A method, system, and medium for intelligent generation of architectural decisions

Abstract

This invention provides an intelligent generation method, system, and medium for architectural decision-making, relating to the field of architectural engineering. The method includes: generating a corresponding construction context vector based on user-inputted problem description information, used to drive at least two agents in parallel to perform adversarial retrieval to obtain at least two sets of evidence; quantifying the credibility of the at least two sets of evidence to obtain a credibility evaluation value for each piece of evidence, used to control at least two agents to conduct multi-round interactive debates to dynamically update the validity status and credibility evaluation value of each piece of evidence; generating a ruling result through a preset ruling model; and generating and outputting decision guidance information adapted to the user's role based on the ruling result. The system is configured to execute this method. The method can also be executed after being read from program instructions stored in a computer-readable storage medium. This invention can more intelligently resolve technical disputes and complete intelligent architectural decision-making.

Description

Technical Field

[0001] This invention relates to the field of building engineering, and in particular to an intelligent method, system and medium for generating building decisions. Background Technology

[0002] In the field of construction engineering, there are often overlaps and even conflicts in content between construction specifications, design drawings, industry standards, and local standards. Furthermore, the applicability of the same clause in different specific scenarios is frequently disputed.

[0003] Currently, resolving such technical disputes mainly relies on manual review of numerous paper or electronic specifications, with decisions made through negotiation based on experience. Furthermore, due to the prevalence of technical disputes, such negotiations often lead to delays in the project timeline. In addition, question-answering systems based on Retrieval-Augmented Generation (RAG) are a primary intelligent approach. Their working mode is: user asks a question → system retrieves the most relevant document → large model generates a unique answer. However, these question-answering systems suffer from problems such as single-perspective bias, lack of validity determination, isolated evidence chains, and insufficient scenario adaptability.

[0004] Therefore, there is a need to provide an intelligent generation method, system, and medium for building decisions, which can more intelligently resolve technical disputes and complete intelligent building decisions. Summary of the Invention

[0005] To address the problem of how to intelligently generate adaptive architectural decisions, this invention provides an intelligent generation method, system, and medium for architectural decisions.

[0006] The invention includes an intelligent generation method for architectural decision-making. The method includes: generating a corresponding construction context vector based on user-inputted problem description information; driving at least two agents in parallel to perform adversarial retrieval based on the construction context vector to obtain at least two sets of evidence, wherein the at least two agents are configured to have dissimilar retrieval preferences; quantifying the credibility of the at least two sets of evidence to obtain a credibility evaluation value for each piece of evidence; controlling the at least two agents to conduct multiple rounds of interactive debate based on the credibility evaluation values ​​to dynamically update the validity status and credibility evaluation value of each piece of evidence; generating a ruling result through a preset ruling model based on each piece of evidence and its updated validity status and updated credibility evaluation value; and generating and outputting decision guidance information adapted to the user's role based on the ruling result.

[0007] The invention includes an intelligent generation system for architectural decision-making. The system comprises: a problem parsing module configured to generate a corresponding construction context vector based on user-inputted problem description information; an adversarial retrieval module configured to drive at least two agents in parallel to perform adversarial retrieval based on the construction context vector, thereby obtaining at least two sets of evidence, wherein the at least two agents are configured to have different retrieval preferences; a credibility quantification module configured to perform credibility quantification on the at least two sets of evidence, obtaining a credibility evaluation value for each piece of evidence in the at least two sets of evidence; an interactive debate module configured to control the at least two agents to conduct multiple rounds of interactive debate based on the credibility evaluation values, dynamically updating the validity status and credibility evaluation value of each piece of evidence over multiple rounds; an adjudication module configured to generate an adjudication result based on each piece of evidence and its updated validity status and updated credibility evaluation value through a preset adjudication model; and a decision-making module configured to generate and output decision guidance information adapted to the user's role based on the adjudication result.

[0008] The invention includes a computer-readable storage medium storing program instructions that, when executed by a processor, implement the intelligent generation method for building decisions as described above.

[0009] The beneficial effects of the above invention include, but are not limited to: (1) By using adversarial retrieval and multi-round interactive debate, the drawbacks of traditional decision-making evidence being one-sided and conclusions being singular are effectively overcome. This method can comprehensively evaluate multiple pieces of evidence and dynamically update their effectiveness, significantly improving the objectivity and robustness of decision-making. Combined with user roles to generate customized guidance, it provides operational and targeted professional advice, greatly improving the efficiency and accuracy of architectural decision-making and optimizing project management processes; (2) By performing structured analysis of user questions, it accurately extracts scene elements such as engineering parts and material types, transforming ambiguous natural language into structured construction context vectors. This method solves the problem that direct vectorization may lose key semantic information, providing a semantic basis for subsequent accurate retrieval. Construction context vectors constructed based on scene elements can more accurately represent user questions, significantly improve retrieval precision, and recall more matching evidence. At the same time, the identification process of scene elements is traceable, enhancing the transparency and credibility of the system; (3) By introducing state transition rules, the controllability and convergence of the debate process are realized. By dynamically updating the effectiveness status and credibility evaluation value of evidence, invalid loops are avoided, and the efficiency of debate is improved. The fact that changes in the validity of each piece of evidence are traceable and verifiable enhances the interpretability of the system and provides arbitrators with a clear and reliable basis for their decisions. Attached Figure Description

[0010] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0011] Figure 1 This is an exemplary block diagram of an intelligent generation system for building decisions, as shown in some embodiments of this specification;

[0012] Figure 2 This is an exemplary flowchart of an intelligent generation method for building decisions according to some embodiments of this specification;

[0013] Figure 3 This is an exemplary flowchart illustrating the generation of adjudication results according to some embodiments of this specification. Detailed Implementation

[0014] The accompanying drawings used in the description of the embodiments will be briefly introduced below. The drawings do not represent all embodiments.

[0015] It should be understood that the terms “system,” “device,” “unit,” and / or “module” used herein are one way to distinguish different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0016] Unless the context clearly indicates an exception, words such as "a," "an," "a kind," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0017] Figure 1 This is an exemplary block diagram of an intelligent generation system for building decisions, as shown in some embodiments of this specification.

[0018] In some embodiments, the intelligent generation system 100 for building decisions may include a problem analysis module 110, an adversarial retrieval module 120, a credibility quantification module 130, an interactive debate module 140, an adjudication module 150, and a decision module 160.

[0019] Problem parsing module 110 refers to a module used to parse problem description information. In some embodiments, problem parsing module 110 is configured to generate a corresponding construction context vector based on the problem description information input by the user.

[0020] The adversarial retrieval module 120 refers to a module used to perform adversarial retrieval. In some embodiments, the adversarial retrieval module 120 is configured to drive at least two agents to perform adversarial retrieval in parallel based on construction context vectors, so as to obtain at least two sets of evidence.

[0021] The credibility quantification module 130 refers to a module that performs credibility quantification on evidence. In some embodiments, the credibility quantification module 130 is configured to perform credibility quantification on at least two evidence sets and obtain a credibility evaluation value for each piece of evidence in at least two evidence sets.

[0022] The interactive debate module 140 refers to a module for performing multiple rounds of interactive debate. In some embodiments, the interactive debate module 140 is configured to control at least two agents to conduct multiple rounds of interactive debate based on credibility evaluation values, so as to dynamically update the validity status and credibility evaluation value of each piece of evidence in multiple rounds.

