambiguous expressions in the order of their appearance in the original policy text to generate a fixed text statement sequence. For each candidate text statement, obtain the original position number of the statement containing ambiguous expressions in the original policy text corresponding to the candidate text statement. Insert the candidate text statement into the fixed text statement sequence at the position corresponding to the original position number, and replace the statement containing ambiguous expressions corresponding to the original position number to generate a complete candidate text statement sequence. Concatenate each statement in the candidate text statement sequence in order, inserting newline characters between statements to generate candidate interpretation text.

8. The method for generating government converged media text based on artificial intelligence according to claim 1, characterized in that, The process of outputting candidate interpretation text is as follows: Multiple candidate interpretation texts are sorted according to the numerical values ​​of their corresponding discrete points to generate a candidate interpretation text list. A unique identifier is generated for each candidate interpretation text in the list. The unique identifier is associated with the storage path of the candidate interpretation text and written into an index file. A corresponding applicable condition description text is attached to each candidate interpretation text in the index file. The applicable condition description text is written directly into the entry position of the index file corresponding to the unique identifier in plain text form. The index file and all candidate interpretation texts are packaged and output as a converged media data package.

Images