Architectural design element question answering method and system based on large language model technology

Abstract

The invention relates to the technical field of building design, in particular to a building design element question and answer method and system based on a large language model technology, and the method comprises the steps: collecting a user question text and a related design element database for a building design scene; obtaining related design element entries including the core keywords and redundant entries not including the core keywords, and determining an effective design element set according to the redundant entries and corresponding filtering coefficients; subtracting the effective design element set from the related design element database to obtain a supplementary element library; and filtering the supplementary element library according to an element association relationship in the question analysis result, and extracting an initial element matching set from the related design element database. According to the method, the user question text is quickly and accurately analyzed and matched, so that the time for a designer to search for related design elements can be greatly shortened, the system can automatically filter and give effective design elements, and the working efficiency is remarkably improved.

Description

Technical Field

[0001] This application relates to the field of architectural design technology, and in particular to a question-and-answer method and system for architectural design elements based on large language model technology. Background Technology

[0002] Architectural design is the process of creating buildings, structures, spaces, and environments that aim to meet functional, aesthetic, technical, and safety requirements. It is not merely an artistic creation process, but also involves engineering technology, environmental protection, social needs, economic costs, and many other factors.

[0003] Currently, with the development of the architectural design field, designers face increasingly diverse needs, often including numerous complex functional requirements, regulatory requirements, and environmental adaptability factors. Addressing these needs requires designers to sift through, analyze, and integrate a large amount of information, making the traditional manual design process cumbersome and inefficient. Furthermore, in traditional design workflows, recommendations for design elements are often generic and fail to accurately match the personalized needs of each project. Designers need to spend considerable time adjusting and optimizing to meet the unique requirements of specific clients or projects.

[0004] Furthermore, the design element library in traditional design methods is updated slowly and often fails to reflect the latest industry standards or technological advancements in a timely manner. Designers need to manually adjust design schemes according to the latest industry trends and standards, but such adjustments are often limited by information lag and manual processing. Summary of the Invention

[0005] The main purpose of this application is to provide a question-and-answer method and system for architectural design elements based on large language model technology, in order to solve the problem that the recommendations of design elements in the existing technology are often too general and difficult to accurately match the personalized needs of each project.

[0006] To achieve the above objectives, this application provides the following technical solution: A question-answering method for architectural design elements based on large language model technology includes: Collect user question texts and related design element databases for architectural design scenarios; The core keywords, element relationships, and intent tendencies in the user's question text are analyzed according to the preset semantic parsing model to obtain the question parsing result; Obtain relevant design element entries including the core keywords and redundant entries excluding the core keywords, and determine the effective design element set based on the redundant entries and the corresponding filtering coefficients; Subtract the set of valid design elements from the database of relevant design elements to obtain a supplementary element library; filter the supplementary element library according to the element relationships in the query parsing results, and extract an initial element matching set from the database of relevant design elements; Based on the intent tendency in the query analysis results, a first matching degree weighting is applied to each element entry in the relevant design element database; the matching degree weighted element entries are then processed based on the core keywords to obtain a weighted element matching set; A target element matching set is generated based on the initial element matching set and the weighted element matching set; the target element matching set is then processed by natural language generation and associated with the user's question text to obtain the architectural design element question and answer results.

[0007] Preferably, a database of user-generated questions and related design elements targeting architectural design scenarios is collected, including: Multi-source raw text data is obtained synchronously through the architectural design community interaction interface, the feedback module embedded in the design tool, and the interface for historical project consultation records. The multi-source raw text data is segmented and denoised to identify user interaction fragments and design specification fragments including architectural terms, resulting in classified preprocessed text. For the preprocessed text of the classification, user interaction fragments containing question marks are extracted as candidate question texts, and design specification clause fragments are extracted to construct an initial element item pool; The candidate question texts are compared with the word vector library in the field of architecture to select texts that match the characteristics of the design scenario as user question texts; Based on the initial element entry pool, structured element units are obtained by associating design specification numbers with element attribute tags using a knowledge graph; The structured element units are clustered according to the building type dimension, and the clustering results are optimized by combining the high-frequency domain words in the user question text to obtain a relevant design element database.

[0008] Preferably, the core keywords, element relationships, and intent tendencies in the user's question text are analyzed according to a preset semantic parsing model to obtain the question parsing result, including: The user's question text is processed to complete the context, and the omitted expressions are filled in by combining the knowledge base of architectural design to obtain the complete semantic text; The complete semantic text is input into the terminology recognition module of the preset semantic parsing model, and the core keywords with design element attributes are identified through domain-specific word vector matching to obtain a keyword sequence; Based on the keyword sequence, a word association graph is constructed using the dependency parsing unit of the semantic parsing model to mark the design logic relationships and obtain the element association network; Based on the element association network, interrogative adverbs, modal words, and sentiment words are extracted from the keyword sequence to obtain an intent feature vector; The keyword sequence, the element association network, and the intent feature vector are fused and encoded to obtain the question parsing result.

[0009] Preferably, relevant design element entries including the core keyword and redundant entries excluding the core keyword are obtained. A valid set of design elements is determined based on the redundant entries and their corresponding filtering coefficients, including: Core keywords are extracted from the query analysis results, and a keyword feature matrix is ​​constructed; where the matrix elements represent the correlation strength of keywords in design element items. Perform matrix matching operation on the keyword feature matrix and the related design element database, and filter out entries whose matrix element values ​​are greater than a preset threshold as related design element entries including core keywords; The remaining entries in the relevant design element database that were not selected as relevant design element entries are marked as redundant entries that do not include core keywords; A secondary evaluation of the domain relevance of the redundant entries is performed, and a corresponding filtering coefficient is assigned to each redundant entry in combination with the implicit requirements of the architectural design scenario; wherein, the filtering coefficient is inversely proportional to the potential value of the entry; Redundant entries are filtered hierarchically based on the filtering coefficient. Redundant entries with filtering coefficients below the threshold are retained and merged with the relevant design element entries including the core keywords to obtain a set of effective design elements.

