Construction method of intelligent question answering system based on lightweight large model

Abstract

The invention discloses a construction method of an intelligent question answering system based on a lightweight large model, and relates to the technical field of natural language processing, and the method comprises the steps: receiving a natural language question of a user, carrying out vector coding through a lightweight BERT model, calculating the cosine similarity of the natural language question and a business index vector, and generating a structured semantic map; converting the structured semantic map into a scene feature vector, and injecting the scene feature vector into an adapter parameter block to construct a lightweight scene adaptation model; the lightweight scene adaptation model is combined with a real-time data interface to obtain a structured multi-modal response and construct an index tracking tree; and based on the unexpanded nodes of the index tracking tree, actively initiating scene migration type questions through a questioning strategy engine, obtaining target scene feature vectors, performing deep analysis, and generating a deep question and answer analysis report. According to the method, the lightweight scene adaptation model is constructed, calculation logic is flexibly adjusted according to different service scene features, and the self-adaptive processing capability of a single model to multiple service scenes is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a method for constructing an intelligent question-answering system based on a lightweight large model. Background Art

[0002] In the field of intelligent question-answering systems, natural language processing technologies based on large models have become a current research hotspot. Traditional approaches typically use general-purpose pre-trained language models (such as BERT and GPT) to directly process user queries and generate responses in an end-to-end manner. These approaches rely on large-scale pre-training and fine-tuning techniques. Language models trained on massive corpora possess strong semantic understanding capabilities and are capable of handling a wide range of natural language questions. In real-world business scenarios, structured business data is obtained by connecting to enterprise databases, and responses are generated using keyword matching or simple template techniques. Existing technical solutions have performed well in general-purpose question-answering tasks, and some commercial systems have already implemented basic functions such as business process query and data report generation.

[0003] However, traditional methods have limitations in specialized question-and-answer scenarios targeting vertical domains. For one thing, general-purpose language models struggle to deeply understand the complex relationships between business metrics, resulting in answers that lack professionalism and interpretability. Furthermore, the large number of model parameters leads to high inference latency, making it difficult to meet the demands of business scenarios with strict real-time requirements. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method for constructing an intelligent question-answering system based on a lightweight large model to solve the problems of insufficient understanding of deep correlation of business indicators and low efficiency of real-time reasoning.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides a method for constructing an intelligent question-answering system based on a lightweight large model, which includes receiving a user's natural language question, using a lightweight BERT model for vector encoding, and calculating the cosine similarity between the natural language question and the business indicator vector to generate a structured semantic graph; Convert the structured semantic graph into a scene feature vector and inject the adapter parameter block to build a lightweight scene adaptation model; A lightweight scenario adaptation model combined with a real-time data interface obtains structured multimodal responses and builds an indicator tracking tree; Based on the unexpanded nodes of the indicator tracking tree, the questioning strategy engine actively initiates scenario migration questions, obtains the target scenario feature vector and conducts in-depth analysis to generate an in-depth question and answer analysis report.

[0007] As a preferred solution of the method for constructing an intelligent question-answering system based on a lightweight large model described in the present invention, wherein: the lightweight BERT model is used for vector encoding, and the steps are as follows: BERT-bas is compressed through soft target distillation and intermediate layer supervision to generate a distilled lightweight BERT model, which is then trained using historical business data. Use the trained lightweight BERT model to encode user natural language questions and generate business semantic embedding vectors.

[0008] As a preferred solution of the method for constructing an intelligent question-answering system based on a lightweight large model described in the present invention, the steps of calculating the cosine similarity between the natural language question and the business indicator vector to generate a structured semantic graph are as follows: Extract business indicator vectors from the enterprise structured data table, calculate the cosine similarity between the natural language question and the business indicator vector, and select the top L business indicator vectors with the highest similarity as high correlation indicators; The semantic association rules between highly correlated indicators are generated through gradient boosting decision tree fitting, and a structured semantic graph is constructed.