[0023] The adjudication module 150 refers to the module used to generate adjudication results. In some embodiments, the adjudication module 150 is configured to generate adjudication results based on each piece of evidence and its corresponding updated validity status and updated credibility evaluation value, using a preset adjudication model.

[0024] Decision module 160 refers to a module used to generate decision guidance information. In some embodiments, decision module 160 is configured to generate and output decision guidance information appropriate to the user role based on the decision result.

[0025] For more information on the configuration of the above modules, please refer to [link / reference]. Figures 2-3 Related descriptions.

[0026] In some embodiments, multiple modules in the intelligent building decision generation system 100 described above can be operated by a processor. In some embodiments, the processor may include a central processing unit (CPU), a digital signal processor (DSP), a microcontroller unit (MCU), a computer, or any combination thereof. In some embodiments, the processor may be centralized or distributed. In some embodiments, the processor may be local or remote. In some embodiments, the processor may be implemented on a cloud platform. By way of example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layered cloud, or any combination thereof.

[0027] It should be noted that the above description of the intelligent generation system for building decisions and its modules is for convenience only and should not be construed as limiting this specification to the scope of the embodiments described. It is understood that those skilled in the art, after understanding the principles of this system, may arbitrarily combine the various modules or construct subsystems connected to other modules without departing from these principles.

[0028] Figure 2 This is an exemplary flowchart of an intelligent generation method for building decisions according to some embodiments of this specification. Figure 2 As shown, process 200 includes the following steps. In some embodiments, process 200 may be implemented by an intelligent generation system for building decisions, specifically executed by a processor.

[0029] Step S210: Generate the corresponding construction context vector based on the problem description information input by the user.

[0030] Users refer to entities that raise questions and seek corresponding solutions. Examples include construction companies, supervisors, designers, and owners. The questions discussed in this manual primarily refer to issues related to architectural decision-making.

[0031] Problem description information refers to natural language input that describes the problem. For example, "What is the minimum thickness of the waterproof protective layer for the exterior wall of a basement?"

[0032] Construction context vectors are structural vectors that represent information describing a problem. Construction context vectors can be constructed in ways including, but not limited to, hybrid structures of "key-value pairs + embedded vectors," pure key-value pair structures, pure embedded vector structures, and combined graph structures.

[0033] In some embodiments, the processor can generate construction context vectors in various ways based on user-inputted problem description information. For example, the processor can directly generate construction context vectors based on user-inputted problem description information using a pre-trained architectural domain sentence vector model. Exemplarily, the pre-trained architectural domain sentence vector model can be a model fine-tuned for the architectural domain based on models such as Sentence-BERT (Bidirectional Encoder Representations from Transformers), SimCSE (Simple Contrastive Learning of Sentence Embeddings), and BGE (BAAI (Beijing Academy of Artificial Intelligence) General Embedding). The input to the pre-trained architectural domain sentence vector model is the problem description information, and the output is a construction context vector with a pure embedding vector structure.

[0034] In some embodiments, the processor can perform structured parsing of the construction scenario based on problem description information to obtain scenario elements; and generate a construction context vector based on the scenario elements.

[0035] Scene elements refer to key information or entities used to describe a construction scene. For example, scene elements may include the location of the project, the type of work process, the type of material, technical parameters, points of contention, and project background.

[0036] The engineering part refers to the specific construction location involved in the user's question, such as "basement exterior wall," "foundation slab," and "roof waterproofing layer." In some embodiments, the intelligent generation system for building decisions maintains a dictionary of approximately 2000 standard engineering part names, supporting the normalization of synonyms for engineering parts (e.g., "basement exterior wall" is normalized to "basement exterior wall"). The process type refers to the construction process, such as "reinforcement binding," "concrete pouring," and "waterproofing construction." The material type refers to the type of building material involved, such as "C30 concrete," "HRB400 steel reinforcement," and "SBS waterproof membrane." The material type can be used for subsequent standard applicability assessment. The technical parameters are the engineering-related parameters mentioned in the user's question, such as numerical or graded parameters like "protective layer thickness," "overlap length," and "seismic grade." The focus of the dispute is the core issue of concern to the user, such as "can it be substituted," "is it necessary," and "what is the allowable deviation?" The project background includes the project location (affecting the applicability of local standards), building type (civil / industrial), and seismic fortification intensity.

[0037] In some embodiments, the processor can perform structured analysis of the construction scene based on the problem description information and a pre-trained architectural named entity recognition model to obtain scene elements.

[0038] A pre-trained architecture-domain named entity recognition model refers to a model obtained by fine-tuning a pre-trained model for the architecture domain. For example, it can be a BERT / RoBERTa variant obtained by fine-tuning BERT and RoBERTa (Robustly Optimized BERT Pretraining Approach) for the architecture domain. The input of a pre-trained architecture-domain named entity recognition model is a problem description, and the output is scene elements.

[0039] In some embodiments, the structured analysis of the construction scene to obtain scene elements can also be achieved through other technologies (e.g., keyword extraction, text classification), which are not limited here.

[0040] In some embodiments, the processor can construct key-value pairs based on scene elements; based on the key-value pairs, it can generate embedding vectors through a pre-trained architectural domain sentence vector model, and then construct construction context vectors, which are represented as a hybrid structure of "key-value pairs + embedding vectors".

[0041] Key-value pairs refer to data pairs consisting of a "key" and a "value," used to store discrete scene elements. For example, scene elements may include engineering location, process type, seismic resistance level, and material type, with corresponding specific contents such as {"foundation slab," "basement exterior wall," "roof," "stairwell," ...}, {"earthwork excavation," "reinforcement binding," "concrete pouring," "waterproofing construction," ...}, {"Level 1," "Level 2," "Level 3," "Level 4"}, {"C30 concrete," "HRB400 steel reinforcement," ...}. For example, for the question description "What is the minimum thickness of the waterproofing protective layer for the basement exterior wall?", the name of the scene element ("engineering location," "process type") is used as the "key," and the corresponding specific contents ("basement exterior wall," "waterproofing construction") are used as the "value," forming the key-value pair {"engineering location": "basement exterior wall," "process type": "waterproofing construction"}. An embedding vector is a 512-dimensional dense vector encoding the question description information. The embedding vector can be used for subsequent semantic similarity retrieval.

[0042] For an explanation of the pre-trained architectural domain sentence vector model, please refer to the relevant description above.

[0043] Some embodiments in this specification employ structured parsing of user questions to accurately extract scene elements such as engineering locations and material types, transforming ambiguous natural language into structured construction context vectors. This method solves the problem of potential loss of key semantic information during direct vectorization, providing a semantic foundation for subsequent accurate retrieval. Construction context vectors constructed based on scene elements can more accurately represent user questions, significantly improving retrieval precision and recalling more relevant evidence. Furthermore, the scene element identification process is traceable, enhancing the system's transparency and credibility.

[0044] Step S220: Based on the construction context vector, at least two agents are driven in parallel to perform adversarial retrieval to obtain at least two sets of evidence, wherein the at least two agents are configured to have different retrieval preferences.

[0045] An intelligent agent is an intelligent model that possesses language understanding and data processing capabilities. For example, a software agent that integrates a large language model and algorithmic logic. Intelligent agents run on processors.

