A customized science popularization communication method and system based on artificial intelligence

Abstract

The application provides a customized popular science communication method and system based on artificial intelligence, and relates to the technical field of artificial intelligence. The application forms a communication context by collecting target audience information and communication constraints, performs semantic analysis on knowledge points and establishes a fact boundary, constructs an acceptability measure in combination with an audience portrait, drives a generator to output candidate expressions, and determines acceptable concepts after checking and evaluation, and dynamically optimizes parameters in combination with feedback signals. Under the premise of ensuring scientific accuracy, customized expression is realized according to different people, and the effectiveness and sustainability of popular science communication are significantly improved through adaptive updating.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a customized science popularization method and system based on artificial intelligence. Background Technology

[0002] Current popular science communication relies heavily on the experience of editors and the recommendation mechanisms of media platforms to select and present content. While this approach can meet the public's information needs to some extent, it often leads to problems such as excessive comprehension burden, inappropriate analogies, or logical jumps when facing audiences of different ages, educational backgrounds, and cultural contexts. Furthermore, existing popular science works generally lack unified control over the granularity of knowledge points, the level of concept expression, and the audience's cognitive threshold, resulting in a lack of balance between accuracy and comprehensibility.

[0003] With the gradual maturation of knowledge graphs, semantic parsing, and content generation control technologies, science popularization is developing towards "structured knowledge, controllable expression, and quantifiable effects." More and more research is attempting to combine cognitive science findings, using cognitive load management and preference modeling to reduce the difficulty of understanding, and leveraging multimodal methods to enhance the intuitiveness and vividness of information delivery. Simultaneously, audience-oriented feedback mechanisms are being gradually introduced to dynamically adjust dissemination strategies, enabling continuous optimization and personalization of the content.

[0004] However, existing methods mostly focus on improvements at the recommendation and copywriting levels, failing to design "generating concepts based on knowledge points that are acceptable to the audience" as a systematic technical approach. The lack of an acceptability model oriented towards audience profiles and a constraint mechanism oriented towards factual boundaries makes the dissemination effect prone to polarization, either oversimplification or over-professionalization. Furthermore, the generation and verification stages are often handled separately, making it difficult to ensure a balance between accuracy and suitability during the content formation stage, thus affecting the stability and promotional value of the dissemination. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a customized science popularization method and system based on artificial intelligence, which can achieve customized expression for different audiences while ensuring scientific accuracy, and significantly improve the effectiveness and sustainability of science popularization through feedback-driven adaptive optimization.

[0006] To achieve the above objectives, the present invention provides the following solution: A customized science popularization method based on artificial intelligence includes: Receive information elements and communication goals from the target audience, obtain the knowledge points to be disseminated and communication constraints, and form a communication context that includes context, length, style, restrictions on sensitive words and risk thresholds; The knowledge points are semantically analyzed and their relationships expanded to generate a set of conceptual elements and dependencies. Fact boundaries are established based on an authoritative knowledge base and a terminology standard base. The fact boundaries define immutable definitions, dimensions, causal relationships, and sources of evidence. Based on the information elements of the target audience, a joint representation of cognitive level, reading habits, cultural context, and risk sensitivity is constructed to establish an acceptability measure; the acceptability measure simultaneously constrains accuracy, consistency, comprehension burden, ambiguity risk, and value fit. Based on the propagation context, the set of conceptual elements, the factual boundaries, and the acceptability measure, the artificial intelligence generator is driven to produce candidate expressions. During the generation process, the factual boundaries are forcibly satisfied, the expression level is selected according to the understanding threshold, and strategies are switched between metaphor, example, analogy, and step-by-step description. The candidate expressions are subjected to factual consistency verification, terminology standardization consistency verification, and acceptability assessment. Only when all three are met is it determined to be an acceptable concept. If the three are not met, the expression level and narrative strategy are adjusted according to the assessment results. The controlled generation and verification are repeated within a preset number of steps until the stopping criterion is met. The acceptable concept is delivered in a way that appeals to the audience and in a length that is appropriate for their preferences. Explicit feedback and implicit interaction signals are collected and used to update the acceptability measurement and strategy selection parameters, thereby improving the matching degree of subsequent generated data without changing the factual boundaries.

[0007] Preferably, the process involves receiving information elements and communication objectives from the target audience, acquiring the knowledge points to be disseminated and communication constraints, and forming a communication context that includes language, length, style, restrictions on sensitive words, and risk thresholds, including: The platform rules, regional policy lists, and citation permission levels in the dissemination constraints are translated into an executable set of constraint rules, and a sensitive word blocking table and initial risk threshold settings are output; the set of constraint rules clearly defines prohibited items, permitted items, evidence level thresholds, and regional scope of application. Based on the terminology standardization database and knowledge relationship network, the knowledge points to be disseminated are expanded and synonymized, and then unitized according to the tags of "definition, principle, application, misunderstanding clarification" to obtain the smallest set of expression units; Based on the information elements and communication goals of the target audience, the communication goals are decomposed into mastery level instructions and expression tone instructions, and an alignment mapping is established with the audience's language habits, reading device types and preferences to generate expression level settings and style preference settings; Using the set of minimum expression units, the set of constraint rules, the setting of expression level, and the setting of style preference as inputs, the parameters of the dissemination context are determined jointly. Specifically, the length is determined by the reading device type and the number of minimum expression units; the style is determined by the style preference setting and the tone indication; the risk threshold is determined by the evidence level threshold and the geographical scope of application in the constraint rule set; sensitive word restrictions are determined by the intersection of the blocking list and platform rules; and the context is determined by the dissemination channel and usage scenario indication. The context, length, style, sensitive word restrictions, and risk thresholds are encapsulated into a structured propagation context object.

[0008] Preferably, the knowledge points are semantically parsed and their relationships expanded to generate a set of conceptual elements and dependencies, and factual boundaries are established based on an authoritative knowledge base and a terminology standardization base, including: The knowledge points to be disseminated are aligned with terms and aliases are merged. The conceptual elements are generated by dividing them into "noun definition units, relational predicate units, quantitative attribute units, and condition / applicable domain units". The dependency template is used to determine the master-slave, limiting, and parallel dependency relationships between the conceptual elements. In the knowledge relationship network, the aforementioned conceptual elements are used as anchor points, and the relationship expansion is limited to four controllable relationship sets: "superior / inferior, composition / part, condition / result, premise / conclusion". Each relationship is labeled with causal role, applicable domain and boundary conditions to form an labeled extended relationship set. The quantity attribute unit is mapped to a unified dimensional system, and the unit of the same physical quantity is standardized and converted to generate dimensional constraint pairs. The dimensional constraint pairs include at least a dimensional identifier, a set of allowed units, a numerical domain, and a precision indicator. Retrieve the normative definitions and evidence sources corresponding to the conceptual elements from the terminology standard library and authoritative knowledge base, extract unique definitions according to preset priorities, and classify and label the evidence according to the authority level of the source and the publication time to form a definition-evidence mapping table. For relationships marked as causal in the extended relation set, cross-constraints are applied with the definition-evidence mapping table, retaining only causal assertions that satisfy definition consistency and evidence level thresholds; when there are conflicts in definition, dimension, or causality, the rules of prioritizing the terminology standard library, evidence level, and timeliness are applied in sequence to adjudicate and fix the result; The adjudicated normative definitions, dimensional constraints, causal assertions, and evidentiary anchors are encapsulated as the fact boundary.