[0009] As a preferred solution of the method for constructing an intelligent question-answering system based on a lightweight large model described in the present invention, the steps of converting the structured semantic graph into a scene feature vector and injecting the adapter parameter block to construct a lightweight scene adaptation model are as follows: A scene vector encoder is built based on a three-layer Transformer, which converts the structured semantic graph into a scene feature vector. Extract training samples from historical business data and initialize the adapter parameter block structure; According to the training samples, the contrastive learning loss function is used to train the initialized adapter parameter block structure to generate the adapter parameter block; Based on the scene feature vector, the adapter parameter block is integrated with the LiteBERT architecture through the parameter plug-in interface to generate a lightweight scene adaptation model.

[0010] As a preferred solution of the method for constructing an intelligent question-answering system based on a lightweight large model according to the present invention, the steps of obtaining a structured multimodal response are as follows: Separate the spatial dimension weights and temporal dimension identities in the scene feature vector; Set up the RPC protocol interface based on the spatial dimension weight and time dimension identifier; The lightweight scene adaptation model receives business data matching the current scene feature vector in real time through the RPC protocol interface, and obtains structured multimodal responses in combination with semantic association rules.

[0011] As a preferred solution of the method for constructing an intelligent question-answering system based on a lightweight large model described in the present invention, the construction of the indicator tracking tree refers to taking the text conclusion in the structured multimodal response as the root node and expanding it layer by layer according to the business indicator association path.

[0012] As a preferred solution of the method for constructing an intelligent question-answering system based on a lightweight large model according to the present invention, the identification process of the unexpanded nodes of the indicator tracking tree is as follows: Use the depth-first search algorithm to traverse the indicator tracking tree and identify the coverage of the business indicator association path in the leaf node; Define coverage thresholds and compare them with the coverage of paths associated with business indicators to identify unexpanded nodes in the indicator tracking tree.

[0013] As a preferred solution of the method for constructing an intelligent question-answering system based on a lightweight large model described in the present invention, wherein: the questioning strategy engine actively initiates scene migration questions, obtains the target scene feature vector and performs in-depth analysis, and generates an in-depth question-answering analysis report, the steps are as follows: Through the questioning strategy engine, we proactively initiate scenario-transfer questions to the unexpanded nodes of the indicator tracking tree, obtain natural language question suggestions with target scenario annotations, re-encode them, and generate target scenario feature vectors. Based on the target scenario feature vector, the lightweight scenario adaptation model loads the adapter parameter block and generates an in-depth question and answer analysis report in combination with the business data that matches the target scenario feature vector.

[0014] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the method for constructing an intelligent question-answering system based on a lightweight large model as described in the first aspect of the present invention.

[0015] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for constructing an intelligent question-answering system based on a lightweight large model as described in the first aspect of the present invention.

[0016] The beneficial effects of the present invention are: through the gradient boosting decision tree, nonlinear association rules are automatically mined from business indicator data, discrete business indicators are converted into a graph structure with clear semantic relationships, and structured representation and explainable reasoning of business knowledge are realized; by constructing a lightweight scenario adaptation model, the calculation logic is flexibly adjusted according to the characteristics of different business scenarios, and the adaptive processing capability of a single model for multiple business scenarios is realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 The flowchart of the construction method of the intelligent question-answering system based on the lightweight large model.

[0019] Figure 2 Flowchart for lightweight BERT model training.

[0020] Figure 3 Flowchart built for structured semantic graphs.

[0021] Figure 4 Flowchart for lightweight scene adaptation model construction. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0025] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides a method for constructing an intelligent question-answering system based on a lightweight large model, comprising the following steps: S1. Receive user natural language questions, use the lightweight BERT model for vector encoding, and calculate the cosine similarity between the natural language questions and the business indicator vectors to generate a structured semantic graph; User natural language questions refer to requests expressed by business personnel in unstructured language, including diverse business intent data such as indicator query, scenario analysis, and root cause tracing.