[0046] At least two agents are used to represent at least two different stakeholders (i.e., different user roles). A user role refers to the identity or functional classification a user plays in a construction project. For example, user roles include those of the construction party (e.g., project manager, construction worker, technician), the supervision party (e.g., chief supervisor, project supervisor), the design party (e.g., designer, reviewer), the owner, or other unknown roles. For example, at least two agents could be Agent A representing the supervision party and Agent B representing the construction party.

[0047] In some embodiments, at least two agents are configured to have different search preferences. Search preferences refer to the rules that agents follow during the search process to determine their inclination towards obtaining evidence.

[0048] In some embodiments, agents corresponding to different user roles are pre-set with different search objective functions and are configured with different search preferences. For example, if Agent A's search objective function is "maximizing compliance and security," then its search preference could be to find clauses with strict compliance. If Agent B's search objective function is "maximizing construction convenience and cost-effectiveness," then its search preference could be to find clauses with flexibility (allowing flexibility, exceptions, and alternatives).

[0049] Adversarial retrieval refers to a retrieval method that uses intelligent agents configured with different retrieval preferences to simulate opposing positions in order to retrieve evidence information from multiple perspectives and in a diverse manner.

[0050] An evidence set refers to the collection of relevant evidence provisions obtained by an agent after retrieval. Each agent has its own evidence set. For example, Agent A's evidence set includes all regulatory provisions retrieved from a "strict compliance" perspective, such as mandatory provisions emphasizing "must" or "must not"; Agent B's evidence set includes all regulatory provisions retrieved from a "flexible" perspective, such as recommended provisions containing "appropriate" or "permissible" elements, explanatory notes, local standards, or special programs.

[0051] In some embodiments, the processor may drive at least two agents in parallel to perform adversarial retrieval based on retrieval preferences, according to the construction context vector, so as to obtain at least two sets of evidence. For example, based on the construction context vector corresponding to the above question description information "What is the minimum thickness of the waterproof protective layer for the basement exterior wall?", the processor drives Agent A and Agent B to perform adversarial retrieval in parallel. According to the retrieval preference, Agent A is configured to prioritize the retrieval of the following types of documents: (1) National standards (GB) and clauses marked as "mandatory clauses"; (2) Construction details marked as "cannot be omitted" in general drawing sets; (3) The main text of the standard rather than "clause explanation" (because the clause explanation usually contains interpretive exemptions), while automatically filtering informal sources (such as technical blogs, manufacturer advertorials, and uncertified industry articles); Agent B is configured to prioritize the retrieval of the following types of documents: (1) The "clause explanation" part of the standard (finding room for interpretation and exemption conditions); (2) Local standards (DB) that are more lenient than GB; (3) "alternative practices" and "simplified construction" in industry drawing sets; (4) Project-level documents such as "technical approval form", "design change form", and "special construction plan".

[0052] Agent A retrieved "the mandatory provisions regarding the thickness of the protective layer in the national standard 'Technical Specification for Waterproofing of Underground Engineering' GB 50108-2008," while Agent B retrieved "the provisions in a local standard that allow the use of thin protective layers under specific conditions" or "the explanations in the relevant explanatory notes regarding flexible handling of special circumstances." After completing their searches, the two agents compiled the relevant evidence provisions they obtained to form an evidence set.

[0053] In some embodiments, the processor can configure a corresponding retrieval bias strategy for each of at least two agents, causing each agent to tend to retrieve evidence that matches its respective retrieval preferences. Dissimilar retrieval preferences include at least two of the following: a preference for retrieving evidence with the strictest compliance requirements, a preference for retrieving evidence with the best construction convenience, and a preference for retrieving evidence with the highest cost-effectiveness.

[0054] The most stringent evidence of compliance refers to evidence that sets the highest standards or mandatory requirements for aspects such as project quality, safety, or compliance with laws and regulations. Examples include clauses explicitly stating "strictly prohibited," "must not," or "mandatory," and evidence involving mandatory national standards and industry norms.

[0055] Evidence of optimal ease of construction refers to evidence that supports simplifying the construction process, reducing construction difficulty, or shortening the construction period. Examples include regulatory provisions that permit alternative construction techniques, simplify construction steps, or recommend easier-to-operate methods.

[0056] Evidence of maximum cost-effectiveness refers to evidence supporting the achievement of the lowest possible cost or highest economic benefit while meeting engineering requirements. Examples include allowing the use of lower-cost, qualified materials, recommending labor-saving construction techniques, and providing design options that reduce overall project expenditure.

[0057] Search bias strategies refer to mechanisms applied to the search process to make search results (i.e., evidence) more aligned with search preferences. Search bias strategies can be bias vectors or bias keywords.

[0058] Biased keywords are keywords that reflect search preferences, such as terms, phrases, or expressions. For example, for a search preference of "preferring to search for evidence with the strictest compliance requirements," biased keywords could include "strictly prohibited," "must not," or "required." For a search preference of "preferring to search for evidence with the best construction convenience," biased keywords could include "convenient," "rapid construction," "alternative process," "low technical requirements," "allowable deviation," "applicable," or "suitable." For a search preference of "preferring to search for evidence with the highest cost-effectiveness," biased keywords could include "economical," "saving," "cost-effectiveness," "optimized solution," or "alternative materials."

[0059] In some embodiments, bias keywords can be obtained in a variety of ways. For example, they can be manually extracted by domain experts, automatically extracted from a corpus, or generated in conjunction with bias vectors (e.g., derived from bias vectors).

[0060] A bias vector is a high-dimensional numerical vector that represents search preferences. Different bias vectors can be constructed based on different search preferences. For example, a compliance strictness bias vector is used to guide searches to "prefer to search for evidence with the strictest compliance requirements"; a cost-effectiveness bias vector is used to guide searches to "prefer to search for evidence with the highest cost-effectiveness"; and a construction convenience bias vector is used to guide searches to "prefer to search for evidence with the best construction convenience".

[0061] In some embodiments, bias vectors can be obtained in various ways. For example, based on keyword set encoding: a large number of bias keywords representing a type of search preference are input into a pre-trained semantic vector model (such as BERT, Sentence-BERT) to calculate its average vector, or the center vector is obtained through clustering methods as the bias vector of that search preference; based on supervised learning generation: a model is trained through labeled data so that it can directly generate the corresponding bias vector according to the user role (such as "supervisor" or "construction party").

[0062] In some embodiments, at least two agents, after configuring retrieval bias strategies, perform hybrid retrieval to obtain evidence that matches their respective retrieval preferences. The hybrid retrieval approach includes at least two of the following: semantic retrieval, keyword retrieval, and graph retrieval.

[0063] In semantic retrieval, the agent weightedly fuses the construction context vector with its configured bias vector to generate a query vector that reflects retrieval preferences, guiding the direction of semantic retrieval. The weight coefficients of the bias vector are adjustable fusion coefficients (e.g., 0.3) to control the strength of the preference. Using the query vector as input, the retrieval is performed using a fine-tuned vector model in the construction domain. For example, an approximate nearest neighbor search is performed in a vectorized database of code provisions, cosine similarity is calculated, and the top N provisions most semantically similar to the query vector are returned as the evidence set.

[0064] In keyword retrieval, the agent can expand the keywords in the construction context vector using the biased keywords configured by the agent, generating a keyword query set to drive keyword retrieval. Using the keyword query set as query input, a ranking algorithm such as BM25 is used in the full-text inverted index to retrieve the top N most relevant entries as the evidence set.

