Decision data intelligent inquiry system and method based on generative large model

Abstract

The invention discloses an intelligent decision data inquiry system and method based on a generative large model, and belongs to the technical field of artificial intelligence. Natural language inquiry input by a user is received, and an inquiry intention and a target field are extracted; extracting structured and unstructured cost data from the real estate project database, and constructing a candidate field set; performing semantic matching by utilizing the generative large model to generate an intention field mapping matrix; constructing a field dependency graph based on a mapping result, and extracting a minimum closed-loop field set; performing credibility evaluation on the fields, and constructing a weighted field set; generating a dynamic query template, outputting a natural language answer, and attaching a field tracing path; if the field is missing or the credibility is insufficient, automatically generating a reverse query until a credibility requirement is met; according to the method, automatic association and accurate question and answer between complex semantic query and heterogeneous cost data are realized, and the method has good interpretability, traceability and intelligent interaction capability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an intelligent query system and method for decision data based on generative large models. Background Technology

[0002] With the increasing number of urban renewal and large-scale real estate development projects, target cost control, as a core component of cost management, plays a crucial role in early project decision-making. Due to the characteristics of real estate projects—long cycles, large investments, and numerous participants—the sources of target cost data are complex, with numerous fields and a lack of standardized format. These data include, but are not limited to, heterogeneous data sources such as tender lists, cost estimates, material unit price databases, contract payment milestones, land transfer costs, and financing cost forecasts. This data is often embedded in unstructured tables, PDF reports, and natural language descriptions, posing significant challenges to project managers' decision-making and inquiry processes.

[0003] Current intelligent query systems are mostly built on fixed templates or preset query statements, making it difficult to handle multi-source, heterogeneous, incomplete, and freely expressive data structures. In particular, they cannot automatically adapt to the semantic mapping and logical dependencies between different data fields. For example, when a user asks, "What percentage of the total cost of the current project is material cost?", the system needs to first identify the "material cost" field in the query, locate its corresponding value in multiple data sources, match it with the numerator and denominator of "total cost," and handle missing fields. In traditional methods, this series of operations often requires manual modeling and field matching, which is inefficient and has a high error rate. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent query system and method for decision data based on generative large models, so as to overcome the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent query method for decision data based on a generative large model, comprising:

[0006] Receive a user query Q in natural language form, and extract the key intent words I and the set of target fields involved. ;

[0007] Extract all structured and unstructured target cost data sources from the real estate project database D, and construct a candidate set of fields. Contains structured fields and unstructured embedded fields ;

[0008] Based on generative large models, Q and Perform semantic matching to generate an intent field mapping matrix M, where Intention words With fields The semantic relevance probability value;

[0009] A field dependency graph G=(V,E) is constructed based on M, where node V corresponds to the identified highly relevant fields, and edge E represents the logical dependencies between fields. Graph traversal is performed on G to extract the minimum closed-loop field set. , as the field group required for the current query;

[0010] right Each field Based on the structural characteristics and data integrity indicators of its source data Perform field credibility assessment to obtain a weighted field set. τ is the credibility threshold;

[0011] Will The fields in the model are used as input features to construct a dynamic query template TQ, where TQ is an intermediate query expression generated by the generative large model based on context autoregression.

[0012] The TQ template generation process is executed, and the answer text R is output, along with the field source path SR, which records the source and data value of each field corresponding to each answer fragment.

[0013] If R contains missing fields or low confidence markers, a reverse question Q' is automatically constructed until the confidence requirement is met.

[0014] Preferably, the generative large model-based pair of Q and Perform semantic matching to generate an intent field mapping matrix M, including:

[0015] Context encoding of Q is performed using a pre-trained large language model to generate the corresponding query vector representation VQ;

[0016] right Each field in Constructing semantic description vectors The It is obtained by concatenating field names, source data fragments, and semantic tags, and then encoding them using a language model.

[0017] VQ is calculated based on cosine similarity with each The matching scores between them are used to construct the intent field mapping matrix M.

[0018] Preferably, a graph traversal is performed on G to extract the minimum closed-loop field set. The fields required for the current query include:

[0019] Fields in M ​​with scores higher than a preset threshold τ Add the node set V to graph G, where τ is the field semantic relevance threshold;

[0020] Based on the co-occurrence frequency and semantic label relationship of fields in the project's historical data, a set of directed edges E between fields is established, and the edge weight represents the dependency strength between fields;

[0021] Perform a weighted depth-first traversal algorithm on graph G to identify the set of closed-loop paths formed between fields;

[0022] The minimum closed-loop field set is obtained by selecting the field-closed-loop path with the minimum sum of dependent edge weights. This is a combination of fields used to meet the semantic requirements of the current query.