[0009] Preferably, based on the information elements of the target audience, a joint representation of cognitive level, reading habits, cultural context, and risk sensitivity is constructed to establish an acceptability metric, including: Based on the information elements of the target audience, the original elements corresponding to cognitive level, reading habits, cultural context and risk sensitivity are mapped to audience indicators, and fixed as a discrete set of indicators with a clear value range; the audience indicators include at least abstract tolerance indicators, vocabulary familiarity indicators, paragraph rhythm preference indicators, rhetorical strategy preference indicators, regional pragmatic variant indicators, taboo topic indicators and controversy avoidance indicators. Using the discrete indicator set as input, a quaternary joint representation including a cognitive level profile, a reading habit profile, a cultural context profile, and a risk sensitivity profile is generated. Based on the joint representation, an expression level interval, a set of allowed rhetorical strategies, and a list of prohibited citations are derived. The list of prohibited citations is compatible with the evidence anchors and dimensional constraints in the fact boundary. Using the aforementioned joint representation as a constraint source, an acceptability measurement object consisting of five components is established, and thresholds or acceptable ranges are set for each component. Among them, the accuracy threshold is a hard constraint on the object of the aforementioned fact boundary and cannot be violated; the consistency threshold is a hard constraint on the pair of terminology norms and dimensional constraints; the comprehension burden index is defined by the density of new terms, the upper limit of sentence length, and the upper limit of reasoning span; the ambiguity risk index is defined by the coverage of referential resolution and the sufficiency of polysemous word resolution; and the value fit index is defined by the matching degree of the rhetorical strategy set, the regional pragmatic variant indication, and the prohibited citation list. The thresholds, indicators, and their thresholds are solidified into executable objects of the acceptability measurement, and a set of conditional parameters for the generation phase is output, including a set of hard constraints and a target range of delimited quantities.

[0010] Preferably, the AI ​​generator is driven to produce candidate representations based on the propagation context, the set of conceptual elements, the factual boundaries, and the acceptability measure. During the generation process, the factual boundaries are forcibly satisfied, the expression level is selected according to a comprehension threshold, and strategies are switched between metaphor, example, analogy, and step-by-step description, including: Using the object of the propagation context, the object of the fact boundary, the set of conceptual elements and their dependencies, and the object of the acceptability measurement as joint input, during the decoding process, terms and sentence structures that are inconsistent with the definition, causality, or dimensionality are intercepted or replaced, and content that references below the level of evidence anchors or triggers sensitive word restrictions is rewritten in a restrictive manner. Limited to the objects of the acceptability measurement, the density of new terms, sentence length and reasoning span of the generated fragments are measured online; when any indicator exceeds the target range, the expression level is automatically reduced, the scope of term substitution and the granularity of explanation are tightened; when the indicator is below the target range, the expression level is increased and the allowed rhetorical depth is opened, so that the indicator returns to the target range. Within the scope and body of the communication context, the choice is made between “exemplary / step-by-step narration” and “analogy / metaphor” based on the set of rhetorical strategies allowed by the acceptability measurement object; when ambiguity risks or regional pragmatic mismatches are triggered, local rewriting and replacement are carried out according to the evidence anchors and terminology norms of the fact boundary object. When the definition / causal / dimensional constraints in the fact boundary two-egg object are not fully satisfied, or the accuracy threshold and consistency threshold in the acceptability measurement object are not simultaneously satisfied, the relevant fragments are partially backtracked and rewritten along the most recent rhetorical selection point, without regenerating the compliant parts; after the constraints and thresholds are satisfied and the measurement index falls into the target range, the fragments are merged to form the candidate expression, and the correspondence between the rhetorical strategy sequence and the evidence anchor point is recorded.

[0011] Preferably, the candidate expressions undergo factual consistency verification, terminology standardization consistency verification, and acceptability assessment. An acceptable concept is determined only when all three criteria are met simultaneously. If the criteria are not met, the expression level and narrative strategy are adjusted based on the assessment results. Controlled generation and verification are repeated within a preset number of steps until a stopping criterion is met, including: a) Based on the object of the fact boundary, align the semantic fragments and conceptual elements in the candidate statements, and sequentially perform boundary predicate matching, evidence anchor level comparison and unit normalization comparison on statements involving definitions, causality and quantitative attributes; any statements that are inconsistent with the definition, causality or dimension, or whose citation level is lower than the evidence anchor threshold, are marked as non-compliant fragments and the start and end positions of the non-compliant fragments are located. b) Based on the terminology standard library and the set of conceptual elements and their dependencies, examine the consistency of term morphology, aliasing and master-detail dependency relationships in candidate expressions; for terms that have cross-level mismatch, term morphology drift or do not conform to dependency relationships, output corresponding standard substitution suggestions and maintain the original dependency relationships unchanged. c) Using the object of the acceptability measurement as the source of constraints, calculate the understanding burden index, ambiguity risk index and value fit index for candidate statements. Only when the accuracy threshold and consistency threshold are met and each index falls within its target range is it determined to be qualified. d) If any of steps a) to c) fails to meet the standard, perform local controlled rewriting on the candidate expression according to the type of failure. The rewriting is limited to adjusting the expression level and narrative strategy without changing the definition, causality and dimension in the fact boundary object, and keeping the terminology and dependency relationships unchanged. After rewriting, the relevant fragments will be generated in a controlled manner and returned for review. The number of iterations is limited by a preset number of steps. When all three types of verification and evaluation meet the standard at the same time or when the preset number of steps is reached but the standard is still not met, output the acceptable concept or failure flag and corresponding position annotation to end this process.

[0012] Preferably, the comprehension burden indicators include neo-term density, sentence length, and inference span; the ambiguity risk indicators include referential resolution coverage and polysemous word resolution sufficiency; and the value fit indicators include rhetorical strategies and regional pragmatic matching degree.

[0013] Preferably, the acceptable concept is delivered using a carrier and length preferred by the audience, and explicit feedback and implicit interaction signals are collected and used to update the acceptability measurement and strategy selection parameters, thereby improving the matching degree of subsequent generation without changing the factual boundaries, including: Based on the object of the aforementioned dissemination context, without changing the meaning of the content and the anchor of evidence, the acceptable concepts are rearranged into delivery units consistent with the carrier type and length limitations, while retaining stylistic limitations and citation annotations for subsequent retrospective purposes; The explicit feedback and implicit interaction signals after delivery are processed into indicators to generate comprehension burden bias indicators, ambiguity risk indicators, and value adaptation bias indicators. The explicit feedback includes evaluation and error correction labels, and the implicit interaction signals include dwell and reread features and interaction trigger locations. Using the object of the acceptability measurement as the source of constraints, without changing the definition, causal and dimensional constraints in the object of the fact boundary, the target range and threshold of the acceptability measurement are fine-tuned according to three types of instructions, and the strategy selection parameters are adjusted simultaneously; the strategy selection parameters include expression level, interpretation granularity, term substitution range and rhetorical strategy weight, and are accompanied by version number and scope of application.