[0065] In graph retrieval, a pre-defined traversal strategy is used by the agent based on its search preferences within the validity knowledge graph. The preliminary relevant entries obtained from the first two steps serve as seed nodes. Starting from these seed nodes, and combining the traversal strategy, the agent selectively explores paths along predefined relation edges (such as "reference," "referenced," "deprecated," "conflict," and "substitution") within the validity knowledge graph. The depth, breadth, and selected relation types are all controlled by the traversal strategy, resulting in the first N entries visited in the traversal order, which form the evidence set. For an explanation of the validity knowledge graph, please refer to [link to relevant documentation]. Figure 3 And related explanations.

[0066] In some embodiments, the processor can calculate a comprehensive score for each piece of evidence from different retrieval methods using a fusion algorithm, and then sort the evidence according to the comprehensive score from high to low to obtain an evidence set. Evidence appearing earlier in the evidence set demonstrates stronger comprehensive performance across multiple dimensions and is considered more core and relevant. The fusion algorithm can be a reverse ranking fusion (RRF) algorithm.

[0067] Some embodiments in this specification configure retrieval bias strategies for agents, enabling each agent to efficiently acquire evidence that aligns with its individual retrieval preferences (such as compliance, ease of construction, and cost-effectiveness). This ensures that the adversarial retrieval process obtains a multi-dimensional and clearly defined set of evidence, providing comprehensive and targeted information support for dispute resolution and significantly improving the accuracy and efficiency of engineering dispute resolution.

[0068] Step S230: Measure the credibility of at least two evidence sets and obtain the credibility evaluation value of each piece of evidence in at least two evidence sets.

[0069] Credibility quantification refers to the process of assessing the quality of retrieved evidence and converting it into a numerical representation.

[0070] A credibility rating is a numerical value that reflects the degree of credibility of evidence. For example, a credibility rating can be represented by a value between 0 and 100.

[0071] In some embodiments, the processor can measure the credibility of at least two sets of evidence in various ways to obtain a credibility rating for each piece of evidence. For example, the processor can train a deep learning model (such as a pre-trained language model based on the Transformer architecture) by taking the text content of the evidence and the construction context vector as input, and having the model directly output the credibility rating of the evidence. This model needs to be trained under supervision using historical adjudication case data annotated by domain experts (including training samples: text data of the evidence and construction context vectors; training labels: credibility ratings annotated by experts) to learn the complex mapping relationship between the text content of the evidence, the construction context vector, and the credibility rating of the evidence.

[0072] In some embodiments, the processor can perform structured metadata extraction on at least two sets of evidence to obtain structured metadata for each piece of evidence; and, based on the structured metadata, obtain a credibility evaluation value.

[0073] Structured metadata refers to formatted data extracted from unstructured evidence. It has clear semantic categories and a fixed format, is easy for machines to process and analyze, and can provide direct and standardized input for subsequent quantitative assessments.

[0074] Structured metadata can include multiple dimensions. For example, the source information of the evidence (such as the name of the standard, the issuing organization, and the date of publication), the location of the evidence (such as the chapter number, clause number, page number, or figure number), the indication of validity (such as whether it is a mandatory provision), the type of wording (such as core verbs such as must, should, should, may, must not, should not, etc.), the conditions of application (such as pre-restrictive modifiers such as when, under the circumstances, for the structure), and the citation chain (such as which other standards or clauses the evidence cites).

[0075] In some embodiments, the processor can perform structured metadata extraction on each piece of evidence to obtain its structured metadata. The extraction methods may include any one or a combination of rule-based and pattern-based extraction (such as extraction using predefined regular expression pattern matching), semantic extraction based on natural language processing (such as extraction using natural language processing models - NLP recognition), and association queries based on external knowledge bases (such as normative information databases, lists of mandatory provisions, validity knowledge graphs, etc.).

[0076] For example, source information and citation chains can be extracted through a combination of rule-based and pattern-based extraction and association queries based on external knowledge bases; evidence location and validity labeling can be extracted through rule-based and pattern-based methods; wording types can be extracted through a combination of rule-based and pattern-based extraction and semantic extraction based on natural language processing; and applicable conditions can be extracted through semantic extraction based on natural language processing.

[0077] In some embodiments, the processor can obtain credibility evaluation values ​​in various ways based on structured metadata. For example, it can construct a six-dimensional evidence credibility assessment model to quantify and score each piece of evidence to obtain a credibility evaluation value.

[0078] The core components of the six-dimensional evidence credibility assessment model can include six assessment dimensions, quantitative scoring rules for each dimension, and a weighted summation formula. Model inputs can include structured metadata and construction context vectors. Model outputs can be credibility ratings and sub-scores for each assessment dimension.

[0079] In some embodiments, the six evaluation dimensions and the quantitative scoring rules under each evaluation dimension can be preset by domain experts. For example, the six evaluation dimensions can be: (1) Source authority score, the quantitative scoring rule of which can be to score according to the authority level of the evidence source, for example, a national mandatory standard (GB) can get 100 points, a national recommended standard (GB / T) can get 90 points, and an industry standard (JGJ / CJJ, etc.) can get 80 points, etc.; (2) Version timeliness score, the quantitative scoring rule of which can be to score according to the newness of the standard version, for example, if the standard is the current valid version, it can get 100 points, if it has been replaced by a new version and the new version has been implemented, it can get 0 points, etc.; (3) Scenario matching score, the quantitative scoring rule of which can be to compare and match the applicable conditions of the evidence (such as "applicable to seismic resistance level I and II") with the construction context vector, and score according to the matching results. For example, a perfect match can get 100 points, a partial match can get 80 points, etc.; (4) Semantic compulsion score, the quantitative scoring rule can be to score according to the compulsion of the word type, such as "must" is 100 points, "should" is 80 points, "appropriate" is 60 points, etc.; (5) Reference chain integrity score, the quantitative scoring rule can be to score according to whether the reference chain of the evidence is complete and valid, for example, if the reference chain is complete and all the referenced specifications are valid, it can get 100 points, if there is a "broken chain" it gets 60 points, if there is a "circular reference" it gets 40 points, etc.; (6) Condition satisfaction score, the quantitative scoring rule can be to evaluate whether the user scenario meets the preconditions of the evidence, if there are no preconditions or all preconditions are met it can get 100 points, if it is clearly not met it can get 0 points, etc.

[0080] In some embodiments, based on the quantitative scoring rules under each evaluation dimension, the sub-scores of the evidence in each evaluation dimension can be obtained. Then, according to the weighted summation formula, the sub-scores of the six evaluation dimensions are weighted and summed to obtain the credibility evaluation value. The weighting is based on the following: in building code disputes, "where it comes from" and "whether it is outdated" are the most critical judgment factors, meaning the source authority score and version timeliness score have the highest weight; secondly, "whether it is applicable to the current scenario," meaning the scenario matching score has the next highest weight; finally, detailed factors such as wording strength and citation relationships have the lowest weight, meaning the semantic forcefulness score, citation chain integrity score, and condition satisfaction score have the lowest weight. For example, the weight of the source authority score can be set to 0.25, the weight of the version timeliness score to 0.2, the weight of the scenario matching score to 0.2, the weight of the semantic forcefulness score to 0.15, the weight of the citation chain integrity score to 0.1, and the weight of the condition satisfaction score to 0.1.

[0081] The credibility rating will serve as a core weight in subsequent debates and rulings. When the evidence presented by two agents conflicts, the evidence with the higher credibility rating will have greater "weight."