[0010] Preferably, the relevant design element database is subtracted from the set of valid design elements to obtain a supplementary element library, including: All element entries in the relevant design element database are uniquely identified and encoded to obtain a complete set of coded element entries. Encode each element entry in the set of valid design elements using unique identifiers according to the same rules to obtain a set of coded valid element entries. The encoded full set of element entries and the encoded valid set of element entries are compared and the differences in encoding are marked in the full set of element entries that do not appear in the valid set of element entries. The initial difference element pool is obtained by matching the original element content in the relevant design element database with the difference coding entries. The initial difference element pool is subjected to design scenario adaptability verification, and element entries that are completely irrelevant to the current architectural design scenario are removed to obtain a supplementary element library.

[0011] Preferably, the supplementary element library is filtered based on the element relationships in the query parsing results, and an initial element matching set is extracted from the relevant design element database, including: Extract the element relationships from the query parsing results and construct a relationship network; wherein, the relationship network includes causal chains, constraint pairs and cooperative groups, and assign scenario-adaptive weights to each relationship type; The element entries in the supplementary element library are decomposed into triples to obtain the element association map; Calculate the topological similarity between the three-dimensional relationship network and the element relationship map, and select element entries with similarity higher than a set threshold as candidate related elements; Based on the strength of the element relationships in the query analysis results, the candidate related elements are prioritized and ranked, and the top N element entries are retained; where N is greater than or equal to 1. The integrity of the retained element entries and their corresponding entries in the relevant design element database is verified to obtain an initial element matching set.

[0012] Preferably, based on the intent tendency in the query parsing results, a first matching degree weighting is applied to each element entry in the relevant design element database, including: The intent tendency features are extracted from the question parsing results, and the dimensions of the intent tendency features are decomposed; wherein, the dimensions of the intent tendency features include functional requirements, standard compliance, cost control and aesthetic preferences, and weight values ​​are assigned to the dimensions of the intent tendency features based on the domain knowledge base; The element entries in the relevant design element database are classified according to attribute tags to obtain feature vectors; wherein, the feature vectors include item function descriptions, compliance basis, cost parameters, and morphological characteristics; Calculate the cosine similarity between the feature vector of each element item and the intention tendency dimension to obtain the basic matching degree value; The basic matching score is weighted and fused using the intention tendency dimension weight value to obtain the first matching score weight value; The first matching degree weighted value is bound to the corresponding feature entry to obtain a set of feature entries.

[0013] Preferably, the element entries weighted by matching degree are processed according to the core keywords to obtain a weighted element matching set, including: Core keywords are extracted from the query parsing results to construct a keyword feature network; wherein, the keyword feature network includes synonym expansion and hyponym / hypernym association, and each node in the keyword feature network is assigned a semantic association strength value; The set of feature entries is decomposed into structured fields; wherein, the structured fields include terms, parameters, and case references, to obtain an entry feature matrix; The semantic overlap between the keyword feature network and the item feature matrix is ​​calculated to obtain the keyword matching coefficient; wherein, the keyword matching coefficient is positively correlated with the embedding depth of the core keyword in the item field; The keyword matching coefficient is multiplied by the first matching degree weighted value of the element entry to obtain the comprehensive matching value; The element entries are sorted in descending order based on the comprehensive matching value, and the entries with comprehensive matching values ​​higher than the dynamic threshold are selected to obtain a weighted element matching set.

[0014] Preferably, generating a target feature matching set based on the initial feature matching set and the weighted feature matching set includes: Feature extraction is performed on the initial feature matching set and the weighted feature matching set to construct dual-set feature maps respectively; wherein, the dual-set feature maps include feature type, association strength and scene adaptability; Calculate the overlap of feature entries in the binary feature map, mark the co-occurring features that exist in both sets, and obtain the co-occurrence confidence value; For unique elements that exist only in the initial element matching set, supplementary scoring is performed by combining the features of similar elements in the weighted element matching set to obtain the unique element fit value. For high-weighted elements that exist only in the weighted element matching set, their rationality is verified by the association rules of the initial element matching set to obtain verification coefficients; The co-occurring elements, the unique elements with supplementary scoring, and the verified high-weight elements are integrated according to a comprehensive priority. After removing redundant and conflicting entries, the target element matching set is obtained.

[0015] To achieve the above objectives, this application also provides the following technical solution: a question-and-answer system for architectural design elements based on large language model technology, applicable to the aforementioned question-and-answer method for architectural design elements based on large language model technology, comprising: The data acquisition unit is used to collect user question texts and related design element databases for architectural design scenarios. The data parsing unit is used to parse the core keywords, element relationships and intent tendencies in the user's question text according to the preset semantic parsing model, and obtain the question parsing result; The element set unit is used to acquire relevant design element entries including the core keyword and redundant entries excluding the core keyword; and to determine the effective design element set based on the redundant entries and the corresponding filtering coefficients. The element filtering unit is used to subtract the set of valid design elements from the relevant design element database to obtain a supplementary element database; filter the supplementary element database according to the element association relationship in the query parsing result; and extract an initial element matching set from the relevant design element database. The element weighting unit is used to perform a first matching degree weighting on each element entry in the relevant design element database based on the intent tendency in the query parsing result; and to process the matching degree weighted element entries according to the core keywords to obtain a weighted element matching set. The element question-and-answer unit is used to generate a target element matching set based on the initial element matching set and the weighted element matching set; after the target element matching set is processed by natural language generation, it is associated with the user's question text to obtain the architectural design element question-and-answer results.

[0016] This application significantly reduces the time designers spend searching for relevant design elements by rapidly and accurately parsing and matching user query text. Designers no longer need to manually sift through redundant information; the system automatically filters and provides effective design elements, significantly improving work efficiency. Furthermore, by parsing and filtering redundant entries, only the set of design elements relevant to user needs is retained, effectively avoiding information overload and interference from repetitive content, allowing designers to obtain the most relevant design suggestions within a limited time. Moreover, by generating and updating a supplementary element library, the design element library can be continuously improved and adapted to new design requirements. This dynamic update mechanism enables the system to provide timely and up-to-date design suggestions in different architectural design scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of an embodiment of the architectural design element question-answering method based on large language model technology in this application. Figure 2 This is a schematic diagram of the system architecture of an embodiment of the architectural design element question-and-answer system based on large language model technology of this application.