[0026] BERT-bas is compressed through soft target distillation and intermediate layer supervision to generate a distilled lightweight BERT model, which is then trained using historical business data. Furthermore, the knowledge transfer process utilizes two methods: soft target distillation and intermediate layer supervision. Soft target distillation utilizes the probability distribution generated by BERT-base for the input text and uses the KL divergence loss to guide the lightweight BERT model to learn the semantic representation capabilities of BERT-base, making the probability distribution of the lightweight BERT model's output as close as possible to that of BERT-base. Intermediate layer supervision requires that the intermediate layer outputs of the lightweight BERT model maintain similarity with the hidden states of the corresponding layers of BERT-base. L2 loss is used to constrain the feature representations of each layer of the lightweight BERT model to align with BERT-base.

[0027] In the implementation, BERT-base is first used to process retail domain data, obtaining probability distributions and hidden states as supervisory signals. During training, the lightweight BERT model simultaneously optimizes the KL divergence loss and L2 loss, enabling it to learn the output distribution of BERT-base while matching its intermediate layer representations. Incorporating annotation information from historical business data, the downstream task loss and distillation loss are jointly optimized, allowing the lightweight BERT model to reduce parameter count while maintaining its ability to handle retail domain tasks. The resulting lightweight BERT model significantly improves inference efficiency while retaining the core functionality of BERT-base.

[0028] Historical business data includes structured business records (such as sales order tables, inventory turnover tables, customer transaction tables, etc.), as well as business indicator data extracted and vectorized using ETL tools (such as indicator vectors such as "GMV growth rate" and "inventory turnover rate").

[0029] Use the trained lightweight BERT model to encode the user's natural language questions and generate business semantic embedding vectors; Furthermore, the user's natural language question is first segmented and tokenized, converted into a word sequence and input into the lightweight BERT model; the lightweight BERT model contextually encodes the word sequence through a multi-layer Transformer structure to generate a context-related representation of each word; for the context-related representation of each word output by the lightweight BERT model, the variable-length sequence is converted into a fixed-dimensional business semantic embedding vector.

[0030] It should be noted that the business semantic embedding vector refers to the business semantic embedding vector generated by the lightweight BERT model by encoding the user's natural language questions. For example, store sales are encoded as a business semantic embedding vector containing region, store type, business indicators, etc.

[0031] Business indicator vectors are extracted from enterprise structured data tables using ETL tools. The cosine similarity between natural language questions and business indicator vectors is calculated using a semantic space alignment algorithm. The top L business indicator vectors with the highest similarity are selected using a K-nearest neighbor algorithm as high-correlation indicators. Furthermore, the ETL tool extracts business indicator vectors from the specified fields of the enterprise structured data table, and the business indicator vectors are stored in the vector database after standardization; the business semantic embedding vectors generated by the lightweight BERT model and the business indicator vectors extracted by the ETL tool are similarity calculated using the semantic space alignment algorithm. The semantic space alignment algorithm uses the cosine similarity measurement and is accelerated by the approximate nearest neighbor search technology; the K-nearest neighbor algorithm selects the top L business indicator vectors with the highest similarity to the business semantic embedding vector as high-correlation indicators based on the similarity results calculated by the semantic space alignment algorithm.

[0032] It should be noted that enterprise structured data is obtained from the normalized tables of ERP, CRM and SCM databases, including numerical indicator fields in the sales fact table, store dimension table and product dimension table.

[0033] Calculate the cosine similarity between the natural language question and the business indicator vector. The expression is: ; in, is a set of business indicator vectors, is the business semantic embedding vector, It is business indicator vectors, is the cosine similarity between the natural language question and the business indicator vector; The semantic association rules between highly correlated indicators are generated through gradient boosting decision tree fitting, and a structured semantic graph is constructed.