[0023] Preferred, for Each field Based on the structural characteristics and data integrity indicators of its source data The field credibility is evaluated to obtain a weighted field set, including:

[0024] Extract each field The corresponding source data structure characteristic parameters include data type consistency, field fill rate, and time update frequency;

[0025] The field data integrity index δᵢ is calculated, which is obtained by weighting the field missing rate, the proportion of outliers, and the trust level of the source file.

[0026] Fields are filtered based on a preset confidence threshold τ. If the value is greater than or equal to τ, the field is considered a reliable field.

[0027] All satisfied The fields form a weighted field set .

[0028] Preferably, Using the fields in the table as input features, a dynamic query template TQ is constructed, including:

[0029] extract Each field Field names, units, field context semantic tags, and their corresponding credibility. Construct a sequence of field semantic vectors;

[0030] Generative large language models are used to jointly encode the user's original query Q and the sequence of field semantic vectors to form a contextual embedding representation;

[0031] Based on the autoregressive language modeling mechanism, an intermediate query expression TQ is generated.

[0032] Preferably, the TQ template generation process is executed, outputting the answer text R, along with the field tracing path SR, including:

[0033] The constructed intermediate query expression TQ is input into the generative large model decoder, and the response text R in natural language form is output based on the context autoregressive generation mechanism.

[0034] While generating R, record the unique identifier, source file path and field value of the referenced field corresponding to each semantic fragment, and construct the field source path SR;

[0035] Add corresponding field labels and data source descriptions to each result value or description in R to form a structured question-and-answer pair;

[0036] Bind R and SR to output a complete answer with interpretable field mappings.

[0037] Preferably, if R contains missing fields or low-confidence markers, a reverse question Q' is automatically constructed until the confidence requirement is met, including:

[0038] Parse the response text R and the field tracing path SR to detect whether there are missing fields or field confidence issues. Cases where the confidence threshold τ is less than the following;

[0039] Extracting missing or low-confidence fields And based on the original query intent, a supplementary reverse question template is constructed to generate reverse questions. ;

[0040] Will The input is fed into the generative large language model, triggering a new round of field semantic matching and credibility evaluation processes; until all generated fields meet the requirements. .

[0041] This invention also provides an intelligent query system for decision data based on generative large models, comprising:

[0042] Query Intent Parsing Module: Receives user queries Q in natural language format, extracts key intent words I and the set of target fields involved. ;

[0043] Multi-source data extraction module: Extracts all structured and unstructured target cost data sources from the real estate project database D, and constructs a candidate set of fields. Contains structured fields and unstructured embedded fields ;

[0044] Semantic matching and mapping module: Based on generative large model for Q and Perform semantic matching to generate an intent field mapping matrix M, where Intention words With fields The semantic relevance probability value;

[0045] Field dependency graph construction module: Based on M, construct a field dependency graph G=(V,E), where node V corresponds to the identified highly relevant fields, and edge E represents the logical dependencies between fields; perform graph traversal on G to extract the minimum closed-loop field set. , as the field group required for the current query;

[0046] Field credibility assessment module: for Each field Based on the structural characteristics and data integrity indicators of its source data Perform field credibility assessment to obtain a weighted field set. τ is the credibility threshold;

[0047] Query template building module: The fields in the model are used as input features to construct a dynamic query template TQ, where TQ is an intermediate query expression generated by the generative large model based on context autoregression.

[0048] Field Origin Module: Executes the TQ template generation process, outputs the answer text R, and attaches the field origin path SR. The SR records the field source and data value corresponding to each answer fragment.

[0049] Reverse questioning control module: If R contains missing fields or low confidence markers, it automatically constructs a reverse question Q' until the confidence requirement is met.

[0050] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0051] 1. This invention achieves deep semantic understanding and structured mapping between natural language queries and real estate target cost data by combining generative large language models with domain data structure modeling. Compared to traditional question-answering systems based on template or keyword matching, this invention can automatically identify user query intent, dynamically match multi-source heterogeneous data fields, and combine field semantic tags and contextual information to generate intermediate query expressions with logical structures, greatly improving the processing capability and automation level of complex semantic queries.

[0052] 2. This invention innovatively introduces a field credibility assessment and field tracing mechanism to ensure that the source of the answer is traceable, the structure is interpretable, and the quality is controllable. In the case of missing data or low credibility, the system can automatically trigger reverse questioning and adaptively repair information gaps, possessing high interactive robustness and intelligent completion capabilities. Therefore, this invention is particularly suitable for intelligent decision support scenarios in the real estate field where fields are diverse, data is complex, and semantic relationships are strong. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0054] Figure 1 This is a flowchart of the method of the present invention.