[0018] Attached figure labels: 1. Data acquisition unit; 2. Data parsing unit; 3. Feature aggregation unit; 4. Feature filtering unit; 5. Feature weighting unit; 6. Feature question and answer unit. Detailed Implementation

[0019] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0020] like Figure 1 As shown, the architectural design element question-answering method proposed in this invention based on large language model technology includes: S1. Collect user question texts and related design element databases for architectural design scenarios; S2. Based on the preset semantic parsing model, analyze the core keywords, element relationships and intent tendencies in the user's question text to obtain the question parsing results; S3. Obtain relevant design element entries including core keywords and redundant entries excluding core keywords, and determine the effective design element set based on the redundant entries and corresponding filtering coefficients. S4. Subtract the valid design element set from the relevant design element database to obtain a supplementary element database; filter the supplementary element database according to the element relationships in the query parsing results, and extract the initial element matching set from the relevant design element database. S5. Based on the intent tendency in the query analysis results, perform a first-degree matching weighting on each element entry in the relevant design element database; process the element entries after matching weighting based on the core keywords to obtain a weighted element matching set; S6. Generate a target element matching set based on the initial element matching set and the weighted element matching set; after processing the target element matching set with natural language, associate it with the user's question text to obtain the architectural design element question and answer results.

[0021] In this invention, when a user asks a question about architectural design, the system automatically records the question. The design element database is a collection containing various architectural design elements (such as materials, structure, spatial layout, environment, etc.). The system analyzes the user's question using a preset semantic parsing model, extracting core keywords, element relationships, and intent. For example, the keywords in the user's question might be "summer climate" and "architectural design." For example, the parsing results might be: Core keywords: summer climate, architectural design; Element relationships: climate considerations, energy-saving design, natural ventilation, building material selection, etc. Based on the parsing results, the system queries the design element database to find entries related to summer climate and architectural design (such as natural ventilation design, thermal insulation materials, etc.), while also finding irrelevant entries (such as winter heating design, sealing materials, etc.). For example: Relevant design elements: ventilation design, shading design, facade materials (such as reflective materials); Redundant design elements: winter heating, excessive insulation design, etc. Redundant entries are removed, and an effective design is obtained based on the core keywords and design intent. The element set is then streamlined. Effective design elements include natural ventilation, shading devices, and green roofs. Ineffective design elements include excessive insulation materials and winter heating systems. The system generates a supplementary element set and filters out more relevant design elements based on the relationships between elements in the user's question. This supplementary element set includes additional elements not in the effective set. The supplementary element set may include UV-resistant coatings, adjustable window designs, and green building certifications. The system weights design elements based on the intent expressed in the question. For example, if the user emphasizes energy efficiency or ventilation design, the system will prioritize these elements. For instance, when the user mentions summer climate, the system may weight ventilation and shading designs, considering them more relevant than other design elements. Based on the filtering and weighting processes described above, the system ultimately generates a target design element matching set containing the most relevant and user-relevant design elements. Finally, the system translates these elements into natural language, generating answers relevant to the user's question to help the user understand how to design.

[0022] In an optional embodiment, a database of user-generated questions and related design elements for architectural design scenarios is collected, including: Multi-source raw text data is obtained synchronously through the architectural design community interaction interface, the feedback module embedded in the design tool, and the interface for historical project consultation records. Multi-source raw text data is segmented and denoised to identify user interaction fragments and design specification fragments including architectural terminology, resulting in classified preprocessed text. For the preprocessed text, extract user interaction fragments with question marks as candidate question texts, and extract design specification clause fragments to construct an initial element item pool; The candidate question texts are compared with the word vector library in the field of architecture to select the texts that match the characteristics of the design scenario as the user question texts; Based on the initial element entry pool, structured element units are obtained by associating design specification numbers with element attribute tags through a knowledge graph; The structured element units are clustered according to the building type dimension, and the clustering results are optimized by combining high-frequency domain words in the user question text to obtain a database of relevant design elements.

[0023] It should be noted that the system collects raw data from multiple sources through the architectural design community interaction interface, the feedback module embedded in the design tool, and the interface for historical project consultation records. This data may include user questions, design scheme feedback, and design highlights from historical projects. For example, user feedback within the design tool may include: how well this material performs at high temperatures; discussions in the design community may discuss the need to select exterior wall materials suitable for humid environments; and historical project records may show that X-type weathering steel was used in a 2015 project. The system processes this multi-source raw text data, including word segmentation and noise reduction. Word segmentation involves dividing the text into individual words, while noise reduction removes irrelevant or redundant parts. Then, the system identifies and separates architectural terminology, user interaction snippets, and design code clauses, generating categorized preprocessed text. The system then extracts data from the categorized text. User interaction fragments marked with question tags serve as candidate question texts. Simultaneously, design specification fragments are extracted to construct an initial element entry pool (including building materials, building technical standards, etc.). The system compares the candidate question texts with a word vector library in the architectural field for similarity. This word vector library contains information such as architectural terminology, design specifications, and technical requirements. Through comparison, the system selects the question texts most relevant to the design scenario. For example, the comparison result shows that the user question "Is glass suitable for high-temperature environments?" has a high similarity to the design specification clause "Glass high-temperature resistance," which fits the scenario and is therefore selected as a valid user question text. The system uses a knowledge graph to associate the design specification numbers and element attribute tags in the initial element entry pool, obtaining structured element units. These element units are standardized design scheme elements that can be further optimized. For example: Glass material → Design code number (e.g., GB / T 8001-2015), attribute tags (e.g., high temperature resistance, transparency, thermal insulation); Exterior glass → Design code number (e.g., GB 50096-2011), attribute tags (e.g., aesthetics, wind pressure resistance); The system clusters structured element units according to building type (e.g., residential, commercial, industrial, etc.) and optimizes the clustering results by combining high-frequency domain terms in user questions (e.g., high temperature, thermal insulation, facade, etc.); This step accurately matches design elements according to actual needs; For example: residential buildings: clustering design elements that are more concerned with thermal insulation performance and temperature control; commercial buildings: clustering design elements that are more concerned with aesthetics and wind pressure resistance; Through the optimized clustering results, the system finally generates an accurate database of relevant design elements; This database contains design elements that meet user needs and can provide reasonable solution suggestions for specific design scenarios.