[0082] Some embodiments in this specification extract structured metadata from evidence sets to obtain structured metadata for each piece of evidence. Based on this structured metadata, a credibility evaluation value is obtained, effectively overcoming the limitation of traditional Retrieval Enhancement Generation (RAG) systems that "only focus on recall, not credibility." This method can quantitatively evaluate the reliability of evidence from multiple dimensions, such as comprehensively considering the authority of the evidence's source, the timeliness of its version, and its relevance to the context, thereby significantly improving the accuracy and reliability of the provided evidence. This provides a more solid and persuasive basis for subsequent decision-making or dispute resolution, improving the overall quality of decision-making and user trust.

[0083] Step S240: Based on the credibility evaluation value, control at least two agents to conduct multiple rounds of interactive debate, so as to dynamically update the validity status and credibility evaluation value of each piece of evidence in multiple rounds.

[0084] Interactive debate refers to a multi-round, adversarial argumentation process in which at least two agents representing different positions engage in debate based on their respective sets of evidence. In each round, each agent has the opportunity to examine the evidence presented by the opposing agent and its credibility rating, and to respond strategically based on their own search preferences and strategies, including questioning, rebuttal, or defense. This process simulates in-depth debate among human experts, allowing the reliability, applicability, and logical completeness of evidence to be fully tested through multiple iterations, rather than simply presenting each agent's viewpoint in a single instance.

[0085] Validity status refers to an internal label indicating the degree of validity of the evidence after interactive debate. In some embodiments, validity status can be represented by a value between 0 and 1, with a higher value indicating greater validity.

[0086] In some embodiments, the processor can control at least two agents to conduct multi-round interactive debates in various ways based on credibility evaluation values. For example, the processor can control multi-round interactive debates between two agents through a hybrid architecture that combines rule-based state machines and language models. This hybrid architecture includes a rule-based control layer and a language model-based execution layer. The rule-based control layer predefines the debate process (such as rounds of argumentation, rebuttal, and counter-argumentation), the triggering conditions and rules for state transitions of evidence validity, and the criteria for determining debate convergence. The language model-based execution layer is responsible for performing specific debate tasks, including: understanding the opponent's arguments, identifying logical flaws, generating natural language content for questioning or defense, and retrieving supplementary evidence from its own evidence base or knowledge graph.

[0087] At the same time, after each update of the validity status, the credibility evaluation value of the corresponding evidence will also be adjusted according to preset rules.

[0088] In some embodiments, the processor may initialize the validity state of each piece of evidence to a valid state; and, in response to challenges raised by at least two agents in a multi-round interactive debate, update the validity state and credibility evaluation value of each piece of evidence based on state transition rules. The validity state may include a valid state, a pending state, and an invalid state.

[0089] A valid state refers to a state in which evidence is considered valid and usable. A valid state means that the evidence has not been effectively challenged or has successfully withstood challenge, and can be used as valid arguments in the final ruling.

[0090] In some embodiments, the validity status of each piece of evidence is initialized to a valid state.

[0091] A pending state refers to a situation where the validity of evidence is questioned but not completely invalidated; it is in an intermediate state where its validity is compromised. For example, if a piece of evidence is shown to only partially match the current scenario, that evidence is in a pending state.

[0092] An invalid status refers to a state where evidence has been deemed unusable. For example, evidence is invalid when the standard it cites has been repealed or proven to be completely inapplicable to the current scenario. Evidence deemed invalid is removed from the evidence set, its credibility rating is set to 0, and it no longer participates in subsequent debate.

[0093] Questioning refers to the act of an agent questioning, challenging, or attempting to overturn the validity of evidence presented by another party.

[0094] In some embodiments, the angles of challenge include, but are not limited to, challenges regarding the timeliness of the version (pointing out that the version of the specification cited in the opposing party's evidence is outdated or has been replaced by a newer version), challenges regarding the applicability of the scenario (pointing out that the applicable conditions of the opposing party's evidence do not fully match the construction context vector), challenges regarding the binding force of the wording (pointing out that the wording used in the opposing party's evidence is recommended rather than mandatory, and therefore lacks binding force), and challenges regarding the integrity of the citations (pointing out that the citation chain on which the opposing party's evidence depends is broken or that the citation source is invalid).

[0095] State transition rules are rules used to define the conditions for the transformation of evidence between different states of validity. In some embodiments, state transition rules include the methods for transitioning the state of validity under different angles of challenge and the methods for updating credibility evaluation values.

[0096] In some embodiments, state transition rules can be predefined based on domain expert knowledge and injected during system development, or calibrated and optimized using historical dispute cases. For example, state transition rules may include the following: when evidence is in a valid state and is subjected to a non-fatal attack (such as pointing out that the applicable conditions do not perfectly match the construction context vector, or that its wording is recommendation-based rather than mandatory), its validity state transitions from valid to pending, and the credibility evaluation value of the evidence is multiplied by a preset attenuation coefficient (e.g., 0.7); when evidence is in a valid state and is subjected to a fatal attack (such as proving that the standard it is based on has been repealed, or that the scenario is clearly inapplicable), its validity state transitions from valid to invalid, and the credibility evaluation value of the evidence becomes 0; when evidence is in a pending state and is subjected to another valid attack (e.g., the cumulative number of attacks reaches a preset threshold, such as 3 times), its validity state transitions from pending to invalid; when evidence is in a pending state and is subjected to a non-fatal attack (i.e., its holder successfully refutes it, such as proving that the other party's attack basis is incorrect), its validity state can be restored to valid, and the credibility evaluation value is also restored accordingly.

[0097] In some embodiments, in response to challenges raised by at least two agents in a multi-round interactive debate, the processor can determine the type of attack (non-lethal attack, lethal attack, etc.) on the evidence based on the challenges, and update the validity status and credibility evaluation value of each piece of evidence based on state transition rules.

[0098] In some embodiments, a convergence mechanism is provided in the multi-round interactive debate. The convergence mechanism is preset manually to ensure that the debate process is controllable and can be terminated. For example, the debate will automatically terminate when the validity of all evidence remains unchanged in two consecutive rounds of debate, or when all evidence has entered a stable valid or invalid state.

[0099] Some embodiments in this specification, by introducing state transition rules, achieve controllability and convergence in the debate process. By dynamically updating the validity status and credibility evaluation value of evidence, invalid cycles are avoided, and debate efficiency is improved. Changes in the validity status of each piece of evidence are traceable and verifiable, enhancing the interpretability of the system and providing arbitrators with clear and reliable grounds for their decisions.

[0100] Step S250: Based on each piece of evidence and its corresponding updated validity status and updated credibility evaluation value, generate a ruling result through a preset ruling model.

[0101] An adjudication result refers to a structured output that includes information such as the final decision conclusion, the reasons for the adjudication, and the confidence level of the result. For example, an adjudication result could be "Support option C", "Support option D", or "Conditionally support option A", along with the corresponding reasons.

[0102] A pre-defined adjudication model refers to a model used to generate an adjudication result. For example, a pre-defined adjudication model can be a quantitative, weight- and rule-based algorithmic model. The input of a pre-defined adjudication model includes each piece of evidence from at least two evidence sets that have undergone multiple rounds of interactive debate, along with their corresponding updated validity status and updated credibility ratings. The output includes the adjudication result.