[0055] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example 1, please refer to Figure 1 As shown in this embodiment, a method for intelligent querying of decision data based on generative large models includes:

[0058] Receive a user query Q in natural language form, and extract the key intent words I and the set of target fields involved. ;

[0059] Extract all structured and unstructured target cost data sources from the real estate project database D, and construct a candidate set of fields. Contains structured fields and unstructured embedded fields ;

[0060] Based on generative large models, Q and Perform semantic matching to generate an intent field mapping matrix M, where Intention words With fields The semantic relevance probability value;

[0061] A field dependency graph G=(V,E) is constructed based on M, where node V corresponds to the identified highly relevant fields, and edge E represents the logical dependencies between fields. Graph traversal is performed on G to extract the minimum closed-loop field set. , as the field group required for the current query;

[0062] right Each field Based on the structural characteristics of its source data and the data integrity index δᵢ, field credibility is assessed to obtain a weighted field set. τ is the credibility threshold;

[0063] Will The fields in the model are used as input features to construct a dynamic query template TQ, where TQ is an intermediate query expression generated by the generative large model based on context autoregression.

[0064] The TQ template generation process is executed, and the answer text R is output, along with the field source path SR, which records the source and data value of each field corresponding to each answer fragment.

[0065] If R contains missing fields or low confidence markers, a reverse question Q' is automatically constructed until the confidence requirement is met.

[0066] In this embodiment of the invention, a query statement Q in natural language form, input by the user, is first received. The query statement Q typically contains the user's specific information requirements regarding the target cost data of a real estate project, and is an intelligent query input geared towards management decision-making. For example, the user might input questions with practical semantics such as: "What percentage of this project is labor cost?" or "Is the trend of total cost affected by fluctuations in material prices?"

[0067] Because this type of query is characterized by flexible expression, complex semantic hierarchy, and ambiguous field references, traditional keyword matching methods struggle to accurately understand its intent and data mapping relationships. Therefore, this invention introduces a generative large language model (such as a Transformer-type pre-trained model) as the core language understanding engine to perform semantic parsing on query Q.

[0068] Specifically, the Q object is first preprocessed with word segmentation, syntactic analysis, and dependency structure analysis to construct a semantic representation vector. Then, a large language model is invoked to perform intent recognition on the Q object, extracting a set of key intent words related to the data fields. , where n is the total number of key intent words; these key intent words are used to characterize the user's inquiry purpose. For example, in "labor cost percentage", keywords might include "labor cost", "percentage", "total cost", etc.

[0069] Next, combining the existing field dictionary or data tagging system in the real estate project database, and using semantic vector matching and contextual semantic expansion methods, high-dimensional similarity matching is performed on each intent word in Q with existing fields in the database to identify a preliminary set of target fields. , m is the total number of target fields, the Each field in the table represents a data field that is likely highly relevant to the user's query target.

[0070] In one embodiment of the present invention, in order to accurately respond to the user's query intent, it is necessary to comprehensively extract data related to the target cost from the real estate project database D to construct a complete set of candidate fields. The set of fields is used to support the subsequent semantic mapping and the establishment of field dependency graphs. The real estate project database D is a comprehensive information platform containing multi-source heterogeneous data, typically composed of the following data:

[0071] Structured data sources, such as project budget sheets, contract cost details sheets, payment schedule sheets, tender control price lists, unit price sheets, and quantitative data from construction logs;

[0072] Unstructured data sources, such as cost analysis reports in PDF format, interim review documents in Word format, funding instructions transmitted in emails, and material lists identified from images;

[0073] Semi-structured data sources, such as formatted Excel lists, HTML reports, and system-exported data logs.

[0074] First, all accessible data content in D is automatically extracted using a pre-defined data interface or file parsing module. Based on this, the extracted data fields are categorized according to their source and parsability, based on the structural characteristics of the data source:

[0075] Structured field collection This collection includes fields derived from standardized data tables, such as "labor costs," "material costs," "total cost," "target profit margin," "project net present value," and "contract payment milestone number," with clearly defined field names, data types, and units, allowing for direct use in query matching and field mapping. It can be obtained through parsing the field names of the data table or extracting metadata.

[0076] Unstructured embedded field collection This collection originates from key field information in unstructured documents, typically not presented as explicit fields but embedded within natural language descriptions or visual objects. Based on technologies such as OCR recognition, text summarization, named entity recognition (NER), and domain keyword recognition, combined with a pre-trained language model, it performs deep analysis of document content, automatically extracting text fragments that may contain field meanings and assigning them semantic labels. For example, it extracts the "labor costs" field and its value "2.8 million yuan" from the sentence "The total labor costs for this project amounted to 2.8 million yuan."