[0024] In an optional embodiment, the core keywords, element relationships, and intent tendencies in the user's question text are analyzed according to a preset semantic parsing model to obtain the question parsing result, including: The contextual information of user questions is completed, and omitted expressions are filled in using a knowledge base in the field of architectural design to obtain complete semantic text. The complete semantic text is input into the terminology recognition module of the preset semantic parsing model, and the core keywords with design element attributes are identified through domain-specific word vector matching to obtain a keyword sequence. Based on the keyword sequence, a word association graph is constructed using the dependency parsing unit of the semantic parsing model to mark the design logical relationships and obtain the element association network; Based on the element association network, interrogative adverbs, modal words, and sentiment words are extracted from the keyword sequence to obtain the intent feature vector; The keyword sequence, element association network, and intent feature vector are fused and encoded to obtain the question parsing result.

[0025] It's important to note that filling in the missing parts of user questions ensures semantic clarity and ease of parsing. Users often omit crucial details in their questions, which are essential for understanding the problem. The semantic parsing model's terminology recognition module identifies design elements in the completed statement and extracts relevant core keywords. Dependency parsing techniques are used to analyze the relationships between keywords and transform them into a relational graph, indicating the logical relationships between design elements. Interrogative adverbs, modal words, and sentiment words that may be present in the user question are extracted; these features help in understanding the user's design intent. The extracted keyword sequences, dependency networks, and intent feature vectors are then fused to obtain the final question parsing result, providing structured suggestions or solutions for design decisions.

[0026] In an optional embodiment, relevant design element entries including core keywords and redundant entries excluding core keywords are obtained. A valid set of design elements is determined based on the redundant entries and their corresponding filtering coefficients, including: Core keywords are extracted from the question analysis results, and a keyword feature matrix is ​​constructed; where the matrix elements represent the correlation strength of keywords in the design element items. Perform matrix matching operation on the keyword feature matrix and the related design element database, and filter out the entries whose matrix element values ​​are greater than a preset threshold as related design element entries that include the core keywords; The matrix matching operation formula is as follows: ; in, Indicates the first The first core keyword and the first The association strength of each design element item, with a value range of [0,1]. The larger the value, the stronger the association. Indicates the first Feature vectors of core keywords Indicates the first Feature vector of each design element item express and The vector dot product, , express and The L2 norm (vector magnitude, used for normalization); The remaining entries in the relevant design element database that were not selected as relevant design element entries are marked as redundant entries that do not include core keywords; A secondary assessment of the domain relevance of redundant entries is conducted, and a corresponding filtering coefficient is assigned to each redundant entry based on the implicit requirements of the architectural design scenario; wherein, the filtering coefficient is inversely proportional to the potential value of the entry. Redundant entries are filtered hierarchically based on the filtering coefficient. Redundant entries with filtering coefficients below the threshold are retained and merged with relevant design element entries including core keywords to obtain a set of effective design elements.

[0027] It should be noted that core design-related keywords are extracted from the user's question analysis results, and a feature matrix is ​​constructed based on these keywords to represent the correlation strength of keywords among design element items. These keywords are then correlated with design items in the database, and the correlation strength of each keyword within an item is calculated. For example, if an item related to the high-temperature resistance of glass materials has a high correlation strength with both glass materials and high-temperature environments, the corresponding element value in the matrix will be larger. The constructed keyword feature matrix is ​​matched with the design element database to filter out design element items containing the core keywords. If the correlation strength between the high-temperature resistance of glass materials and the core keywords "glass materials and high-temperature environments" is greater than a set threshold (e.g., 0.8), then that item will be selected as a relevant design element. Other items, such as aesthetic considerations in facade design, have a weaker correlation strength with high-temperature environments and are therefore not selected. The remaining items not selected as relevant design element items are then further processed... Rows are marked as redundant entries; these entries, although existing in the design element database, have low relevance to the core keywords. A secondary assessment of the domain relevance of redundant entries is performed, and each redundant entry is assigned a filtering coefficient based on the implicit requirements of the architectural design. The filtering coefficient is inversely proportional to the potential value of the entry. Redundant entries are then graded and filtered according to their filtering coefficients, retaining those with filtering coefficients below a critical value (e.g., 0.5). These entries will be merged with relevant design element entries to form the final effective design element set. For example, considering the aesthetics of the facade design, a filtering coefficient of 0.2 is used, which is below the critical value of 0.5, so it is retained and merged with other relevant design elements. If there are redundant entries, such as those related to the wind load-bearing capacity of the facade design, their filtering coefficients are high and they may be filtered out. Finally, an effective design element set is obtained, including design elements related to the core keywords and the filtered redundant entries, forming a comprehensive element set that meets the design requirements.

[0028] In an optional embodiment, the relevant design element database is subtracted from the set of valid design elements to obtain a supplementary element library, including: All element entries in the relevant design element database are uniquely identified and encoded to obtain a complete set of coded element entries; Encode each element entry in the valid design element set with a unique identifier according to the same rules to obtain a coded set of valid element entries. Compare the coded full set of feature entries and the coded valid set of feature entries, and mark the coded difference entries in the full set of features that do not appear in the valid set of feature entries. The initial pool of differential elements is obtained by matching the original element content in the relevant design element database with the differential coding entries. The initial pool of differential elements is tested for its suitability to the current architectural design scenario. Elements that are completely irrelevant to the current architectural design scenario are removed to obtain a supplementary element library.