[0034] Furthermore, the gradient boosting decision tree uses highly correlated indicators as input features and constructs a decision tree set through multiple rounds of iterative training. Each gradient boosting decision tree learns the nonlinear relationship between indicators. During the training process, the gradient boosting decision tree uses the logarithmic loss function to optimize the node splitting rule to capture the conditional dependency between indicators. After training, frequently occurring indicator combination patterns are extracted from the decision paths of the gradient boosting decision tree (for example, "the proportion of paths where GMV growth rate and inventory turnover rate appear simultaneously is greater than 60%") and converted into semantic association rules through a rule induction algorithm. Based on the extracted semantic association rules, with the core business indicators as the root nodes and the associated indicators as the child nodes, the edge weights are determined by the rule coefficients to construct a structured semantic graph in JSON-LD format. The structured semantic graph nodes contain indicator metadata (definition, unit of measurement), and the edges store association types (derivation / dependence).

[0035] S2, convert the structured semantic graph into a scene feature vector and inject the adapter parameter block to build a lightweight scene adaptation model; A scene vector encoder is built based on a three-layer Transformer, which converts the structured semantic graph into a scene feature vector. Furthermore, the three-layer Transformer architecture receives input in the JSON-LD format of a structured semantic graph. The first layer converts the structured semantic graph (business indicator vector) and edges (semantic association rules) into a token sequence. The second layer maps each node and edge to an initial vector representation, while adding positional encoding to preserve the structural information of the structured semantic graph. The third layer uses the Transformer's multi-head attention mechanism to identify dependencies between nodes. After three layers of encoding, the output vector of the [CLS] tag is linearly transformed to generate a scene feature vector. It should be noted that the scene feature vector includes four types of semantic features: subject type weight (such as "store = 0.7"), spatial dimension intensity (such as "Hangzhou = 0.8"), time decay factor (such as "this month = -0.2") and core indicator correlation (such as "GMV growth rate = 0.82").

[0036] Extract training samples from historical business data through approximate nearest neighbor search based on locality sensitive hashing and initialize the adapter parameter block structure; Furthermore, each record in the historical business data is first converted into a scene encoding vector containing a timestamp, numerical indicators and classification indicators; the scene encoding vector is bucketed using the local sensitive hashing algorithm to extract training samples; a hash index is constructed, and historical business data with similar high correlation indicators are quickly retrieved from the hash index through approximate nearest neighbor search, and the numerical fields and classification fields in the historical business data are retrieved based on the hash index, and the adapter parameter block structure containing the attention head correction matrix, feedforward network bias item and layer normalization parameters is initialized.

[0037] According to the training samples, the contrastive learning loss function is used to train the initialized adapter parameter block structure to generate the adapter parameter block; Furthermore, based on the business scenario data in the training samples, a contrastive learning loss function is used to train the initialized adapter parameter block structure. During training, the contrastive learning loss function is used to narrow the distance between the scenario feature vectors for positive sample pairs, while the distance between the scenario feature vectors for negative sample pairs is pushed further away. Simultaneously, an optimization algorithm is used to iteratively optimize the adapter parameter block structure, ultimately generating an adapter parameter block that meets the business scenario adaptation requirements. Positive samples refer to scenario encoding vectors from the same business scenario (e.g., sales promotion diagnosis), while negative samples refer to scenario encoding vectors from different business scenarios (e.g., sales promotion diagnosis and channel optimization).

[0038] Based on the scene feature vector, the adapter parameter block is integrated with the LiteBERT architecture through the parameter plug-in interface to generate a lightweight scene adaptation model.

[0039] Furthermore, the scene feature vector is matched with the corresponding adapter parameter block through the parameter plug-in interface. The attention head correction matrix in the adapter parameter block is inserted into the multi-head attention layer of the LiteBERT architecture in the form of a low-rank matrix. The feedforward network bias term is added element-by-element to the feedforward network layer bias of LiteBERT. The layer normalization parameters replace the original parameters of the corresponding layer of LiteBERT, and finally a lightweight scene adaptation model is generated.