[0077] Each field in the dataset also needs to undergo structured transformation to generate corresponding embedded field tags, semantic categories, and data source identifiers to support subsequent unified semantic modeling and credibility assessment. Specifically, to ensure consistency of fields with different formats within the same field space, the system introduces a field standardization process at this stage. and The fields in the dataset undergo name normalization, unit conversion, and time stamp binding. Finally, the above process constructs a complete set of candidate fields. ,in: , In the formula, It is an empty set; Each field in the dataset includes auxiliary information such as its semantic tag, source type, and data credibility level.

[0078] In this embodiment of the invention, to achieve a deep semantic association between natural language queries and real estate cost data fields, a semantic matching method based on a generative large language model is adopted to match the intent words in the user query Q with the field candidate set. Each field in the matrix is ​​encoded and matched to ultimately generate an intent field mapping matrix M. The process specifically includes the following steps:

[0079] First, the natural language query Q input by the user is fed into a pre-trained large language model to obtain its semantic representation vector. The large language model can be a pre-trained language model based on the Transformer architecture, such as BERT or RoBERTa. Preferably, this embodiment uses a domain-fine-tuned BERT model to give it professional knowledge of real estate industry corpora.

[0080] Specifically, the query Q is segmented into words and sentences, and then input into the encoder of the BERT model. The [CLS] tag vector of the output layer is extracted as the overall semantic representation vector, denoted as VQ. This vector VQ is a high-dimensional dense vector of dimension d (e.g., d=768), used to represent the overall semantic features of the user query.

[0081] For the candidate set of fields any field in To improve semantic matching accuracy, a multi-dimensional information splicing strategy is adopted to construct the field description input text, including: field name, data fragments from which the field originates (such as sample data or context statements), and semantic tags for the field (such as "cost category", "resource cost", "time type field", etc.).

[0082] By field For example, if the value is "labor costs", the concatenated content could include: "Field name: labor costs; Example data: 285,000 yuan; Category: direct costs". This concatenated text is then input into the pre-trained language model, and its [CLS] tag output is extracted as the field semantic description vector. .

[0083] This method can be used to obtain The vector representation of each field in the set Each vector dimension is consistent with VQ, possessing a unified embedding semantic space.

[0084] To achieve semantic matching between the query vector and the field vector, cosine similarity is used as the similarity metric. For the query vector VQ and the field vector... The similarity score between VQ and VQ is defined as follows: Let each be represented as a d-dimensional vector. The cosine similarity is defined as: Similarity = (VQ and ...) (dot product) divided by (VQ norm multiplied by) (norm of); that is: .

[0085] Considering that a user query Q may contain multiple intent words or semantic fragments (e.g., "labor costs", "percentage", "total cost"), the keywords or phrases in Q are represented as sets during the preprocessing stage. And for each intent word Encode into sub-vectors respectively Then, respectively with the field vector Calculate the cosine similarity.

[0086] Finally, an intent field mapping matrix M is constructed, which is an n-row × m-column matrix, where the element in the j-th row and k-th column is... Intention words With fields The semantic correlation probability value between them. This probability value is obtained after normalizing the cosine similarity above, and its range is [0, 1].

[0087] In this embodiment of the invention, after generating the intent field mapping matrix M, a field dependency graph G=(V,E) is further constructed based on the matrix, and the minimum closed-loop field set used to respond to the current query is extracted by graph traversal. The process includes the following steps:

[0088] First, the intent field mapping matrix M is filtered to extract all words that match the user's query intent. Fields with semantic relevance scores greater than a preset threshold This threshold is a field semantic relevance filtering parameter, typically set as a floating-point number between 0.65 and 0.85, with a preferred value of 0.70. The above score is derived from the aforementioned cosine similarity normalization result.

[0089] Fields that meet the above conditions Add it to the node set V of graph G, that is: if Then ∈V. Each field in node V represents a highly semantically relevant field that may be involved in the current user query.

[0090] Next, based on the co-occurrence relationships and semantic label dependencies of fields in historical real estate project data, a set of directed edges E is constructed between fields. Each edge e(i,j)∈E represents a field. Pointer field There is some kind of logical or computational dependency.

[0091] The dependency strength between fields is calculated by edge weight w(i,j) as follows: First, the field pairs are statistically analyzed. and The frequency of co-occurrence in historical data forms, reports, or computational models is p(i,j); secondly, the semantic dependency weight λ(i,j) is defined by combining the semantic label types of the two fields (e.g., "labor costs" is a sub-category field and "total cost" is a general category field), where the smaller the label hierarchy difference, the higher the value of λ.