[0014] A customized science popularization and dissemination system based on artificial intelligence includes: The audience-target analysis unit is used to receive information elements and communication goals from the target audience, obtain the knowledge points to be disseminated and communication constraints, and form a communication context that includes context, length, style, sensitive word restrictions and risk thresholds; The knowledge point semantic parsing unit is used to perform semantic parsing and relation expansion on the knowledge points, generate a set of conceptual elements and dependency relationships, and establish fact boundaries based on an authoritative knowledge base and terminology standard base; the fact boundaries define immutable definitions, dimensions, causal relationships and sources of evidence. The audience acceptability modeling unit is used to construct a joint representation of cognitive level, reading habits, cultural context and risk sensitivity based on the information elements of the target audience, and to establish an acceptability measure; the acceptability measure simultaneously constrains accuracy, consistency, comprehension burden, ambiguity risk and value fit; The controlled generation unit is used to drive the artificial intelligence generator to generate candidate expressions based on the propagation context, the set of conceptual elements, the factual boundaries, and the acceptability measure. During the generation process, the factual boundaries are forcibly satisfied, the expression level is selected according to the understanding threshold, and strategies are switched between metaphor, example, analogy, and step-by-step description. The verification-evaluation unit is used to perform factual consistency verification, terminology standardization consistency verification, and acceptability evaluation on the candidate statements. It is determined to be an acceptable concept only when all three are met. If they are not met, the expression level and narrative strategy are adjusted according to the evaluation results. The controlled generation and verification are repeated within a preset number of steps until the stopping criterion is met. The delivery-feedback update unit is used to deliver the acceptable concept in a way that meets the audience's preferred format and length, collect explicit feedback and implicit interaction signals, and use them to update the acceptability measurement and strategy selection parameters, thereby improving the matching degree of subsequent generation without changing the factual boundaries.

[0015] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: First, by establishing a communication context and factual boundaries, this invention solidifies the definition, causality, dimension, and source of evidence of knowledge points into strong constraints, thereby fundamentally ensuring the accuracy and authority of the generated popular science statements and avoiding the problems of oversimplification or arbitrary expansion caused by the lack of factual boundaries in existing methods.

[0016] Second, this invention models the information elements of the target audience to form a joint representation of cognitive level, reading habits, cultural context and risk sensitivity, and constructs an acceptability metric based on this. This solves the problem in the prior art that it fails to effectively combine the differences in audience cognition, resulting in the content being difficult for different groups to fully understand and accept.

[0017] Third, this invention introduces a mechanism for scheduling understanding thresholds and switching rhetorical strategies during the generation process, which allows the generated results to be flexibly adjusted between metaphors, examples, analogies and step-by-step descriptions. This reduces the burden of understanding while ensuring scientific rigor, and overcomes the shortcomings of traditional methods that have a single mode of expression and difficulty in balancing comprehensibility and completeness.

[0018] Fourth, this invention, through a joint mechanism of factual consistency verification, terminology standardization consistency verification, and acceptability assessment of candidate expressions, places quality control in advance and embeds it into the generation process. This avoids the fragmented "generation-post-verification" model in existing methods and significantly improves the stability and reusability of disseminated content in terms of accuracy and adaptability.

[0019] Fifth, this invention collects explicit feedback and implicit interaction signals during the delivery process and dynamically updates the acceptability measurement and strategy selection parameters accordingly. It achieves adaptive optimization without changing the factual boundaries, effectively overcoming the shortcomings of traditional science popularization communication, which lacks closed-loop adjustment and makes it difficult to continuously improve the communication effect. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart of the method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention. Detailed Implementation

[0022] 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.

[0023] The purpose of this invention is to provide a customized science popularization method and system based on artificial intelligence. Through the collaborative design of fact boundary constraints and audience modeling, it achieves a dynamic balance between accurate delivery of scientific content and personalized expression, and significantly improves the pertinence and continuous optimization capability of science popularization under the feedback-driven update mechanism.

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a customized science popularization method based on artificial intelligence, including: Step 100: Receive information elements and communication goals from the target audience, obtain the knowledge points to be disseminated and communication constraints, and form a communication context that includes context, length, style, sensitive word restrictions and risk thresholds; Step 200: Perform semantic analysis and relation expansion on knowledge points to generate a set of conceptual elements and dependencies, and establish fact boundaries based on an authoritative knowledge base and terminology standard base; the fact boundaries define immutable definitions, dimensions, causal relationships and sources of evidence; Step 300: Construct a joint representation of cognitive level, reading habits, cultural context and risk sensitivity based on the information elements of the target audience, and establish an acceptability measure; the acceptability measure simultaneously constrains accuracy, consistency, comprehension burden, ambiguity risk and value fit; Step 400: Drive the AI ​​generator to produce candidate expressions based on the propagation context, set of conceptual elements, factual boundaries, and acceptability measures. During the generation process, the factual boundaries are forcibly satisfied, the expression level is selected according to the understanding threshold, and strategies are switched between metaphor, example, analogy, and step-by-step narrative. Step 500: Perform factual consistency verification, terminology standardization consistency verification, and acceptability assessment on candidate expressions. Only when all three are met is it determined to be an acceptable concept. If the three are not met, adjust the expression level and narrative strategy according to the assessment results. Repeat controlled generation and verification within a preset number of steps until the stopping criterion is met. Step 600: Deliver acceptable concepts using media and length preferred by the audience, collect explicit feedback and implicit interaction signals, and use them to update the acceptability measurement and strategy selection parameters to improve the matching degree of subsequent generation without changing the factual boundaries.

[0026] In a preferred embodiment, step 100 of the present invention first receives information elements and dissemination goals from the target audience, and obtains the medical knowledge points to be disseminated and dissemination constraints. These dissemination constraints include platform rules, a list of regional policies, and citation license levels, which are translated into an executable set of constraint rules. The "constraint rule set" refers to a structured set of rules that can be automatically processed after compilation, used to clarify prohibited items, permitted items, evidence level thresholds, and regional applicability in medical science popularization content. For example, in the dissemination of knowledge about tumor screening, the constraint rule set can restrict certain drug descriptions to come from authoritative literature approved by the national drug regulatory authority, while blocking sensitive descriptions that have not been clinically verified, thereby outputting a sensitive word blocking table and initial risk threshold settings.

[0027] Furthermore, this invention, based on a terminology standardization database and a knowledge relationship network, performs semantic analysis and dependency expansion on the medical knowledge points to be disseminated, forming a set of minimal expression units. The so-called "set of minimal expression units" refers to breaking down complex medical knowledge points into fine-grained concepts that can be independently expressed, supported by a knowledge graph or knowledge relationship network. Examples include unitized content under four categories: "definition, principle, application, and clarification of misunderstandings." In the context of clinical nutrition knowledge dissemination, the nutrient "vitamin D" can be broken down into: "Definition: Vitamin D is a fat-soluble vitamin," "Principle: It affects bone health by regulating calcium and phosphorus metabolism," "Application: It can be used for osteoporosis prevention," and "Clarification of misunderstandings: Long-term excessive supplementation can lead to poisoning." This approach ensures that medical content is disseminated without omitting core points or using vague expressions.

[0028] In another implementation step, based on the information elements and communication goals of the target audience, this invention further decomposes the communication goals into mastery level indicators and expression tone indicators. The mastery level indicators determine the depth of medical knowledge presented, while the expression tone indicators control the objectivity and rigor of the language style. For example, for undergraduate medical students, the mastery level can be set to "intermediate," allowing the use of professional terminology with explanations; while for community residents, it is set to "beginner," using everyday metaphors instead of complex professional vocabulary. Simultaneously, this invention generates corresponding expression level and stylistic preference settings by aligning and mapping with the audience's language habits, reading device type (such as mobile phone or tablet), and content preferences.