[0029] It should be noted that assigning a unique identifier (such as ID) to each entry in the design element database effectively manages the information of each entry. Suppose the design element database has three entries: thermal insulation performance of the glass material (code: A001); aesthetic considerations of the facade design (code: B002); and wind load-bearing capacity of the facade design (code: C003). Therefore, the complete set of element entries would be: A001: thermal insulation performance of the glass material; B002: aesthetic considerations of the facade design; C003: wind load-bearing capacity of the facade design. The same rule applies to the valid design element set, encoding these element entries and generating a coded set of valid element entries. Suppose the valid design element set only contains: thermal insulation performance of the glass material (code: A001); and aesthetic considerations of the facade design (code: B002). Therefore, the set of valid element entries would be: A001: thermal insulation performance of the glass material; B002: aesthetic considerations of the facade design. Comparing the coded set of complete element entries with the coded set of valid element entries identifies those that are not present in the valid element entries. Concentrated entries (i.e., difference entries); for example: Full set of element entries: A001: Thermal insulation performance of glass materials; B002: Aesthetic considerations of facade design; C003: Wind load-bearing capacity of facade design; Valid set of element entries: A001: Thermal insulation performance of glass materials; B002: Aesthetic considerations of facade design; Difference-coded entry: C003 (Wind load-bearing capacity of facade design) is not in the valid element set, therefore it is marked as a difference-coded entry; Based on the difference-coded entry, find the content of the original entry in the relevant design element database, and form... A preliminary pool of difference elements is created; these entries are those that do not match the current set of valid design elements; for example, the difference code entry: C003; by matching the original database, the original content of this entry is found: wind load capacity of the facade design; therefore, the initial difference element pool is: wind load capacity of the facade design; the entries in the initial difference element pool are subjected to adaptability verification to check whether these entries are relevant to the current architectural design scenario; entries that are completely irrelevant to the design scenario are removed, and the remaining entries form a supplementary element library as alternative entries that can be used in the current design scenario.

[0030] In an optional embodiment, the supplementary feature library is filtered based on the feature relationships in the query parsing results, and an initial feature matching set is extracted from the relevant design feature database, including: Extract the relationships between elements from the query analysis results and construct a relationship network; the relationship network includes causal chains, constraint pairs and cooperative groups, and assign scenario-adaptive weights to each relationship type; The element entries in the supplementary element library are broken down into triples to obtain the element association map; Calculate the topological similarity between the 3D relationship network and the feature relationship map, and select feature items with similarity higher than a set threshold as candidate related features; The formula for calculating topological similarity is as follows: ; in, This represents the topological similarity (value range [0,1], the larger the value, the more similar the two network structures are). This represents the set of nodes in the element relationship network of the query analysis results, where the nodes are the core keywords. This represents the set of nodes in the feature association map of the supplementary feature library, where nodes are feature entry terms. This represents the edge set of a network of relationships between elements, where each edge represents a logical relationship between keywords. This represents the edge set of an element association graph, where edges represent the association relationships between element entries. This represents the size of the intersection of two node sets and the number of co-occurring nodes. This represents the size of the union of two node sets, and the total number of nodes. This represents the node matching weight and edge matching weight; Based on the strength of the element relationships in the query analysis results, the candidate related elements are prioritized and ranked, and the top N element entries are retained; where N is greater than or equal to 1. The integrity of the retained element entries and their corresponding entries in the relevant design element database is verified to obtain the initial element matching set.

[0031] It should be noted that the relationships extracted from the questions or design elements form causal chains, constraint pairs, and synergistic groups. These relationships can help analyze the dependencies and mutual influences between elements. Assume the design element library contains the following elements: A001: Thermal insulation performance of glass materials; B002: Aesthetic considerations of facade design; C003: Wind load-bearing capacity of facade design. These elements may have the following relationships: Causal chain: The thermal insulation performance of glass materials affects the heat regulation of the facade design; Constraint pair: Aesthetic design and thermal insulation performance may conflict in some cases, for example, aesthetic design requires transparency, while thermal insulation requires thicker glass; Synergistic group: Aesthetic considerations and wind load-bearing capacity can work together to improve the overall design effect of the facade. These relationships are constructed into a network of relationships: A001 → B002 (causal chain); B002 C003 (Collaboration Group); A001 B002 (Constraint Pairs); Assign appropriate weights to each type of association (causal chain, constraint pair, synergy group) based on its importance in the current design scenario; for example: assuming the current design scenario is in a cold region, the thermal insulation of the facade is more important than aesthetic design; therefore, the causal chain (the impact of thermal insulation performance on facade design) will be given a higher weight, while aesthetic considerations (constraint pairs) will be given a lower weight; assuming the weights are as follows: causal chain: 0.7; constraint pair: 0.2; synergy group: 0.1; Decompose each entry in the supplementary element library into triplets (i.e., the relationship between one element and two other elements) to form an element association map; this helps to further analyze the interrelationships between elements; for example: assuming the entries in the supplementary element library include: D004: cost of thermal insulation materials; E005: stability of building structures under high temperature environments; these entries can be decomposed into triplets, such as: D004 → A001 (thermal insulation materials affect the thermal insulation performance of glass); E005 → C003 (structural stability may affect wind load capacity); D004 B002 (Related to material cost and aesthetic design of the facade); Calculate the topological similarity between the 3D relationship network and the element relationship map, and select elements with similarity higher than a set threshold as candidate related elements; for example: assuming the 3D relationship network contains the aforementioned causal chains, constraint pairs, and cooperative group relationships, and the element relationship map includes new supplementary elements (such as D004, E005, etc.); by calculating similarity, it is found that the relationship between D004 (cost of thermal insulation material) and A001 (thermal insulation performance of glass material) is very similar, therefore D004 is selected as a candidate related element; based on the element relationship in the question analysis results... The strength of the correlation is ranked, and the candidate related elements are sorted, with the top N elements retained. For example, assuming the strength of the correlation is defined by the scenario adaptation weight, the causal chain relationship is stronger, so the candidate elements related to thermal insulation performance (such as D004) will be retained first. For example, D004 (cost of thermal insulation material) may rank first, while E005 (stability of building structure under high temperature environment) ranks second. The retained candidate related elements are then checked against the corresponding entries in the relevant design element database to ensure that these elements are complete and applicable in the design scenario, thus forming the initial element matching set.

[0032] In an optional embodiment, based on the intent tendency in the query parsing results, a first matching degree weighting is applied to each element entry in the relevant design element database, including: The intent tendency features are extracted from the query analysis results and their dimensions are broken down. The intent tendency feature dimensions include functional requirements, standard compliance, cost control and aesthetic preferences. Weight values ​​are assigned to the intent tendency feature dimensions based on the domain knowledge base. The element entries in the relevant design element database are classified according to attribute tags to obtain feature vectors; the feature vectors include the functional description of the entry, compliance basis, cost parameters and morphological characteristics. Calculate the cosine similarity between the feature vector of each element item and the intention tendency dimension to obtain the basic matching degree value; The formula for calculating cosine similarity is as follows: ; in, Indicates the first The element entry and the first The basic matching degree of each intention tendency dimension (value range [0,1]). Indicates the first Feature vectors of each element entry Indicates the first A vector of intentional tendencies; The first weighted matching score is obtained by weighting and fusing the basic matching score by the intention tendency dimension weight value; The formula for calculating the weighted value of the first matching degree is as follows: ; in, Indicates the first The first matching degree weighted value of each element item (the value ranges from [0,1], and the larger the value, the better it matches the user's intent). Indicates the first The weights of each intention orientation dimension; The first matching degree weighted value is bound to the corresponding feature entry to obtain the feature entry set.