[0092] The final edge weight calculation method is as follows: , representing the reciprocal of the dependency strength, with smaller values ​​indicating stronger dependencies. Each edge in E contains attributes such as direction, edge weight, and historical origin, forming a directed weighted edge set of graph G.

[0093] After graph G is constructed, the weighted depth-first search (W-DFS) algorithm is executed. Starting from any node in V, all reachable nodes are recursively visited, and the traversed paths and cumulative edge weights are recorded.

[0094] A path is considered a closed loop if its starting and ending nodes coincide (i.e., the path's start and end points are the same). The set of all identifiable closed loop paths is denoted as […]. Each path It consists of several field nodes, and there are directed dependency paths between the fields.

[0095] For each path in the path set P Calculate its path weight , which is the sum of the edge weights of all edges in the path, i.e.: , where (i,j)∈ .

[0096] Select the path with the minimum weight from all closed-loop paths. The path P_min, the set of field nodes involved in the path, is the minimum closed-loop field set required for the current query. ,Right now: This field collection It boasts the highest information closure and the lowest field dependency cost, effectively supporting subsequent field credibility assessment and query result generation, and avoiding redundant field interference or information loss.

[0097] In a preferred embodiment of the present invention, based on the aforementioned minimum closed-loop field set fields in Perform a field credibility assessment to remove low-quality or incomplete fields, ensuring the reliability of the data supporting the output answer. This process includes the following steps:

[0098] First of all, Each field This involves identifying the structured features of the data in the original data source and extracting parameters, including but not limited to the following three dimensions:

[0099] Data type consistency: The data type matching rate of a statistical field in historical records. For example, if a field is defined as numeric, and more than 90% of the samples match this type, the consistency is 0.9.

[0100] Field fill rate: The percentage of non-null values ​​in a field out of the total number of samples, used to reflect the coverage of a field in practical applications;

[0101] Time update frequency: The frequency with which this field has been updated or reassigned in the most recent N items in the database, used to measure its currentity and activity level.

[0102] All the above features are expressed in numerical form and uniformly normalized to the interval [0, 1] to form fields. . Structural feature vectors.

[0103] Based on the above structural characteristics, and combined with the actual content quality of the field data, the field is calculated. Data integrity metrics . Taking into account the following three factors:

[0104] Field missing rate This is defined as the proportion of a field that is missing in its data source; a higher value indicates a lower degree of completeness.

[0105] Outlier ratio This refers to a field falling within an outlier range in historical samples (e.g., a Z-score greater than 3 or less than 3). The proportion of the value of );

[0106] Source document trust level Based on the original document type to which the field data belongs, for example, the credibility level of "Approved Budget Table" is higher than that of "Provisional Estimation Document," and can be represented by an empirical score (such as 0.9, 0.7, 0.5). The field integrity index δᵢ is calculated using the following weighted formula: ;in, For the weighting coefficients, satisfying The preferred configuration is w1=0.4, w2=0.3, w3=0.3. The value ranges from 0 to 1; a larger value indicates higher field quality and better completeness. A confidence threshold τ is set, preferably between 0.6 and 0.85, to define the lower limit of a confident field. [Add all] Fields greater than or equal to τ Marked as "Trusted Field". That is: if ,but .

[0107] All satisfied The fields form a new field set. This represents a subset of fields that can be used to generate responses in the current round of semantic queries. Each field in F3 can be accompanied by its corresponding confidence value δᵢ, which serves as a weighting factor for fields in subsequent response generation, increasing the model's attention to confident fields. That is: ,in For fields The integrity index is τ, where τ is the credibility threshold.

[0108] In this embodiment of the invention, in order to achieve effective utilization of highly reliable fields in a generative question-answering system, it is necessary to base it on a weighted field set. A dynamic query template (TQ) is constructed to guide the generative large language model in referencing and expressing structured data, thereby improving the accuracy and controllability of question answering. The TQ is an intermediate query expression generated autoregressively by the generative large model under context-driven conditions, and its construction process includes the following steps:

[0109] First, from Each field in Extracting key attribute information from it is used to construct a semantic expression structure, which includes the following four types of attributes:

[0110] Field names, such as "labor costs" and "target profit margin";

[0111] Unit information, such as "yuan", "%", "ten thousand yuan / square meter", etc.;

[0112] Field context semantic tags indicate the semantic category of the field in the data structure (such as cost, ratio, time, total, etc.).