[0029] Finally, this invention uses the aforementioned set of minimum expression units, set of constraint rules, expression level settings, and stylistic preference settings as inputs to jointly determine various parameters of the dissemination context. Specifically, the length is jointly determined by the reading device type and the number of minimum expression units; for example, on a smartphone, each paragraph is limited to no more than 100 characters. The style is jointly determined by stylistic preference settings and expression tone indicators; for example, a colloquial style is used when addressing the elderly. The risk threshold is determined by the evidence level threshold and geographical applicability in the constraint rule set; for example, the citation level of medical guidelines varies in different countries. Sensitive word restrictions are determined by the intersection of the blocking list and platform rules to prevent the dissemination of unapproved drug names or efficacy claims. The context is determined by the dissemination channel and usage scenario indicators; for example, different context labels are set in hospital health education columns and social media platforms. Ultimately, the system encapsulates the context, length, style, sensitive word restrictions, and risk thresholds into a structured dissemination context object. The so-called "dissemination context object" refers to a structured data unit that integrates the above parameters and can be directly called by subsequent generation modules to drive the personalized generation of medical science popularization content.

[0030] In a preferred embodiment, step 200 of the present invention performs semantic parsing and relation expansion on the medical knowledge points to be disseminated. First, terminology alignment and alias merging are performed to generate a set of conceptual elements and establish dependency relationships. The "set of conceptual elements" refers to breaking down medical knowledge points into basic units with clear functions, including noun definition units, relational predicate units, quantitative attribute units, and conditional / applicable domain units. For example, in the dissemination of cardiovascular disease knowledge, "blood pressure" serves as a noun definition unit, "elevation" as a relational predicate unit, "120 / 80 mmHg" as a quantitative attribute unit, and "resting state" as a conditional unit. Dependency templates further determine the master-slave, limiting, and parallel relationships between these units, thereby ensuring the semantic structure of the knowledge points is complete and readily accessible.

[0031] Furthermore, this invention uses conceptual elements as anchors within a knowledge relationship network, limiting relationship expansion to four controllable relationship sets: "superior / inferior," "component / part," "condition / result," and "premise / conclusion." A "dependency relationship" refers to the logical connection between different conceptual elements; for example, "hypertension (premise) — increased cardiac load (result)" is a condition / result relationship. During the expansion process, each relationship is labeled with its causal role, applicable domain, and boundary conditions, resulting in an expanded relationship set. For example, in diabetes education, the relationship "long-term elevated blood sugar (condition) — retinopathy (result)" would be explicitly labeled as a causal expanded relationship.

[0032] In another implementation step, the present invention performs a unified dimensionalization process on the quantitative attribute units, forming dimensional constraint pairs. A "dimensional constraint pair" refers to mapping medical measurement data to a unified physical quantity framework, ensuring standardized conversions between different units and numerical domains. For example, blood glucose values ​​can be expressed in mmol / L or mg / dL. The present invention maps them to the same dimensional system through dimensional constraint pairs, and limits the allowed unit sets, numerical domains, and precision ranges to ensure consistent representation and avoid ambiguity when disseminating data from different sources.

[0033] Finally, this invention retrieves the normative definitions and evidence sources corresponding to the conceptual elements from a terminology standardization database and an authoritative knowledge base. Based on a preset priority (the terminology standardization database takes precedence over the authoritative knowledge base, which in turn takes precedence over other references), unique definitions are extracted and graded according to the authority level and timeliness of the source, forming a definition-evidence mapping table. For relationships marked as causal in extended relationships, they are cross-checked against the definition-evidence mapping table, retaining only causal assertions that satisfy definition consistency and evidence level thresholds. When conflicts arise, the principle of prioritizing the terminology standardization database, evidence level, and timeliness is applied sequentially to obtain a final, fixed result. All adjudicated normative definitions, dimensional constraints, causal assertions, and evidence anchors are encapsulated as fact boundaries. The so-called "fact boundaries" refer to hard constraints that cannot be broken during content generation; their core function is to ensure that medical science popularization content does not deviate from authoritative definitions or incorrectly cite evidence during interpretation. For example, in the popularization of osteoporosis, the medical definition "bone density below a certain threshold is osteoporosis" and its corresponding international guideline evidence are included in the fact boundaries and are forcibly maintained during dissemination.

[0034] In a preferred embodiment, step 300 of the present invention first maps raw data corresponding to cognitive level, reading habits, cultural context, and risk sensitivity based on the information elements of the target audience, obtaining a set of audience indicators. The so-called "audience indicators" refer to quantifiable markers reflecting the specific characteristics of the audience's understanding of medical content. Their value range is clear and discrete, facilitating subsequent calculations. Specifically, they include: abstraction tolerance indicators, vocabulary familiarity indicators, paragraph rhythm preference indicators, rhetorical strategy preference indicators, regional pragmatic variant indicators, taboo topic indicators, and controversy avoidance indicators. For example, for elderly audiences attending community health lectures, their "vocabulary familiarity indicators" tend to favor commonly used medical terms, while their "abstraction tolerance indicators" tend to be low, avoiding excessive use of abstract models or complex pathological mechanisms.

[0035] Furthermore, this invention uses the discrete audience indicator set as input to generate a quaternary joint representation comprising a cognitive level profile, a reading habit profile, a cultural context profile, and a risk sensitivity profile. The so-called "quaternary joint representation" refers to combining the four profiles with a unified time base and weight parameters to obtain a multi-dimensional vector structure that reflects the overall acceptance characteristics of the audience. This joint representation not only describes the static characteristics of the audience but can also derive expression level intervals, a set of permitted rhetorical strategies, and a list of prohibited citations. For example, in a medical science popularization scenario, the joint representation targeting young medical students might derive higher-level intervals and a wider set of rhetorical strategies; while for ordinary patients, it derives lower-level intervals and generates a list of prohibited citations including "avoid using drug names that have not been clinically approved."

[0036] In another embodiment, the present invention utilizes the aforementioned quaternary joint representation as a constraint source to establish an acceptability measurement object consisting of five components. The so-called "acceptability measurement object" refers to a parameterized structure that uniformly encapsulates evaluation indicators of different dimensions, used for direct invocation during the generation and verification stages. The five components defined in this invention include: accuracy threshold, consistency threshold, comprehension burden indicator, ambiguity risk indicator, and value fit indicator. Specifically, the accuracy threshold uses factual boundary objects as hard constraints, allowing no deviations; the consistency threshold is determined by terminology standardization and dimensional constraints; the comprehension burden indicator is defined by new term density, sentence length upper limit, and reasoning span upper limit; the ambiguity risk indicator is measured by referential resolution coverage and polysemous word resolution sufficiency; and the value fit indicator is defined by the matching degree of rhetorical strategy set, regional pragmatic variant indication, and prohibited citation list.

[0037] In practical applications, taking the dissemination of diabetes prevention and treatment knowledge as an example, the acceptability metrics differ for different audiences. For undergraduate medical students, the accuracy threshold requires close adherence to the latest international clinical guidelines, while the consistency threshold allows the use of professional international units. The upper limit for sentence length in the burden of understanding indicator can be set at 40 words, suitable for reading highly technical paragraphs. The ambiguity risk indicator allows for a certain degree of polysemy, such as the word "metabolism." The value fit indicator allows for the use of professional metaphors. In contrast, for community residents, the accuracy threshold still needs to follow the guidelines, but the upper limit for sentence length in the burden of understanding indicator is lowered to 20 words, and citations involving the efficacy of controversial drugs are prohibited, thus ensuring comprehensibility and safety.