[0033] It should be noted that intent-oriented features are extracted from the question analysis and broken down into multiple dimensions. Each dimension reflects different aspects of the design goals, such as functional requirements, regulatory compliance, cost control, and aesthetic preferences. Entries in the design element database are categorized by attribute tags, and a feature vector is generated for each entry. Each entry's feature vector includes a functional description, compliance basis, cost parameters, and morphological characteristics. For example, suppose there is an entry in the design element database: Entry A: Material Option—Glass; Functional Description: Good thermal insulation; Compliance Basis: Meets building thermal performance standards; Cost Parameter: Moderate cost; Morphological Characteristics: Transparent, suitable for modern architectural styles; The corresponding feature vector is: Functional Description: Good thermal insulation performance; Compliance Basis: Meets thermal performance standards; Cost Parameter: Moderate; Morphological Characteristics: Transparent, modern. The cosine similarity between the feature vector of each design element and the intent-oriented feature dimensions is calculated to measure their similarity in functional requirements, regulatory compliance, and aesthetic preferences. The degree of matching in terms of compliance, cost control, and aesthetic preferences; for example: Suppose we want to calculate the similarity between item A (glass material) and the intentional preference dimension: Functional requirements (temperature difference resistance): Item A's thermal insulation performance and temperature difference resistance are similar; Regulatory compliance (building regulations): Item A meets thermal performance standards, so it has a high similarity in regulatory compliance; Cost control (material saving): Item A has a moderate cost and a weaker ability to save materials compared to other materials; Aesthetic preference (modern design): Item A is suitable for modern architectural styles, so it has a high aesthetic similarity; The calculated cosine similarity may be as follows: Functional requirements: 0.8; Regulatory compliance: 0.9; Cost control: 0.5; Aesthetic preference: 0.7; Multiply the cosine similarity value of each dimension by the weight value of that dimension, and then perform weighted fusion to obtain a weighted matching value; Bind the weighted matching value to the corresponding design element item to obtain a set containing multiple element items with high matching degrees.

[0034] In an optional embodiment, the matching-weighted element entries are processed based on core keywords to obtain a weighted element matching set, including: Core keywords are extracted from the query analysis results, and a keyword feature network is constructed. The keyword feature network includes synonym expansion and hyponym / hypernym association, and each node in the keyword feature network is assigned a semantic association strength value. The feature entry set is decomposed into structured fields; these structured fields include terms, parameters, and case references, resulting in an entry feature matrix. The semantic overlap between the keyword feature network and the item feature matrix is ​​calculated to obtain the keyword matching coefficient; among which, the keyword matching coefficient is positively correlated with the embedding depth of the core keywords in the item field; The keyword matching coefficient is multiplied by the weighted value of the first matching degree of the element entry to obtain the comprehensive matching value. The feature entries are sorted in descending order based on the comprehensive matching value. Entries with comprehensive matching values ​​higher than the dynamic threshold are then selected from the sorted results to obtain a weighted feature matching set.

[0035] It should be noted that core keywords are extracted from the question parsing, and a keyword feature network is constructed. This network includes synonym expansion and hyponym / hypernym association, and a semantic association strength value is assigned to each keyword node. Synonyms of the core keywords are added to the feature network to increase keyword diversity. Semantic understanding is enhanced through hyponym / hypernym relationships. The set of element entries is decomposed into structured fields to ensure that the information of each entry is expressed in a clear and standardized format. The structured fields include terms, parameters, and case references. The semantic overlap between the keyword feature network and the entry feature matrix is ​​calculated; that is, the semantic overlap in the entry feature matrix is ​​checked. The degree of overlap between terminology, parameters, and case references and the keyword network yields a matching coefficient; this coefficient measures the degree of matching between keywords and entries; the keyword matching coefficient is positively correlated with the embedding depth of the core keyword in the entry field; that is, the deeper the keyword is embedded in the entry field, the higher the matching coefficient will be; the keyword matching coefficient is multiplied by the first matching degree weighted value of the entry to obtain a comprehensive matching value; this comprehensive matching value is the overall matching degree between the entry and the demand; the entries are sorted in descending order according to the comprehensive matching value, and the entries with comprehensive matching values ​​higher than the dynamic threshold are selected to form a weighted element matching set.

[0036] In an optional embodiment, generating a target feature matching set based on an initial feature matching set and a weighted feature matching set includes: Feature extraction is performed on the initial feature matching set and the weighted feature matching set to construct dual-set feature maps; the dual-set feature maps include feature type, association strength and scene adaptability; Calculate the overlap of feature entries in the biset feature map, mark the co-occurring features that exist in both sets, and obtain the co-occurrence confidence value; For unique elements that exist only in the initial element matching set, supplementary scoring is performed by combining the features of similar elements in the weighted element matching set to obtain the unique element fit value. For high-weighted elements that exist only in the weighted element matching set, their rationality is verified by the association rules of the initial element matching set, and the verification coefficient is obtained. By integrating co-occurring elements, unique elements with supplementary scoring, and verified high-weight elements according to comprehensive priority, and removing redundant and conflicting entries, the target element matching set is obtained.