[0103] The workflow of the pre-defined adjudication model includes the following steps. First, the pre-defined adjudication model obtains the updated credibility evaluation value for each piece of evidence. Next, for each set of evidence, based on the updated credibility evaluation value of each piece of evidence, the pre-defined adjudication model obtains the result confidence level of the corresponding solution for that set of evidence. If a solution has the highest result confidence level and exceeds the confidence level threshold, a ruling of "supporting the solution" is generated. If the result confidence levels of all solutions are below the confidence level threshold, a ruling of "conditionally supporting the solution" is generated for the solution with the highest result confidence level. The confidence level threshold can be set manually.

[0104] In some embodiments, in response to entering the adjudication phase, the processor determines cleaned evidence pairs based on each piece of evidence and its corresponding updated validity status; identifies the semantic coercivity coefficients of conflicting evidence in the cleaned evidence pairs based on a semantic disambiguation engine; and calculates the semantic adjudication score of the conflicting evidence based on its updated credibility evaluation value and semantic coercivity coefficients. For details of the above process, please refer to [link to relevant documentation]. Figure 3 And related explanations.

[0105] Step S260: Based on the ruling result, generate decision guidance information that is appropriate to the user's role and output it.

[0106] For an explanation of user roles, please refer to the relevant description in step S220.

[0107] In some embodiments, the processor can identify user roles in a variety of ways. For example, the processor can identify user roles by explicit statements in the conversation (such as "I am from the project team"), by analyzing the wording of questions (such as "Can we save money?" which tends to favor the construction team, and "Is it compliant?" which tends to favor the supervision team), by referring to historical conversation records or project configuration information of the user's organization.

[0108] Decision guidance information refers to personalized, professional advice that can be implemented immediately. In some embodiments, decision guidance information tailored to user roles can be provided in user-friendly natural language. Decision guidance information may include a conclusion, action steps, communication scripts, and risk warnings, and offers different emphases for different user roles. The conclusion directly and clearly answers the user's core question, such as "Support / Do not support / Conditionally support". If the conclusion is "Support" or "Conditionally support", the action steps are the steps for "How to comply with regulations"; if the conclusion is "Do not support", the action steps are the compliance path for "How to obtain an exemption".

[0109] In some embodiments, the processor can generate and output decision guidance information adapted to the user's role based on the adjudication result in various ways. For example, the adjudication result can be structured using a pre-trained language model to generate decision guidance information.

[0110] In some embodiments, the processor may identify user roles; select corresponding script templates from a preset script template library based on user roles; and fill the script templates with the adjudication results to generate decision guidance information.

[0111] A pre-defined script template library refers to a collection of script templates corresponding to various user roles and / or user intentions. A script template is a predefined text structure or framework. It contains communication scripts and / or action suggestions for communicating with user roles, and can be used to generate complete decision guidance information by filling in the decision results.

[0112] User intent refers to a user's intrinsic purpose, specific goal they hope to achieve, or problem they need to solve. User intent can include seeking exemptions, seeking compliance confirmation, seeking communication strategies, seeking risk assessments, or seeking optimal solutions. Seeking exemptions means that the user wants to find evidence to avoid complying with a strict requirement.

[0113] In some embodiments, the processor can retrieve a corresponding script template from a preset script template library based on the identified user role and / or user intent, and use it to populate the adjudication result.

[0114] In some embodiments, the script template includes one or more placeholders. The processor automatically fills in the placeholders in the selected script template based on the final decision conclusion, reasoning, confidence level of the result, and main supporting evidence contained in the adjudication result, in order to generate and output decision guidance information.

[0115] Some embodiments in this specification achieve role-adaptive generation of decision guidance information by identifying user roles, selecting matching script templates, and filling in the adjudication results. This method can provide customized communication scripts and action suggestions based on the specific needs of different user roles, solving the implementation difficulties of traditional "one-size-fits-all" suggestions. This method significantly improves the practicality and relevance of decision information, lowers the threshold for users to understand and implement adjudication results, and effectively enhances practical value and user experience.

[0116] This specification provides an intelligent method for generating architectural decisions through several embodiments. By employing adversarial retrieval and multi-round interactive debate, it effectively overcomes the shortcomings of traditional decision-making methods, which often suffer from limited evidence and singular conclusions. This method comprehensively evaluates diverse evidence and dynamically updates its validity, significantly improving the objectivity and robustness of decisions. Combined with user role-based generation of customized guidance, it provides actionable and targeted professional advice, greatly improving the efficiency and accuracy of architectural decisions and optimizing project management processes.

[0117] Figure 3 This is an exemplary schematic diagram illustrating the calculation of semantic decision scores according to some embodiments of this specification.

[0118] In some embodiments, such as Figure 3 As shown, process 300 includes: in response to entering the adjudication stage, the processor can determine the cleaned evidence pair 320 based on each piece of evidence and its corresponding updated validity status 310; identify the semantic coercivity coefficient 340 of conflicting evidence in the cleaned evidence pair based on the semantic disambiguation engine 330; and calculate the semantic adjudication score 360 ​​of the conflicting evidence based on the updated credibility evaluation value 350 and the semantic coercivity coefficient 340 of the conflicting evidence.

[0119] The adjudication phase refers to the stage where the final adjudication result is generated after multiple rounds of interactive debate. In some embodiments, the processor determines whether the debate has ended and proceeds to the adjudication phase through a convergence mechanism. For a description of the convergence mechanism, please refer to [link to documentation]. Figure 2 The relevant description of step S240.

[0120] Cleaned evidence pairs refer to the structured final set of evidence used to support different positions or viewpoints after processing and organization. In some embodiments, cleaned evidence pairs may include a pair of valid evidence lists (including valid evidence lists from two agents), a pair of invalid evidence lists (including invalid evidence lists from two agents), attack summary records, and focus of disputed issues.

[0121] The list of valid evidence includes all evidence in a valid or pending state, along with the initial and updated credibility ratings for each piece of evidence. The list of invalid evidence includes all evidence in an invalid state, along with the reasons for its invalidation. The attack summary record is a summary of key attack points during the debate, used to explain the debate process to the user in the final decision. For example, a summary of key attack points that caused a change in the status of evidence (especially from valid to pending or invalid). The focus of the dispute refers to the true core points of disagreement between the two sides after the debate.

[0122] In some embodiments, once the adjudication phase begins, the processor can traverse each piece of evidence involved in the debate and classify and structure it according to its final updated validity status to generate cleaned evidence pairs.

[0123] A semantic disambiguation engine is a software module used to analyze and identify the specific semantics, level of force, etc., of the wording in the text content of evidence.

[0124] In some embodiments, the semantic disambiguation engine can identify the level of enforceability of clause wording through steps such as core verb extraction, wording pattern matching, context correction, and overlay of mandatory clause markers.

[0125] Conflicting evidence refers to evidence that shows inconsistencies or contradictions regarding the same technical specifications or requirements. For example, if the evidence provided by the supervisor and the contractor comes from national and local standards respectively, then this evidence can be considered conflicting evidence.

[0126] The semantic coercivity coefficient refers to the numerical value used to quantify semantic coercivity.