[0038] Finally, this invention solidifies the aforementioned thresholds, indicators, and their thresholds into executable acceptability measurement objects and outputs conditional parameters for the generation stage. These "conditional parameters" are a set of parameters consisting of a set of hard constraints and a defined target range, which can directly influence the selection and scheduling of candidate expressions during the controlled generation process. For example, when generating medical content, if the density of new terminology exceeds the comprehension burden threshold of the community audience, the system will automatically lower the expression level and adopt a more straightforward rhetoric; if the cited content is found to be below the level requirement of the evidence anchor point, the candidate expression will be eliminated. In this way, this invention ensures that the generated content is both scientifically sound and comprehensible, thereby achieving precise dissemination to different groups within the medical field.

[0039] In a preferred embodiment, step 400 of the present invention uses a propagation context object, a fact boundary object, a set of conceptual elements and their dependencies, and an acceptability measurement object as joint inputs to constrain the decoding process of the artificial intelligence generator. The "propagation context object" refers to the structured data unit established in step 100, which includes context, length, style, sensitive word restrictions, and risk thresholds; the "fact boundary object" is the set of hard constraints generated in step 200, which includes definitions, dimensions, causality, and sources of evidence. This step, by calling these objects during the decoding process, intercepts or replaces terms and sentence structures inconsistent with definitions, causality, or dimensions, and performs constrained rewriting on content that references below the level of evidence anchors or triggers sensitive word restrictions, thereby ensuring the scientific accuracy and compliance of the output text.

[0040] Furthermore, this invention introduces an online measurement mechanism during the generation process. Using an acceptability measurement object as a constraint, it continuously calculates the density of new terms, sentence length, and reasoning span of the generated fragments. The "acceptability measurement object" refers to the parameterized structure established in step 300, which includes an accuracy threshold, a consistency threshold, a comprehension burden index, an ambiguity risk index, and a value fit index. For example, in the dissemination of diabetes diet science, when the generated content contains too many new terms (such as "glycated hemoglobin" and "low-density lipoprotein"), causing the density of new terms to exceed the target range for the target audience, the system automatically lowers the expression level and uses more intuitive expressions such as "long-term blood glucose levels" to replace them, ensuring that the comprehension burden is controllable.

[0041] In another embodiment, this invention introduces a rhetorical strategy gating mechanism. Within the limited space and text of the communication context, it switches between "exemplary / step-by-step narration" and "analogy / metaphor" based on the set of rhetorical strategies allowed in the acceptability measurement object. The "rhetorical strategy set" refers to a predefined set of expressions designed to ensure the readability and audience acceptance of medical knowledge. For example, in popularizing osteoporosis knowledge, if the target audience is elderly residents, the system will prioritize "exemplary / step-by-step narration," such as "three steps to daily calcium supplementation"; if the audience is medical undergraduates, it can use "analogy / metaphor," such as "bones are like the steel bars of a house; a lack of calcium makes them brittle," to enhance learning interest and comprehension efficiency.

[0042] Furthermore, this invention introduces a local backtracking and rewriting mechanism. When it detects that the definition, causal, or dimensional constraints in the fact boundary object are not fully satisfied, or that the accuracy threshold and consistency threshold in the acceptability measurement object are not simultaneously satisfied, the relevant segment is locally backtracked along the most recent rhetorical strategy switching point, without regenerating the compliant part. "Local backtracking" means regenerating only the problematic segment, rather than starting from scratch, to ensure efficiency and stability. For example, in cardiovascular health popularization, if the generated sentence "High blood pressure always leads to heart disease" violates the causal constraints in the fact boundary (because the causal relationship is not absolute), the system will backtrack and rewrite that segment as "Long-term high blood pressure may increase the risk of heart disease," maintaining scientific accuracy.

[0043] Finally, after meeting the requirements of fact boundary constraints and acceptability metrics, this invention merges the generated fragments to form candidate statements. A "candidate statement" refers to the text result before controlled generation and verification, which has passed fact constraints and metric scheduling but requires further evaluation. This invention simultaneously records the correspondence between rhetorical strategy sequences and evidence anchors. The rhetorical strategy sequence reflects the rhetorical path used during text generation (e.g., first using "examples," then switching to "metaphors"); the evidence anchor correspondence maps key arguments in the text to evidence sources defined in the fact boundaries. In specific applications in the medical field, such as cancer screening education, candidate statements retain "World Health Organization guidelines" as evidence anchors and cite them accordingly in sentences to ensure authority and traceability.

[0044] In a preferred embodiment, step 500 of the present invention first performs a factual consistency check on the candidate statements. The so-called "factual consistency check" refers to aligning the semantic segments in the candidate statements with the set of conceptual elements, using the fact boundary object formed in step 200 as a benchmark. For statements involving definitions, causality, and quantitative attributes, boundary predicate matching, evidence anchor level comparison, and unit normalization comparison are performed sequentially. For example, in the popularization of medical science regarding hypertension, if a candidate statement contains the description "blood pressure higher than 150 / 100 mmHg inevitably leads to heart disease," the present invention will detect that this causal statement is inconsistent with the causal constraint in the fact boundary "elevated blood pressure only increases the risk of heart disease," thereby marking the segment as a non-compliant segment and accurately locating the start and end positions of the non-compliant content in the statement.

[0045] Furthermore, this invention performs terminology standardization consistency verification on candidate expressions. "Terminology standardization consistency verification" refers to checking the consistency of terminology morphology, aliasing, and master-slave dependency relationships in candidate expressions based on a terminology standardization library and a set of conceptual elements and their dependencies. For example, in the medical communication of diabetic dietary guidelines, "diabetic retinopathy" cannot be arbitrarily replaced with "diabetic eye disease" because their master-slave dependency relationships in the terminology standardization library are different. This invention, through terminology standardization consistency verification, ensures that candidate expressions do not drift in terms of word choice and relational structure, outputs alternative suggestions when non-standard situations occur, and maintains the original dependency relationships without being broken.

[0046] In another implementation step, the present invention uses an acceptability measurement object as a constraint source to conduct an acceptability assessment of candidate statements. The "acceptability assessment" refers to calculating the comprehension burden index, ambiguity risk index, and value fit index of candidate statements based on the measurement object formed in step 300. When both the accuracy threshold and consistency threshold are met, and each index falls within the target range, the candidate statement is deemed acceptable. For example, in medical communication for cancer screening, if a candidate statement uses excessively long sentences, causing the comprehension burden index to exceed the target range for the elderly population, the statement will be deemed unacceptable and requires adjustment of the expression.

[0047] Finally, if any of the factual consistency check, terminology standardization consistency check, or acceptability assessment fails to meet the standards, this invention triggers a partial controlled rewriting mechanism. "Partial controlled rewriting" refers to adjusting the expression level and narrative strategy only for the substandard fragments, without changing the definitions, causal relationships, and dimensional constraints within the factual boundary objects, while maintaining the terminology and dependency relationships unchanged. The rewritten fragments re-enter the controlled generation stage and return to this step for review until all three types of checks and assessments simultaneously meet the standards or the preset number of steps is reached. For example, in medical communication about osteoporosis, when a candidate statement cites an unverified folk remedy that violates the prohibited citation list, the system will automatically replace it with a recommended treatment plan from authoritative medical guidelines. The final output is an "acceptable concept," or, if the preset number of iterations has been reached and the standards are still not met, a "failure flag" is output, along with a non-compliant location annotation, to ensure the scientific accuracy, comprehensibility, and compliance of the disseminated content.

[0048] In a preferred embodiment, step 600 of the present invention first rearranges acceptable concepts into delivery units consistent with the carrier type and length constraints, based on the communication context, without changing the meaning of the medical content and the anchor of evidence. A "delivery unit" refers to the smallest communication block after length segmentation and stylistic constraints adjustment, which can be directly presented on different media. For example, in medical applications, for paper posters in hospital health information boards, delivery units are formatted as short paragraphs, each not exceeding 80 words; while in mobile app push notifications, delivery units are presented in a graphic format, with each push card corresponding to a key knowledge point, and citations are retained for later reference.