[0037] It should be noted that features are extracted from the initial feature matching set and the weighted feature matching set to construct dual-set feature maps. Each feature map includes feature type, association strength, and scene adaptability. Feature type refers to the type of each feature, such as material, function, or performance. Association strength measures the correlation between features; for example, from a technical perspective, some materials are strongly correlated with certain functions. Scene adaptability refers to the degree of adaptability of a feature in a specific scene, such as the applicability of a certain material in cold regions. The overlap of feature entries in the dual-set feature maps is calculated, and co-occurring features that exist simultaneously in both sets are marked to obtain co-occurrence confidence values. Overlap refers to the degree of overlap of a feature in the initial feature matching set and the weighted feature matching set. Co-occurrence confidence measures the reliability or credibility of the co-occurrence of the same feature in the two sets. For unique features that exist only in the initial feature matching set, supplementary scoring is performed by combining similar feature features from the weighted feature matching set, thereby obtaining the final score. The matching value of unique elements; unique elements exist only in the initial matching set but not in the weighted matching set; the score is supplemented by the features of similar elements in the weighted matching set to improve the matching degree of unique elements; for high-weight elements that only exist in the weighted matching set, their rationality is verified by the association rules of the initial element matching set to ensure their rationality and obtain the verification coefficient; high-weight elements refer to elements with high importance in the weighted element matching set; their rationality is verified by the rules of the initial matching set, and if they meet the rules, a high verification coefficient is given; co-occurring elements, unique elements with supplemented scores, and verified high-weight elements are integrated according to the comprehensive priority, and redundant or conflicting entries are removed to finally obtain the target element matching set; redundant and conflicting entries refer to elements that may be listed repeatedly or conflictingly in multiple steps and need to be removed; comprehensive priority refers to the final priority of each element based on co-occurrence confidence, supplemented scores, and verification coefficients, and then integrated.

[0038] like Figure 2 As shown, the architectural design element question-answering system proposed in this invention, based on large language model technology, is applicable to the aforementioned architectural design element question-answering method based on large language model technology, and includes: Data acquisition unit 1 is used to collect user question texts and related design element databases for architectural design scenarios; Data parsing unit 2 is used to parse the core keywords, element relationships and intent tendencies in the user's question text according to the preset semantic parsing model, and obtain the question parsing results; Element set unit 3 is used to obtain relevant design element entries including core keywords and redundant entries that do not include core keywords; and to determine the effective design element set based on the redundant entries and the corresponding filtering coefficients. Element filtering unit 4 is used to subtract the set of valid design elements from the relevant design element database to obtain a supplementary element library; the supplementary element library is filtered according to the element association relationship in the query parsing result, and an initial element matching set is extracted from the relevant design element database; Element weighting unit 5 is used to perform first-degree matching weighting on each element entry in the relevant design element database based on the intent tendency in the query parsing results; and to process the element entries after matching weighting based on the core keywords to obtain a weighted element matching set. Element question and answer unit 6 is used to generate a target element matching set based on the initial element matching set and the weighted element matching set; after the target element matching set is processed by natural language generation, it is associated with the user's question text to obtain the architectural design element question and answer results.

[0039] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A question-and-answer method for architectural design elements based on large language model technology, characterized in that: include: Collect user question texts and related design element databases for architectural design scenarios; The core keywords, element relationships, and intent tendencies in the user's question text are analyzed according to the preset semantic parsing model to obtain the question parsing result; Obtain relevant design element entries including the core keywords and redundant entries excluding the core keywords, and determine the effective design element set based on the redundant entries and the corresponding filtering coefficients; Subtract the set of valid design elements from the database of relevant design elements to obtain a supplementary element database; Based on the element relationships in the question parsing results, the supplementary element library is filtered, and an initial element matching set is extracted from the relevant design element database; Based on the intent tendency in the question parsing results, a first matching degree weighting is applied to each element entry in the relevant design element database; Based on the core keywords, the element entries weighted by matching degree are processed to obtain a weighted element matching set; A target element matching set is generated based on the initial element matching set and the weighted element matching set; the target element matching set is then processed by natural language generation and associated with the user's question text to obtain the architectural design element question and answer results.

2. The question-and-answer method for architectural design elements based on large language model technology according to claim 1, characterized in that, Collect user question texts and related design element databases for architectural design scenarios, including: Multi-source raw text data is obtained synchronously through the architectural design community interaction interface, the feedback module embedded in the design tool, and the interface for historical project consultation records. The multi-source raw text data is segmented and denoised to identify user interaction fragments and design specification fragments including architectural terms, resulting in classified preprocessed text. For the preprocessed text of the classification, user interaction fragments containing question marks are extracted as candidate question texts, and design specification clause fragments are extracted to construct an initial element item pool; The candidate question texts are compared with the word vector library in the field of architecture to select texts that match the characteristics of the design scenario as user question texts; Based on the initial element entry pool, structured element units are obtained by associating design specification numbers with element attribute tags using a knowledge graph; The structured element units are clustered according to the building type dimension, and the clustering results are optimized by combining the high-frequency domain words in the user question text to obtain a relevant design element database.

3. The question-and-answer method for architectural design elements based on large language model technology according to claim 2, characterized in that, The core keywords, element relationships, and intent tendencies in the user's question text are analyzed based on a preset semantic parsing model to obtain the question parsing results, including: The user's question text is processed to complete the context, and the omitted expressions are filled in by combining the knowledge base of architectural design to obtain the complete semantic text; The complete semantic text is input into the terminology recognition module of the preset semantic parsing model, and the core keywords with design element attributes are identified through domain-specific word vector matching to obtain a keyword sequence; Based on the keyword sequence, a word association graph is constructed using the dependency parsing unit of the semantic parsing model to mark the design logic relationships and obtain the element association network; Based on the element association network, interrogative adverbs, modal words, and sentiment words are extracted from the keyword sequence to obtain an intent feature vector; The keyword sequence, the element association network, and the intent feature vector are fused and encoded to obtain the question parsing result.

4. The question-and-answer method for architectural design elements based on large language model technology according to claim 3, characterized in that, Obtain relevant design element entries including the core keyword and redundant entries excluding the core keyword. Determine the effective design element set based on the redundant entries and their corresponding filtering coefficients, including: Core keywords are extracted from the query analysis results, and a keyword feature matrix is ​​constructed; where the matrix elements represent the correlation strength of keywords in design element items. Perform matrix matching operation on the keyword feature matrix and the related design element database, and filter out entries whose matrix element values ​​are greater than a preset threshold as related design element entries including core keywords; The remaining entries in the relevant design element database that were not selected as relevant design element entries are marked as redundant entries that do not include core keywords; A secondary evaluation of the domain relevance of the redundant entries is performed, and a corresponding filtering coefficient is assigned to each redundant entry in combination with the implicit requirements of the architectural design scenario; wherein, the filtering coefficient is inversely proportional to the potential value of the entry; Redundant entries are filtered hierarchically based on the filtering coefficient. Redundant entries with filtering coefficients below the threshold are retained and merged with the relevant design element entries including the core keywords to obtain a set of effective design elements.