[0127] In some embodiments, the processor may utilize a semantic disambiguation engine to identify the semantic coercivity coefficient of conflicting evidence in the cleaned evidence pair through the following steps: extracting core verbs from the conflicting evidence, such as "should not be less than"; determining the coercivity level of the core verbs based on a word pattern library, which is manually preset and includes five coercivity levels corresponding to the core verbs; correcting the coercivity level, including correction according to context, such as setting the coercivity level of a core verb with a lower coercivity level directly to 5 if it is followed by a national mandatory standard, or correcting it according to preset rules, such as setting the coercivity level of the core verb directly to 5 if the conflicting evidence originates from the "Mandatory Clauses of Engineering Construction Standards"; and determining the semantic coercivity coefficient of the conflicting evidence based on the coercivity level of the core verbs.

[0128] The semantic adjudication score refers to the final quantitative evaluation score assigned to evidence after comprehensively considering its semantic enforceability and other relevant factors.

[0129] In some embodiments, the processor may determine the semantic adjudication score of the conflicting evidence by multiplying the updated credibility evaluation value of the conflicting evidence with its corresponding semantic coercivity coefficient.

[0130] In some embodiments, the processor can sum the updated credibility evaluation values ​​of all non-conflicting evidence supporting a certain scheme and the semantic adjudication scores of conflicting data to obtain the total score of the scheme; based on the total scores of different schemes, the preset adjudication model outputs the final adjudication result.

[0131] Some embodiments in this specification use the validity status of evidence after debate to determine cleaned evidence pairs, and utilize a semantic disambiguation engine to identify the semantic coercivity coefficients of conflicting evidence. Finally, a semantic adjudication score is calculated by combining the credibility evaluation value and the semantic coercivity coefficients. This scheme effectively overcomes the limitations of traditional adjudication that relies solely on the quantity of evidence or a single credibility evaluation value, making the adjudication results more accurate and objective, and significantly improving the accuracy and persuasiveness of decisions in complex dispute scenarios.

[0132] In some embodiments, such as Figure 3 As shown, based on the validity knowledge graph 370, the processor can obtain the validity coverage relationship 380 and the graph path coefficient 390 of the conflicting evidence; and, based on the validity coverage relationship 380, the graph path coefficient 390 and the semantic adjudication score 360, determine the weighted adjudication score 3010 of the conflicting evidence.

[0133] A validity knowledge graph is a knowledge graph used to represent the validity relationships between different normative sources or normative clauses, and to support querying and reasoning about these relationships. A validity knowledge graph can adopt a directed graph structure, including nodes and edges connecting the nodes.

[0134] Node types include specification nodes, clause nodes, and project document nodes. Specification nodes include attributes such as specification number, specification name, issuing authority, publication date, implementation date, repeal date (if repealed), level of validity (e.g., nationally mandatory, nationally recommended, industry-specific, local, enterprise-specific), and relevant field (e.g., structure, architecture, water supply and drainage, HVAC). Clause nodes include attributes such as clause number, clause text, wording type, applicable conditions, and level of mandatory enforcement. Project document nodes include attributes such as document name, approval date, and scope of application.

[0135] The edge types in the validity knowledge graph include "Overwrites," "Updates," "References," "Conflicts," and "SpecificTo." "Overwrites" indicates a relationship where a higher-level law overrides a lower-level law. The edge direction points from higher validity to lower validity. The rules for determining validity can include national mandatory standards > national recommended standards > industry standards > local standards > enterprise standards. "Updates" indicates a relationship where a new version of a specification replaces an old version. The edge direction points from the new version to the old version. The processor can automatically compare the prefixes and implementation dates of the two specifications to make the determination. "References" indicates a relationship where one clause references another clause. The edge direction points from the referrer to the referenced. "Conflicts" indicates a relationship where two clauses have inconsistencies on the same technical indicators. These are undirected edges, labeled manually or semi-automatically. "SpecificTo" indicates a relationship where a specific document takes precedence over a general specification within its scope of application. The edge direction points from the specific document to the general specification, with an attached scope attribute. In some embodiments, the validity knowledge graph can be constructed by staff based on publicly available laws and regulations, normative texts, internal corporate documents, and the knowledge of domain experts.

[0136] A validity overwrite relationship refers to a relationship where validity is prioritized over other validity nodes. In a validity knowledge graph, validity overwrite relationships are represented by "Overwrites" edges, which point from high-efficiency nodes to low-efficiency nodes. In some embodiments, validity overwrite relationships can be used to determine source validity coefficients, where the source validity coefficient of a high-efficiency node is greater than that of a low-efficiency node.

[0137] In some embodiments, the processor can obtain the validity coverage relationship of conflicting evidence based on a validity knowledge graph. For example, when two conflicting pieces of evidence are received (e.g., evidence X comes from a national standard, and evidence Y comes from a local standard), the processor performs a path search along the "Overwrites" relationship edges in the validity knowledge graph to determine the validity coverage relationship between evidence X and evidence Y. For example, if a path exists from the national standard to the local standard, then the validity of evidence X is determined to be higher than that of evidence Y, and the source validity coefficient of evidence X is greater than that of evidence Y (e.g., 0.8 and 0.2 respectively); if no path exists from the national standard to the local standard, the processor can further determine the validity coverage relationship based on the hierarchy of the conflicting evidence in the validity knowledge graph, with higher levels indicating higher validity and larger source validity coefficients.

[0138] The graph path coefficient is a quantitative coefficient derived from the completeness or validity of the citation chain of evidence in the knowledge graph of validity.

[0139] In some embodiments, the processor can obtain the graph path coefficient of conflicting evidence based on a validity knowledge graph. For example, conflicting evidence corresponds to a clause node, and the processor can trace its reference chain along the "References" edge. If the reference chain is complete and all nodes in it are valid, the graph path coefficient can be set to 1. If there are deprecated nodes or loops in the reference chain, the processor will generate a "broken chain warning" or a "circular reference warning" and set the graph path coefficient to 0.7 or 0.5 to reflect the validity of the reference chain. Deprecated nodes can be determined by the processor checking for updated versions along the Updates edge. If an updated version exists and has been implemented, the original node is marked as a deprecated node.

[0140] The weighted adjudication score is a quantitative evaluation value obtained by comprehensively considering multiple factors such as the validity coverage relationship of conflicting evidence, the graph path coefficient, and the semantic adjudication score, and then calculating them in a weighted manner.

[0141] In some embodiments, the weighted adjudication score of conflicting evidence can be implemented using an adjudication calculation formula. An exemplary adjudication calculation formula is as follows: Weighted adjudication score = Source validity coefficient × Semantic adjudication score × Graph path coefficient.

[0142] In some embodiments, generating a ruling also involves applying a conflict resolution decision tree. For example, for conflicting evidence between Agent A and Agent B regarding the same technical indicator, the processor can adjudicate according to the following decision tree: First, determine whether there is a validity coverage relationship; if so, the party with higher validity wins; if not, determine whether only one party has a mandatory provision; if so, the party with the mandatory provision wins. If neither of the above can be adjudicated, compare the semantic force coefficients and scenario matching degrees of the conflicting evidence from both parties. Finally, if still unable to adjudicate, compare the weighted adjudication scores of the conflicting evidence from both parties; the party with the higher score wins. If the score difference is still small (e.g., <10%), it is determined as "cannot be automatically adjudicated," and "conditionally feasible" is output along with a list of the conditions for both parties. Scenario matching degree refers to the degree of matching between the scenario and scenario elements in the text of the evidence. Scenario matching degree can be determined based on a large language model. For an explanation of scenario elements, see [link to relevant documentation]. Figure 1 And its related descriptions.