[0049] Furthermore, this invention collects explicit feedback and implicit interaction signals from the audience after delivery. "Explicit feedback" refers to user-expressed evaluations and corrections, such as clicking the "easy / difficult to understand" button or manually modifying medical terminology. "Implicit interaction signals" refer to behavioral characteristics that can be obtained without active input from the audience, such as dwell time, number of rereads, and interaction trigger locations in popular science articles. The system processes this feedback and interaction data into indicators, generating three types of quantifiable metrics: comprehension burden bias indicator, ambiguity risk indicator, and value fit bias indicator. For example, in medical communication of diabetes diet guidance, if the audience spends too much time on the "carbohydrate metabolism" section and rereads it multiple times, a "comprehension burden bias indicator" will be generated, suggesting that this section needs simplification.

[0050] In another embodiment, the present invention uses the acceptability measurement object as a constraint source. Without altering the definitions, causal relationships, and dimensional constraints within the fact boundary object, it fine-tunes the target range and threshold of the acceptability measurement according to the aforementioned three types of indicators. The "acceptability measurement object" refers to a parameterized structure composed of accuracy thresholds, consistency thresholds, comprehension burden indicators, ambiguity risk indicators, and value fit indicators, which plays a constraining role in the generation process. In a medical and health education scenario, if elderly audiences report an excessive comprehension burden, the system will lower the sentence length upper limit and reduce the new terminology density threshold; if feedback from medical students indicates that the expression is too simplistic, the system will increase the reasoning span upper limit and abstraction tolerance.

[0051] Finally, this invention updates the strategy selection parameters simultaneously while adjusting the acceptability metric. The "strategy selection parameters" refer to the set of parameters controlling the expression method of the generation process, including expression level, explanation granularity, terminology substitution range, and rhetorical strategy weights. Each update is accompanied by a version number and applicable scope to indicate the conditions for the updated effect. For example, in medical communication for cancer screening, if feedback indicates that the population prefers a more relatable, everyday narrative style, the rhetorical strategy weights will shift towards "example / step-by-step narration," forming a set of strategy parameters with a version number in the system for subsequent generation stages. Through this mechanism, this invention achieves continuous adaptation and dynamic optimization of science communication content to the target audience while ensuring that factual boundaries are not breached.

[0052] Corresponding to the above method, this embodiment also provides a customized science popularization and dissemination system based on artificial intelligence, including: The audience-target analysis unit is used to receive information elements and communication goals from the target audience, obtain the knowledge points to be disseminated and communication constraints, and form a communication context that includes context, length, style, sensitive word restrictions and risk thresholds; The knowledge point semantic parsing unit is used to perform semantic parsing and relation expansion on the knowledge points, generate a set of conceptual elements and dependency relationships, and establish fact boundaries based on an authoritative knowledge base and terminology standard base; the fact boundaries define immutable definitions, dimensions, causal relationships and sources of evidence. The audience acceptability modeling unit is used to construct a joint representation of cognitive level, reading habits, cultural context and risk sensitivity based on the information elements of the target audience, and to establish an acceptability measure; the acceptability measure simultaneously constrains accuracy, consistency, comprehension burden, ambiguity risk and value fit; The controlled generation unit is used to drive the artificial intelligence generator to generate candidate expressions based on the propagation context, the set of conceptual elements, the factual boundaries, and the acceptability measure. During the generation process, the factual boundaries are forcibly satisfied, the expression level is selected according to the understanding threshold, and strategies are switched between metaphor, example, analogy, and step-by-step description. The verification-evaluation unit is used to perform factual consistency verification, terminology standardization consistency verification, and acceptability evaluation on the candidate statements. It is determined to be an acceptable concept only when all three are met. If they are not met, the expression level and narrative strategy are adjusted according to the evaluation results. The controlled generation and verification are repeated within a preset number of steps until the stopping criterion is met. The delivery-feedback update unit is used to deliver the acceptable concept in a way that meets the audience's preferred format and length, collect explicit feedback and implicit interaction signals, and use them to update the acceptability measurement and strategy selection parameters, thereby improving the matching degree of subsequent generation without changing the factual boundaries.

[0053] The beneficial effects of this invention are as follows: First, this invention introduces "fact boundary objects" (including normative definitions, dimensional constraint pairs, causal assertions, and evidence anchors) to impose rigid constraints on definitions, units, and causality during the generation stage, preventing arbitrary rewriting or inappropriate inferences and suppressing the distortion and exaggeration of popular science content from the source. In medical scenarios, guideline-level evidence and units of measurement (such as the conversion between mmol / L and mg / dL) are forced to be consistent, significantly reducing the risk of confusion in terminology and incorrect citations, and enhancing scientific rigor and authority.

[0054] Second, this invention employs a chain-like modeling approach—"audience indicators—quaternary joint representations—acceptability measurement objects"—to provide quantifiable thresholds and target ranges (such as new terminology density, sentence length limits, and reasoning spans) for different population groups based on their cognitive levels, reading habits, cultural contexts, and risk sensitivities. Consequently, the selection of expression levels and rhetorical strategies (examples, analogies, step-by-step narration) has clear triggering conditions and fallback rules, reducing the burden of understanding without sacrificing scientific accuracy and significantly improving acceptability for medical audiences (such as elderly residents and medical students).

[0055] Third, this invention proposes a controlled generation mechanism of "boundary constraint decoding + understanding threshold scheduling + rhetorical gating + local backtracking," forming an online scheduling closed loop oriented towards quality goals (without changing the factual boundaries). When an indicator goes out of bounds or a pragmatic mismatch is detected, the system only performs local backtracking and rewriting on the problematic segment instead of a full recalculation, maintaining output stability while significantly improving generation efficiency. In medical descriptions, "absolute causality" can be automatically corrected to compliant expressions such as "increased risk," balancing rigor and readability.

[0056] Fourth, this invention embeds factual consistency verification, terminology standardization consistency verification, and acceptability assessment in parallel into the post-generation admission process. Only when all three meet the standards can an "acceptable concept" be output. This "pre-emptive quality gate" design replaces traditional post-production manual revisions, significantly reducing the cost of manual review and the risk of publication. It is particularly effective for highly sensitive medical information (drugs, efficacy, contraindications), ensuring the stability and reusability of cross-platform, multi-batch publications through engineering.

[0057] Fifth, after delivery, this invention performs indicative processing on explicit feedback and implicit interaction signals (understanding burden bias, ambiguity risk, value fit bias), and fine-tunes the acceptability measurement objects and strategy selection parameters (expression level, interpretation granularity, terminology substitution range, rhetorical weight) accordingly, while also managing them through version numbers and scope of application. This approach ensures continued alignment with the preferences of specific audience groups without violating the hard constraints of "factual boundary objects," guaranteeing scientific accuracy while achieving long-term optimization for different target audiences and channels.

[0058] Sixth, this invention constructs a traceable, auditable, and portable engineering system using a modular data structure consisting of "propagation context object—conceptual element set and its dependencies—fact boundary object—acceptability measurement object—rhetorical strategy sequence—evidence anchor point correspondence." This structure facilitates integration into multiple systems such as hospital education platforms, medical science popularization apps, and hospital WeChat official accounts, enabling rapid, large-scale reuse and version management. While increasing content productivity, it significantly reduces the probability of compliance and quality incidents.