[0040] S3, a lightweight scenario adaptation model combined with a real-time data interface, obtains structured multimodal responses and builds an indicator tracking tree; A semantic dimension parser is used to separate the spatial dimension weights and temporal dimension identifiers in the scene feature vector; Furthermore, the semantic dimension parser receives the scene feature vector as input and extracts the spatial dimension weight and temporal dimension identifier through the dimension mask matrix (the spatial dimension mask covers dimensions 32-63, and the temporal dimension mask covers dimensions 64-95). The spatial dimension weight is normalized by Softmax to generate the regional importance distribution (such as "Hangzhou = 0.7, Shanghai = 0.3"), and the temporal dimension identifier is mapped to a temporal granularity label through a classifier.

[0041] It should be noted that the dimension mask matrix is ​​a binary matrix defined by business scenario analysis requirements (the spatial dimension position is set to 1, and the rest are set to 0).

[0042] Set up the RPC protocol interface based on the spatial dimension weight and time dimension identifier; The lightweight scenario adaptation model receives business data matching the current scenario feature vector in real time through the RPC protocol interface, and combines semantic association rules to obtain a structured multimodal response containing text conclusions and business indicator association paths; Furthermore, the lightweight scenario adaptation model takes the scenario feature vector as an input parameter and initiates a gRPC call to the business data service through the RPC protocol interface. The RPC protocol interface generates query conditions based on the spatial dimension weights and time dimension identifiers in the scenario feature vector, retrieves matching business indicator values ​​from the enterprise data warehouse (such as "GMV growth rate in Hangzhou this month = 18%"), and logically combines them with the semantic association rules generated by the gradient boosting decision tree (such as "GMV growth rate → SKU turnover rate") to generate a structured multimodal response containing natural language text conclusions (such as "sales growth mainly comes from SKU12345") and business indicator association paths (such as "GMV growth rate 18% → SKU turnover rate increased by 25%").

[0043] The text conclusion in the structured multimodal response is used as the root node, and it is expanded layer by layer according to the business indicator association path to build an indicator tracking tree.

[0044] Furthermore, the textual conclusions in the structured multimodal response are used as the root node, and branches are then expanded layer by layer according to the semantic association rules generated by the gradient boosting decision tree. Each branch node contains the name of the business indicator, its specific value, and the calculation basis. The connecting edges between nodes are annotated with the weight coefficient of the semantic association rule. Nodes that do not meet the expansion threshold (such as "distribution cycle 7 days") are marked as pending expansion and the missing associated indicators are recorded. Finally, a breadth-first traversal is used to construct the complete tree structure.

[0045] It should be noted that the expansion threshold is defined based on business scenario analysis requirements and the value range is: 0.2-0.5.

[0046] S4. Based on the unexpanded nodes of the indicator tracking tree, the questioning strategy engine actively initiates scenario migration questions, obtains the target scenario feature vector and performs in-depth analysis to generate an in-depth question and answer analysis report.

[0047] Use the depth-first search algorithm to traverse the indicator tracking tree and identify the coverage of the business indicator association path in the leaf node; Furthermore, the depth-first search algorithm recursively visits each leaf node starting from the root node of the indicator tracking tree. During the visit, the number of business indicator association paths that have been expanded in the leaf node is counted (by querying the complete semantic association rule set generated by the gradient boosting decision tree to obtain the total number of business indicator association paths that can theoretically be expanded in the leaf node (for example, "inventory turnover rate" can be associated with three paths). The coverage rate of the business indicator association paths in the leaf node is identified by calculating the ratio of the number of expanded paths in the leaf node to the total number of expandable paths.