5. The question-and-answer method for architectural design elements based on large language model technology according to claim 4, characterized in that, Subtracting the set of valid design elements from the database of relevant design elements yields a supplementary element library, including: All element entries in the relevant design element database are uniquely identified and encoded to obtain a complete set of coded element entries. Encode each element entry in the set of valid design elements using unique identifiers according to the same rules to obtain a set of coded valid element entries. The encoded full set of element entries and the encoded valid set of element entries are compared and the differences in encoding are marked in the full set of element entries that do not appear in the valid set of element entries. The initial difference element pool is obtained by matching the original element content in the relevant design element database with the difference coding entries. The initial difference element pool is subjected to design scenario adaptability verification, and element entries that are completely irrelevant to the current architectural design scenario are removed to obtain a supplementary element library.

6. The question-and-answer method for architectural design elements based on large language model technology according to claim 5, characterized in that, Based on the element relationships in the query parsing results, the supplementary element library is filtered, and an initial element matching set is extracted from the relevant design element database, including: Extract the element relationships from the query parsing results and construct a relationship network; wherein, the relationship network includes causal chains, constraint pairs and cooperative groups, and assign scenario-adaptive weights to each relationship type; The element entries in the supplementary element library are decomposed into triples to obtain the element association map; Calculate the topological similarity between the three-dimensional relationship network and the element relationship map, and select element entries with similarity higher than a set threshold as candidate related elements; Based on the strength of the element relationships in the query analysis results, the candidate related elements are prioritized and ranked, and the top N element entries are retained; where N is greater than or equal to 1. The integrity of the retained element entries and their corresponding entries in the relevant design element database is verified to obtain an initial element matching set.

7. The question-and-answer method for architectural design elements based on large language model technology according to claim 6, characterized in that, Based on the intent tendency in the query parsing results, a first-degree matching weighting is applied to each element entry in the relevant design element database, including: The intent tendency features are extracted from the question parsing results, and the dimensions of the intent tendency features are decomposed; wherein, the dimensions of the intent tendency features include functional requirements, standard compliance, cost control and aesthetic preferences, and weight values ​​are assigned to the dimensions of the intent tendency features based on the domain knowledge base; The element entries in the relevant design element database are classified according to attribute tags to obtain feature vectors; wherein, the feature vectors include item function descriptions, compliance basis, cost parameters, and morphological characteristics; Calculate the cosine similarity between the feature vector of each element item and the intention tendency dimension to obtain the basic matching degree value; The basic matching score is weighted and fused using the intention tendency dimension weight value to obtain the first matching score weight value; The first matching degree weighted value is bound to the corresponding feature entry to obtain a set of feature entries.

8. The question-and-answer method for architectural design elements based on large language model technology according to claim 7, characterized in that, Based on the core keywords, the element entries are processed using a matching degree weighting method to obtain a weighted element matching set, including: Core keywords are extracted from the query parsing results to construct a keyword feature network; wherein, the keyword feature network includes synonym expansion and hyponym / hypernym association, and each node in the keyword feature network is assigned a semantic association strength value; The set of feature entries is decomposed into structured fields; wherein, the structured fields include terms, parameters, and case references, to obtain an entry feature matrix; The semantic overlap between the keyword feature network and the item feature matrix is ​​calculated to obtain the keyword matching coefficient; wherein, the keyword matching coefficient is positively correlated with the embedding depth of the core keyword in the item field; The keyword matching coefficient is multiplied by the first matching degree weighted value of the element entry to obtain the comprehensive matching value; The element entries are sorted in descending order based on the comprehensive matching value, and the entries with comprehensive matching values ​​higher than the dynamic threshold are selected to obtain a weighted element matching set.

9. The question-and-answer method for architectural design elements based on large language model technology according to claim 8, characterized in that, Generating a target feature matching set based on the initial feature matching set and the weighted feature matching set includes: Feature extraction is performed on the initial feature matching set and the weighted feature matching set to construct dual-set feature maps respectively; wherein, the dual-set feature maps include feature type, association strength and scene adaptability; Calculate the overlap of feature entries in the binary feature map, mark the co-occurring features that exist in both sets, and obtain the co-occurrence confidence value; For unique elements that exist only in the initial element matching set, supplementary scoring is performed by combining the features of similar elements in the weighted element matching set to obtain the unique element fit value. For high-weighted elements that exist only in the weighted element matching set, their rationality is verified by the association rules of the initial element matching set to obtain verification coefficients; The co-occurring elements, the unique elements with supplementary scoring, and the verified high-weight elements are integrated according to a comprehensive priority. After removing redundant and conflicting entries, the target element matching set is obtained.

10. A question-and-answer system for architectural design elements based on large language model technology, applicable to the question-and-answer method for architectural design elements based on large language model technology as described in any one of claims 1-9, characterized in that, include: The data acquisition unit is used to collect user question texts and related design element databases for architectural design scenarios. The data parsing unit is used to parse the core keywords, element relationships, and intent tendencies in the user's question text according to a preset semantic parsing model, and obtain the question parsing result; The element set unit is used to acquire relevant design element entries including the core keyword and redundant entries excluding the core keyword; and to determine the effective design element set based on the redundant entries and the corresponding filtering coefficients. The element filtering unit is used to subtract the set of valid design elements from the relevant design element database to obtain a supplementary element database. Based on the element relationships in the question parsing results, the supplementary element library is filtered, and an initial element matching set is extracted from the relevant design element database; The element weighting unit is used to perform a first matching degree weighting on each element entry in the relevant design element database based on the intent tendency in the query parsing result; Based on the core keywords, the element entries weighted by matching degree are processed to obtain a weighted element matching set; The element question-and-answer unit is used to generate a target element matching set based on the initial element matching set and the weighted element matching set; after the target element matching set is processed by natural language generation, it is associated with the user's question text to obtain the architectural design element question-and-answer results.

Images