[0113] The field credibility index δᵢ represents the completeness level of the field in the credibility assessment process, and its value ranges from 0 to 1.

[0114] The above attributes are concatenated into a structured field description fragment, for example: "Field: Labor Costs, Unit: Yuan, Tag: Direct Costs, Credibility: 0.91". Using natural language processing, the structured descriptions of all fields in F3 are concatenated into an ordered sequence of field semantic vectors S_F, which is used as input to the subsequent model.

[0115] The user's original query Q and the semantic vector sequence S_F of the above fields are used as input and concatenated into a joint input statement using a specific template structure. For example: "User query: What percentage of the total cost is labor cost?; Candidate field information: [Field: Labor cost, Unit: Yuan, Tag: Direct cost, Confidence: 0.91], [Field: Total cost, Unit: Yuan, Tag: Total item, Confidence: 0.94] …".

[0116] The joint input text is fed into a pre-trained generative large language model for context embedding representation processing. The model performs embedding encoding on the joint input, generating a set of context vectors that represent the relevant path relationships between the original query intent and field semantics, providing an embedding basis for subsequent template generation.

[0117] After context embedding is completed, the autoregressive generation mechanism of the language model is used, that is, the output of the previous time step is used as the input of the next time step, word-by-word prediction is performed to dynamically generate an intermediate query expression TQ with structure and logical relationship.

[0118] The TQ (Transcription Query) does not directly output the answer. Instead, it serves as an intermediary for the model after understanding the logical path between the user's query intent and the fields, presenting the query process in structured natural language or pseudocode. For example: "Calculation expression: Labor cost ÷ Total cost × 100, unit is percentage; Field source: [Labor cost budget list.xlsx, line 3], [Total cost contract summary.pdf, page 1]". The TQ may include the following: field aliases, logical operation relationships between fields, confidence ranking of referenced fields, expected unit of result, and field tracing path, etc.

[0119] In a preferred embodiment of the present invention, to achieve controllable generation of natural language response text R and transparent tracing of field origins, the entire process of response generation and field tracing is executed based on the aforementioned generated dynamic query template TQ. The specific steps are as follows:

[0120] The intermediate query expression TQ is taken as input and fed into the decoder module of the generative large language model. The module adopts an autoregressive generation mechanism to predict the next word or phrase step by step based on the generated context output, and finally outputs a complete natural language response text R.

[0121] Generative models preferentially utilize pre-trained language models with similar structures, possessing contextual understanding and structured text expression capabilities. The model, based on the field operation relationships and logical templates described by TQ, references a set of highly reliable fields. Based on this, output answers that are factually consistent and semantically coherent.

[0122] While generating R, the field reference records in the template TQ are queried to trace the field references for each numerical fragment, conclusive description, or quantitative / qualitative judgment in the answer, constructing a field source path (SR). The SR is a structured data tracking table where each record corresponds to a content unit in the answer text and includes the following fields: a unique identifier for the corresponding field (e.g., field ID or field name); the file path and page / line number of the field's source (e.g., "Budget Summary.xlsx, line 5"); the data value or field result used (e.g., "Labor Costs: 280,000 yuan"); and the field credibility δᵢ and reference weight (used for result auditing). For example, the SR record corresponding to "Labor Costs are 280,000 yuan" generated in R is: "Field: Labor Costs; Source: Project Budget Table.xlsx, line 8; Data Value: 280,000 yuan; Credibility: 0.91".

[0123] Insert field labels and source descriptions into the R response text. For example, automatically add parenthetical notes where data values ​​are involved, such as: "The labor cost for this project is 280,000 yuan [Field: Labor Cost, Source: Budget List.xlsx, Line 8]".

[0124] At the same time, each semantic fragment in R is bound to its corresponding S_R record to generate a structured question-answer pair, including: question, answer text fragment, field reference, data source, credibility level, etc., which facilitates subsequent auditing, reuse or training feedback.

[0125] Finally, the natural language answer R is bound to its field tracing path SR and output to form a complete question-and-answer response result with traceability and clear semantics.

[0126] In a preferred embodiment of the present invention, to address the potential issues of missing fields or insufficient field credibility in the initially generated response text R, an automatic reverse questioning mechanism is introduced. This mechanism dynamically constructs supplementary questions Q′ and triggers a new round of field semantic recognition and credibility assessment until all fields referenced in the output result meet the credibility requirements. This process includes the following steps:

[0127] After generating the response text R, the system automatically parses its associated field tracing path SR and analyzes the credibility of each field. With structural integrity.