[0143] In some embodiments, the processor can sum the updated credibility scores of all non-conflicting evidence supporting a particular scheme and the weighted adjudication scores of the conflicting data to obtain the scheme's total score. Based on the total scores of different schemes, a preset adjudication model outputs the final adjudication result.

[0144] Some embodiments in this specification, by constructing and utilizing a knowledge graph of validity, can efficiently and accurately obtain the validity coverage relationship and citation chain reliability between conflicting evidence. Combined with semantic analysis, this quantifies the enforceability of evidence, thereby generating a comprehensive, multi-dimensional weighted adjudication score. This significantly improves the objectivity and accuracy of adjudication, reduces the burden of manual review, and, especially when dealing with complex normative conflicts, can more effectively assist in adjudication decisions.

[0145] This specification provides one or more embodiments of a computer-readable storage medium storing program instructions that, when executed by a processor, implement the intelligent generation method for building decisions as described above.

[0146] The basic concepts have been described above. It is clear that the detailed disclosure above is merely illustrative and does not constitute a limitation of the present invention. Although not explicitly stated herein, various modifications, improvements, and corrections may be made to the present invention by those skilled in the art. Such modifications, improvements, and corrections are suggested in this invention and therefore remain within the spirit and scope of the exemplary embodiments of the present invention.

[0147] Certain features, structures, or characteristics in one or more embodiments of the present invention may be appropriately combined.

[0148] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this invention are not intended to limit the order of the inventive processes and methods. Although the foregoing disclosure has discussed some currently considered useful embodiments of the invention through various examples, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the spirit and scope of the embodiments of the invention. For example, while the system components described above can be implemented using hardware devices, they can also be implemented solely through software solutions, such as installing the described system on existing servers or mobile devices.

Claims

1. A method for intelligently generating architectural decisions, characterized in that, include: Based on the user's input of the problem description information, a corresponding construction context vector is generated; Based on the construction context vector, at least two agents are driven in parallel to perform adversarial retrieval in order to obtain at least two sets of evidence, wherein the at least two agents are configured to have different retrieval preferences. The credibility of the at least two sets of evidence is quantified to obtain the credibility evaluation value of each piece of evidence in the at least two sets of evidence. Based on the credibility evaluation value, control the at least two agents to conduct multiple rounds of interactive debate, so as to dynamically update the validity status of each piece of evidence and the credibility evaluation value in multiple rounds. In response to entering the adjudication stage, based on each piece of evidence and its corresponding updated validity status, a cleaned evidence pair is determined. The cleaned evidence pair is a structured evidence pair that has been processed and organized to support different positions. Based on the semantic disambiguation engine, the semantic coercivity coefficient of conflicting evidence in the cleaned evidence pair is identified; Based on the updated credibility evaluation value of the conflicting evidence and the semantic coercivity coefficient, the semantic adjudication score of the conflicting evidence is calculated. Based on the validity knowledge graph, the validity coverage relationship and graph path coefficient of the conflicting evidence are obtained. The validity knowledge graph is a knowledge graph used to characterize the validity relationship between different normative sources or normative clauses. The validity coverage relationship is a relationship of coverage according to the degree of validity. The graph path coefficient is a quantitative coefficient derived from the completeness or validity of the chain of reference of the evidence in the validity knowledge graph. Based on the effectiveness coverage relationship, the graph path coefficient, and the semantic adjudication score, the weighted adjudication score of the conflicting evidence is determined; Based on the weighted adjudication score, an adjudication result is generated using a preset adjudication model; as well as, Based on the ruling, decision guidance information adapted to the user's role is generated and output.

2. The method according to claim 1, characterized in that, The generation of the corresponding construction context vector based on the user-inputted problem description information includes: Based on the problem description information, the construction scenario is structured and analyzed to obtain scenario elements; and, The construction context vector is generated based on the scene elements.

3. The method according to claim 1, characterized in that, The parallel driving of at least two agents to perform adversarial retrieval includes: Configure a corresponding retrieval bias strategy for each of the at least two agents, so that each agent tends to retrieve evidence that matches its own retrieval preferences; The different search preferences include at least two of the following: a preference for searching evidence with the strictest compliance requirements, a preference for searching evidence with the best construction convenience, and a preference for searching evidence with the highest cost-effectiveness.

4. The method according to claim 1, characterized in that, The step of quantifying the credibility of the at least two sets of evidence to obtain a credibility evaluation value for each piece of evidence in the at least two sets of evidence includes: Metadata structure extraction is performed on the at least two sets of evidence to obtain structured metadata for each piece of evidence; and, The credibility evaluation value is obtained based on the structured metadata.

5. The method according to claim 1, characterized in that, The step of controlling at least two agents to conduct multi-round interactive debates based on the credibility evaluation value, and dynamically updating the validity status of each piece of evidence and the credibility evaluation value over multiple rounds, includes: Initialize the validity state of each piece of evidence to a valid state; and, In response to challenges raised by the at least two agents during the multi-round interactive debate, the validity state and credibility evaluation value of each piece of evidence are updated based on state transition rules. The validity state includes the valid state, the pending state, and the invalid state.

6. The method according to claim 1, characterized in that, The step of generating decision guidance information adapted to the user's role based on the ruling result includes: Identify the user role; Based on the user role, select the corresponding script template from the preset script template library; and, The ruling result is filled into the script template to generate the decision guidance information.

7. An intelligent generation system for architectural decision-making, characterized in that, include: The problem parsing module is configured to generate a corresponding construction context vector based on the problem description information input by the user; The adversarial retrieval module is configured to drive at least two agents to perform adversarial retrieval in parallel based on the construction context vector, so as to obtain at least two sets of evidence, wherein the at least two agents are configured to have different retrieval preferences. The credibility quantification module is configured to perform credibility quantification on the at least two evidence sets and obtain a credibility evaluation value for each piece of evidence in the at least two evidence sets. An interactive debate module is configured to control at least two agents to conduct multiple rounds of interactive debate based on the credibility evaluation value, so as to dynamically update the validity status of each piece of evidence and the credibility evaluation value in multiple rounds. The adjudication module is configured as follows: In response to entering the adjudication stage, based on each piece of evidence and its corresponding updated validity status, a cleaned evidence pair is determined. The cleaned evidence pair is a structured evidence pair that has been processed and organized to support different positions. Based on the semantic disambiguation engine, the semantic coercivity coefficient of conflicting evidence in the cleaned evidence pair is identified; Based on the updated credibility evaluation value of the conflicting evidence and the semantic coercivity coefficient, the semantic adjudication score of the conflicting evidence is calculated. Based on the validity knowledge graph, the validity coverage relationship and graph path coefficient of the conflicting evidence are obtained. The validity knowledge graph is a knowledge graph used to characterize the validity relationship between different normative sources or normative clauses. The validity coverage relationship is a relationship of coverage according to the degree of validity. The graph path coefficient is a quantitative coefficient derived from the completeness or validity of the chain of reference of the evidence in the validity knowledge graph. Based on the effectiveness coverage relationship, the graph path coefficient, and the semantic adjudication score, the weighted adjudication score of the conflicting evidence is determined; as well as, Based on the weighted adjudication score, an adjudication result is generated using a preset adjudication model; and... The decision-making module is configured to generate and output decision guidance information appropriate to the user's role based on the ruling result.

8. A computer-readable storage medium, characterized in that, The system stores program instructions, which, when executed by a processor, implement the intelligent generation method for building decisions as described in claim 1.

Images