[0059] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0060] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A customized science popularization method based on artificial intelligence, characterized in that, include: The system receives information elements and communication goals from the target audience, obtains the knowledge points to be disseminated and communication constraints, and forms a communication context that includes context, length, style, sensitive word restrictions, and risk thresholds. The platform rules, regional policy lists, and citation license levels in the communication constraints are translated into an executable set of constraint rules. The risk thresholds are determined by the evidence level threshold and regional applicability in the set of constraint rules. The knowledge points are semantically analyzed and their relationships expanded to generate a set of conceptual elements and dependencies. Fact boundaries are established based on an authoritative knowledge base and a terminology standard base. The fact boundaries define immutable definitions, dimensions, causal relationships, and sources of evidence. Based on the information elements of the target audience, a joint representation of cognitive level, reading habits, cultural context, and risk sensitivity is constructed to establish an acceptability measure. The acceptability measure simultaneously constrains accuracy, consistency, comprehension burden, ambiguity risk, and value fit. The value fit index is defined by the matching degree of a set of rhetorical strategies, regional pragmatic variant indicators, and a list of prohibited citations. The set of rhetorical strategies refers to a predefined set of expressions to ensure the readability of medical knowledge and the acceptability of the audience. Based on the aforementioned communication context, set of conceptual elements, factual boundaries, and acceptability metrics, an AI generator is driven to produce candidate expressions. During the generation process, the factual boundaries are forcibly satisfied, the expression level is selected according to the comprehension threshold, and strategies are switched between metaphor, example, analogy, and step-by-step narration. The expression level is generated based on the information elements of the target audience and the communication goals, decomposing the communication goals into mastery level indicators and expression tone indicators, and establishing an alignment mapping with the audience's language habits, reading device types, and preferences. The candidate expressions are subjected to factual consistency verification, terminology standardization consistency verification, and acceptability assessment. Only when all three are met is it determined to be an acceptable concept. If the three are not met, the expression level and narrative strategy are adjusted according to the assessment results. The controlled generation and verification are repeated within a preset number of steps until the stopping criterion is met. The acceptable concept is delivered in a way that appeals to the audience and in a length that is appropriate for their preferences. Explicit feedback and implicit interaction signals are collected and used to update the acceptability measurement and strategy selection parameters, thereby improving the matching degree of subsequent generated data without changing the factual boundaries.

2. The customized science popularization method based on artificial intelligence according to claim 1, characterized in that, Receive information elements and communication goals from the target audience, acquire the knowledge points to be disseminated and communication constraints, and form a communication context that includes language, length, style, restrictions on sensitive words, and risk thresholds, including: The platform rules, regional policy lists, and citation permission levels in the dissemination constraints are translated into an executable set of constraint rules, and a sensitive word blocking table and initial risk threshold settings are output; the set of constraint rules clearly defines prohibited items, permitted items, evidence level thresholds, and regional scope of application. Based on the terminology standard library and knowledge relationship network, the knowledge points to be disseminated are expanded and synonymized, and then unitized according to the tags of "definition, principle, application, clarification of misunderstanding" to obtain the smallest set of expression units; Based on the information elements and communication goals of the target audience, the communication goals are decomposed into mastery level instructions and expression tone instructions, and an alignment mapping is established with the audience's language habits, reading device types and preferences to generate expression level settings and style preference settings; Using the set of minimum expression units, the set of constraint rules, the setting of expression level, and the setting of language style preference as inputs, the parameters of the dissemination context are determined jointly. Specifically, the length is determined by the reading device type and the number of minimum expression units; the language style is determined by the language style preference setting and the expression tone indication; the risk threshold is determined by the evidence level threshold and the geographical scope of application in the constraint rule set; sensitive word restrictions are determined by the intersection of the blocking list and platform rules; and the context is determined by the dissemination channel and usage scenario indication. The context, length, style, sensitive word restrictions, and risk thresholds are encapsulated into a structured propagation context object.

3. The customized science popularization method based on artificial intelligence according to claim 1, characterized in that, The knowledge points are semantically parsed and their relationships expanded to generate a set of conceptual elements and dependencies. Fact boundaries are established based on an authoritative knowledge base and a terminology standardization database, including: The knowledge points to be disseminated are aligned with terms and aliases are merged. The conceptual element set is generated by dividing the knowledge points into "noun definition unit, relational predicate unit, quantitative attribute unit, condition unit and applicable domain unit". The dependency template is used to determine the master-slave, limiting and parallel dependency relationships between the conceptual elements. In the knowledge relationship network, the aforementioned conceptual elements are used as anchor points, and the relationship expansion is limited to four controllable relationship sets: "superior and subordinate, composition and part, condition and result, premise and conclusion". Each relationship is labeled with causal role, applicable domain and boundary conditions to form a labeled extended relationship set. The quantity attribute unit is mapped to a unified dimensional system, and the unit of the same physical quantity is standardized and converted to generate dimensional constraint pairs. The dimensional constraint pairs include at least a dimensional identifier, a set of allowed units, a numerical domain, and a precision indicator. Retrieve the normative definitions and evidence sources corresponding to the conceptual elements from the terminology standard library and authoritative knowledge base, extract unique definitions according to preset priorities, and classify and label the evidence according to the authority level of the source and the publication time to form a definition-evidence mapping table. For relationships marked as causal in the extended relation set, cross-constraints are applied with the definition-evidence mapping table, retaining only causal assertions that satisfy definition consistency and evidence level thresholds; when there are conflicts in definition, dimension, or causality, the rules of prioritizing the terminology standard library, evidence level, and timeliness are applied in sequence to adjudicate and fix the result; The adjudicated normative definitions, dimensional constraints, causal assertions, and evidentiary anchors are encapsulated as the fact boundary.

4. The customized science popularization method based on artificial intelligence according to claim 1, characterized in that, Based on the information elements of the target audience, a joint representation of cognitive level, reading habits, cultural context, and risk sensitivity is constructed to establish an acceptability metric, including: Based on the information elements of the target audience, the original elements corresponding to cognitive level, reading habits, cultural context and risk sensitivity are mapped to audience indicators, and fixed as a discrete set of indicators with a clear value range; the audience indicators include at least abstract tolerance indicators, vocabulary familiarity indicators, paragraph rhythm preference indicators, rhetorical strategy preference indicators, regional pragmatic variant indicators, taboo topic indicators and controversy avoidance indicators. Using the discrete indicator set as input, a quaternary joint representation including a cognitive level profile, a reading habit profile, a cultural context profile, and a risk sensitivity profile is generated. Based on the joint representation, an expression level interval, a set of allowed rhetorical strategies, and a list of prohibited citations are derived. The list of prohibited citations is compatible with the evidence anchors and dimensional constraints in the fact boundary. Using the aforementioned joint representation as a constraint source, an acceptability measurement object consisting of five components is established, and thresholds or acceptable ranges are set for each component. Among them, the accuracy threshold is a hard constraint on the object of the aforementioned fact boundary and cannot be violated; the consistency threshold is a hard constraint on the pair of terminology norms and dimensional constraints; the comprehension burden index is defined by the density of new terms, the upper limit of sentence length, and the upper limit of reasoning span; the ambiguity risk index is defined by the coverage of referential resolution and the sufficiency of polysemous word resolution; and the value fit index is defined by the matching degree of the rhetorical strategy set, the regional pragmatic variant indication, and the prohibited citation list. The thresholds, indicators, and their thresholds are solidified into executable objects of the acceptability measurement, and a set of conditional parameters for the generation phase is output, including a set of hard constraints and a target range of delimited quantities.