[0128] The following two types of conditions trigger a follow-up question:

[0129] Field missing: Field content exists in R that is presented as a placeholder, a vague expression (such as "incomplete data"), or "unknown", but the corresponding field in SR has no value or no source;

[0130] Insufficient field credibility: Fields in records in SR If the value is less than the preset confidence threshold τ, preferably with τ ranging from 0.70 to 0.85, then the field is marked. If it is an abnormal field, proceed to the reverse questioning construction process.

[0131] For each exception field Extract the field names, reasons for missing values ​​(e.g., no value, outlier, insufficient credibility), and source data identifiers, and combine them with the intent word set I in the original user query Q to construct a reverse question template. For example: the original query Q is: "What is the percentage of labor costs in this project?"; if the "total cost" field is missing or δ<τ, then the reverse question Q′ is generated: "Please supplement the total cost information of this project for calculating the percentage of labor costs."

[0132] The construction of the reverse question template relies on the domain language template library and contextual guidance logic. It dynamically generates supplementary questions in natural language form through the prompt learning ability of the generative large language model, which preserves the original semantics while highlighting data gaps, ensuring that Q′ has contextual consistency and clear query direction.

[0133] Use the automatically generated Q′ as new input to assess the field's credibility. Recalculate and compare with the threshold τ; if the field still exists... Then, Q′ is constructed repeatedly, allowing the system to iterate for up to N rounds of reverse questioning, preferably N is 3 to 5 rounds.

[0134] When all fields satisfy If the maximum number of follow-up questions N is reached, the reverse follow-up questioning process will terminate, and the latest generated answer text R and field tracing path SR will be returned.

[0135] Example 2, please refer to Figure 2 As shown in this embodiment, a decision data intelligent query system based on a generative large model includes:

[0136] Query Intent Parsing Module: Receives user queries Q in natural language format, extracts key intent words I and the set of target fields involved. ;

[0137] Multi-source data extraction module: Extracts all structured and unstructured target cost data sources from the real estate project database D, and constructs a candidate set of fields. Contains structured fields and unstructured embedded fields ;

[0138] Semantic matching and mapping module: Based on generative large model for Q and Perform semantic matching to generate an intent field mapping matrix M, where Intention words With fields The semantic relevance probability value;

[0139] Field dependency graph construction module: Based on M, construct a field dependency graph G=(V,E), where node V corresponds to the identified highly relevant fields, and edge E represents the logical dependencies between fields; perform graph traversal on G to extract the minimum closed-loop field set. , as the field group required for the current query;

[0140] Field credibility assessment module: for Each field Based on the structural characteristics and data integrity indicators of its source data Perform field credibility assessment to obtain a weighted field set. τ is the credibility threshold;

[0141] Query template building module: The fields in the model are used as input features to construct a dynamic query template TQ, where TQ is an intermediate query expression generated by the generative large model based on context autoregression.

[0142] Field Origin Module: Executes the TQ template generation process, outputs the answer text R, and attaches the field origin path SR. The SR records the field source and data value corresponding to each answer fragment.

[0143] Reverse questioning control module: If R contains missing fields or low confidence markers, it automatically constructs a reverse question Q' until the confidence requirement is met.

[0144] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent querying of decision data based on generative large models, characterized in that: include: Receive a user query Q in natural language form, and extract the key intent words I and the set of target fields involved. ; Extract all structured and unstructured target cost data sources from the real estate project database D, and construct a candidate set of fields. Contains structured fields and unstructured embedded fields ; Based on generative large models, Q and Perform semantic matching to generate an intent field mapping matrix M, where Intention words With fields The semantic relevance probability value; A field dependency graph G=(V,E) is constructed based on M, where node V corresponds to the identified highly relevant fields, and edge E represents the logical dependencies between fields. Graph traversal is performed on G to extract the minimum closed-loop field set. , as the field group required for the current query; right Each field Based on the structural characteristics and data integrity indicators of its source data Perform field credibility assessment to obtain a weighted field set. τ is the credibility threshold; Will The fields in the model are used as input features to construct a dynamic query template TQ, where TQ is an intermediate query expression generated by the generative large model based on context autoregression. The TQ template generation process is executed, and the answer text R is output, along with the field source path SR, which records the source and data value of each field corresponding to each answer fragment. If R contains missing fields or low confidence markers, a reverse question Q' is automatically constructed until the confidence requirement is met.

2. The intelligent query method for decision data based on generative large models according to claim 1, characterized in that: in, The generative large model-based Q and Perform semantic matching to generate an intent field mapping matrix M, including: Context encoding of Q is performed using a pre-trained large language model to generate the corresponding query vector representation VQ; right Each field in Constructing semantic description vectors The It is obtained by concatenating field names, source data fragments, and semantic tags, and then encoding them using a language model. VQ is calculated based on cosine similarity with each The matching scores between them are used to construct the intent field mapping matrix M.