5. The customized science popularization method based on artificial intelligence according to claim 1, characterized in that, Based on the aforementioned propagation context, set of conceptual elements, factual boundaries, and acceptability metrics, an AI generator is driven to produce candidate representations. During the generation process, the factual boundaries are forcibly satisfied, the expression level is selected according to a comprehension threshold, and strategies are switched between metaphor, example, analogy, and step-by-step description, including: Using the object of the propagation context, the object of the fact boundary, the set of conceptual elements and their dependencies, and the object of the acceptability measurement as joint input, during the decoding process, terms and sentence structures that are inconsistent with the definition, causality, or dimensionality are intercepted or replaced, and content that references below the level of evidence anchors or triggers sensitive word restrictions is rewritten in a restrictive manner. Limited to the objects of the acceptability measurement, the density of new terms, sentence length and reasoning span of the generated fragments are measured online; when any indicator exceeds the target range, the expression level is automatically reduced, the scope of term substitution and the granularity of explanation are tightened; when the indicator is below the target range, the expression level is increased and the allowed rhetorical depth is opened, so that the indicator returns to the target range. Within the scope and body of the communication context, the choice is made between "exemplification and step-by-step narration" and "analogy and metaphor" based on the set of rhetorical strategies allowed by the acceptability measurement object; when ambiguity risks or regional pragmatic mismatches are triggered, local rewriting and replacement are carried out according to the evidence anchors and terminology norms of the fact boundary object. When the definition constraints, causal constraints, and dimensional constraints in the fact boundary object are not all satisfied, or the accuracy threshold and consistency threshold in the acceptability measurement object are not simultaneously satisfied, a local backtracking rewrite is performed on the relevant fragments along the most recent rhetorical selection point, without regenerating the compliant parts; after all constraints and thresholds are satisfied and the measurement indicators fall into the target range, the fragments are merged to form the candidate statements, and the correspondence between the rhetorical strategy sequence and the evidence anchor point is recorded.

6. The customized science popularization method based on artificial intelligence according to claim 1, characterized in that, The candidate expressions undergo factual consistency verification, terminology standardization consistency verification, and acceptability assessment. An expression is deemed acceptable only if all three criteria are met simultaneously. If none are met, the expression level and narrative strategy are adjusted based on the assessment results. Controlled generation and verification are repeated within a preset number of steps until a stopping criterion is met, including: a) Based on the object of the fact boundary, align the semantic fragments and conceptual elements in the candidate statements, and sequentially perform boundary predicate matching, evidence anchor level comparison and unit normalization comparison on statements involving definitions, causality and quantitative attributes; any statements that are inconsistent with the definition, causality or dimension, or whose citation level is lower than the evidence anchor threshold, are marked as non-compliant fragments and the start and end positions of the non-compliant fragments are located. b) Based on the terminology standard library and the set of conceptual elements and their dependencies, examine the consistency of term morphology, aliasing and master-detail dependency relationships in candidate expressions; for terms that have cross-level mismatch, term morphology drift or do not conform to dependency relationships, output corresponding standard substitution suggestions and maintain the original dependency relationships unchanged. c) Using the object of the acceptability measurement as the source of constraints, calculate the understanding burden index, ambiguity risk index and value fit index for candidate statements. Only when the accuracy threshold and consistency threshold are met and each index falls within its target range is it determined to be qualified. d) If any of steps a) to c) fails to meet the standard, perform local controlled rewriting on the candidate expression according to the type of failure. The rewriting is limited to adjusting the expression level and narrative strategy without changing the definition, causality and dimension in the fact boundary object, and keeping the terminology and dependency relationships unchanged. After rewriting, the relevant fragments will be generated in a controlled manner and returned for review. The number of iterations is limited by a preset number of steps. When all three types of verification and evaluation meet the standard at the same time or when the preset number of steps is reached but the standard is still not met, output the acceptable concept or failure flag and corresponding position annotation to end this process.

7. The customized science popularization method based on artificial intelligence according to claim 6, characterized in that, The comprehension burden indicators include neo-term density, sentence length, and inference span; the ambiguity risk indicators include referential resolution coverage and polysemous word resolution sufficiency; and the value fit indicators include rhetorical strategies and regional pragmatic matching degree.

8. The customized science popularization method based on artificial intelligence according to claim 1, characterized in that, Delivering the acceptable concept using media and length preferred by the audience, collecting explicit feedback and implicit interaction signals, and using these to update the acceptability measurement and strategy selection parameters, thereby improving the matching accuracy of subsequently generated content without altering the factual boundaries, including: Based on the object of the aforementioned dissemination context, without changing the meaning of the content and the anchor of evidence, the acceptable concepts are rearranged into delivery units consistent with the carrier type and length limitations, while retaining stylistic limitations and citation annotations for subsequent retrospective purposes; The explicit feedback and implicit interaction signals after delivery are processed into indicators to generate comprehension burden bias indicators, ambiguity risk indicators, and value adaptation bias indicators. The explicit feedback includes evaluation and error correction labels, and the implicit interaction signals include dwell and reread features and interaction trigger locations. Using the object of the acceptability measurement as the source of constraints, without changing the definition, causal and dimensional constraints in the object of the fact boundary, the target range and threshold of the acceptability measurement are fine-tuned according to three types of instructions, and the strategy selection parameters are adjusted simultaneously; the strategy selection parameters include expression level, interpretation granularity, term substitution range and rhetorical strategy weight, and are accompanied by version number and scope of application.

9. A customized science popularization and dissemination system based on artificial intelligence, characterized in that: For implementing the AI-based customized science popularization dissemination method as described in any one of claims 1 to 8, the AI-based customized science popularization dissemination system comprises: The audience and target analysis unit is used to receive information elements and communication goals from the target audience, obtain the knowledge points to be disseminated and communication constraints, and form a communication context that includes context, length, style, sensitive word restrictions and risk thresholds; The knowledge point semantic parsing unit is used to perform semantic parsing and relation expansion on the knowledge points, generate a set of conceptual elements and dependency relationships, and establish fact boundaries based on an authoritative knowledge base and terminology standard base; the fact boundaries define immutable definitions, dimensions, causal relationships and sources of evidence. The audience acceptability modeling unit is used to construct a joint representation of cognitive level, reading habits, cultural context and risk sensitivity based on the information elements of the target audience, and to establish an acceptability measure; the acceptability measure simultaneously constrains accuracy, consistency, comprehension burden, ambiguity risk and value fit; The controlled generation unit is used to drive the artificial intelligence generator to generate candidate expressions based on the propagation context, the set of conceptual elements, the factual boundaries, and the acceptability measure. During the generation process, the factual boundaries are forcibly satisfied, the expression level is selected according to the understanding threshold, and strategies are switched between metaphor, example and analogy and step-by-step description. The verification-evaluation unit is used to perform factual consistency verification, terminology standardization consistency verification, and acceptability evaluation on the candidate statements. It is determined to be an acceptable concept only when all three are met. If they are not met, the expression level and narrative strategy are adjusted according to the evaluation results. The controlled generation and verification are repeated within a preset number of steps until the stopping criterion is met. The delivery-feedback update unit is used to deliver the acceptable concept in a way that meets the audience's preferred format and length, collect explicit feedback and implicit interaction signals, and use them to update the acceptability measurement and strategy selection parameters, thereby improving the matching degree of subsequent generation without changing the factual boundaries.

Images