3. The intelligent query method for decision data based on generative large models according to claim 1, characterized in that: in, Perform a graph traversal on G to extract the minimum closed-loop field set. The fields required for the current query include: Fields in M ​​with scores higher than a preset threshold τ Add the node set V to graph G, where τ is the field semantic relevance threshold; Based on the co-occurrence frequency and semantic label relationship of fields in the project's historical data, a set of directed edges E between fields is established, and the edge weight represents the dependency strength between fields; Perform a weighted depth-first traversal algorithm on graph G to identify the set of closed-loop paths formed between fields; The minimum closed-loop field set is obtained by selecting the field-closed-loop path with the minimum sum of dependent edge weights. This is a combination of fields used to meet the semantic requirements of the current query.

4. The intelligent query method for decision data based on generative large models according to claim 1, characterized in that: right Each field Based on the structural characteristics and data integrity indicators of its source data The field credibility is evaluated to obtain a weighted field set, including: Extract each field The corresponding source data structure characteristic parameters include data type consistency, field fill rate, and time update frequency; The field data integrity index δᵢ is calculated, which is obtained by weighting the field missing rate, the proportion of outliers, and the trust level of the source file. Fields are filtered based on a preset confidence threshold τ. If the value is greater than or equal to τ, the field is considered a reliable field. All satisfied The fields form a weighted field set .

5. The intelligent query method for decision data based on generative large models according to claim 1, characterized in that: Will Using the fields in the table as input features, a dynamic query template TQ is constructed, including: extract Each field Field names, units, field context semantic tags, and their corresponding credibility. Construct a sequence of field semantic vectors; Generative large language models are used to jointly encode the user's original query Q and the sequence of field semantic vectors to form a contextual embedding representation; Based on the autoregressive language modeling mechanism, an intermediate query expression TQ is generated.

6. The intelligent query method for decision data based on generative large models according to claim 1, characterized in that: in, Execute the TQ template generation process, output the answer text R, and attach the field tracing path SR, including: The constructed intermediate query expression TQ is input into the generative large model decoder, and the response text R in natural language form is output based on the context autoregressive generation mechanism. While generating R, record the unique identifier, source file path and field value of the referenced field corresponding to each semantic fragment, and construct the field source path SR; Add corresponding field labels and data source descriptions to each result value or description in R to form a structured question-and-answer pair; Bind R and SR to output a complete answer with interpretable field mappings.

7. The intelligent query method for decision data based on generative large models according to claim 1, characterized in that: If R contains missing fields or low confidence markers, a follow-up question Q' is automatically constructed until the confidence requirement is met, including: Parse the response text R and the field tracing path SR to detect whether there are missing fields or field confidence issues. Cases where the confidence threshold τ is less than the following; Extracting missing or low-confidence fields And based on the original query intent, a supplementary reverse question template is constructed to generate reverse questions. ; Will The input is fed into the generative large language model, triggering a new round of field semantic matching and credibility evaluation processes; until all generated fields meet the requirements. .

8. A decision data intelligent query system based on generative large models, used to implement the decision data intelligent query method based on generative large models as described in any one of claims 1-7, characterized in that: include: Query Intent Parsing Module: Receives user queries Q in natural language format, extracts key intent words I and the set of target fields involved. ; Multi-source data extraction module: Extracts all structured and unstructured target cost data sources from the real estate project database D, and constructs a candidate set of fields. Contains structured fields and unstructured embedded fields ; Semantic matching and mapping module: Based on generative large model for Q and Perform semantic matching to generate an intent field mapping matrix M, where Intention words With fields The semantic relevance probability value; Field dependency graph construction module: Based on M, construct a field dependency graph G=(V,E), where node V corresponds to the identified highly relevant fields, and edge E represents the logical dependencies between fields; perform graph traversal on G to extract the minimum closed-loop field set. , as the field group required for the current query; Field credibility assessment module: for Each field Based on the structural characteristics and data integrity indicators of its source data Perform field credibility assessment to obtain a weighted field set. τ is the credibility threshold; Query template building module: The fields in the model are used as input features to construct a dynamic query template TQ, where TQ is an intermediate query expression generated by the generative large model based on context autoregression. Field Origin Module: Executes the TQ template generation process, outputs the answer text R, and attaches the field origin path SR. The SR records the field source and data value corresponding to each answer fragment. Reverse questioning control module: If R contains missing fields or low confidence markers, it automatically constructs a reverse question Q' until the confidence requirement is met.

Images