[0048] Based on historical interaction data analysis, define coverage thresholds and compare them with the coverage of paths associated with business indicators to identify unexpanded nodes in the indicator tracking tree; Furthermore, historical interaction data analysis statistics the average expansion depth of users for indicator tracking tree nodes in different business scenarios, and defines the coverage threshold (value range: 0.15-0.6) in combination with business priority; when the depth-first search algorithm traverses the indicator tracking tree, the complete business indicator association path set generated by the gradient boosting decision tree is queried for each leaf node, and the ratio of the current number of expanded paths to the total number of expandable paths is calculated; when the business indicator association path coverage of a leaf node is lower than the coverage threshold, the current node is marked as an unexpanded node.

[0049] It should be noted that historical interaction data includes complete conversation records between users and the question-answering system (such as question text, follow-up path), scene feature vector call logs, adapter parameter block loading records, and user feedback data (explicit ratings / implicit behaviors).

[0050] Through the questioning strategy engine, we proactively initiate scenario-transfer questions to the unexpanded nodes of the indicator tracking tree, obtain natural language question suggestions with target scenario annotations, and re-encode them through the scenario vector encoder to generate the target scenario feature vector; Furthermore, the follow-up question strategy engine proactively initiates scenario-transfer questions, parses the leaf nodes marked as unexpanded in the indicator tracking tree, and combines the gradient boosting decision tree to generate natural language follow-up question suggestions with target scenario annotations (such as "Should we analyze the impact of procurement cycles on inventory turnover? [Target scenario: supply chain efficiency analysis]"). Through word segmentation processing and word vectorization, the natural language follow-up question suggestion mapping is converted into a word sequence input into the scenario vector encoder for re-encoding to generate the target scenario feature vector.

[0051] It should be noted that the target scenario refers to the next stage of analysis scenario triggered by the user's follow-up behavior (for example, migrating from "sales dynamics diagnosis" to "supply chain efficiency analysis").

[0052] Based on the target scenario feature vector, the lightweight scenario adaptation model loads the adapter parameter block and generates an in-depth question and answer analysis report in combination with the business data that matches the target scenario feature vector.

[0053] Furthermore, the target scene feature vector is matched with the corresponding adapter parameter block through the parameter plug-in interface. The attention head correction matrix in the adapter parameter block adjusts the multi-head attention mechanism weights of the lightweight scene adaptation model, so that the model focuses on the key fields of the target scene; the feedforward network bias term corrects the activation function output to enhance the representation ability of the target scene features; the lightweight scene adaptation model obtains the business data matching the target scene feature vector through the RPC protocol interface, and performs multiple rounds of reasoning based on the business logic rules in the adapter parameter block to generate a structured report containing text conclusions, data evidence and indicator association paths.

[0054] This embodiment also provides a computer device, which is suitable for the method of constructing an intelligent question-answering system based on a lightweight large model, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions, thereby implementing the method of constructing an intelligent question-answering system based on a lightweight large model proposed in the above embodiment.

[0055] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0056] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing an intelligent question-answering system based on a lightweight large model as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0057] In summary, the present invention achieves structured representation and explainable reasoning of business knowledge through: gradient boosting decision tree, automatically mining nonlinear association rules from business indicator data, converting discrete business indicators into graph structures with clear semantic relationships; by constructing a lightweight scenario adaptation model, flexibly adjusting the calculation logic according to the characteristics of different business scenarios, realizing the adaptive processing capability of a single model for multiple business scenarios.

[0058] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for constructing an intelligent question-answering system based on a lightweight large model, characterized by: include, Receive user natural language questions, use the lightweight BERT model for vector encoding, and calculate the cosine similarity between the natural language questions and business indicator vectors to generate a structured semantic graph; Convert the structured semantic graph into a scene feature vector and inject the adapter parameter block to build a lightweight scene adaptation model; A lightweight scenario adaptation model combined with a real-time data interface obtains structured multimodal responses and builds an indicator tracking tree; Based on the unexpanded nodes of the indicator tracking tree, the questioning strategy engine actively initiates scenario migration questions, obtains the target scenario feature vector and conducts in-depth analysis to generate an in-depth question and answer analysis report.

2. The method for constructing an intelligent question-answering system based on a lightweight large model according to claim 1, characterized in that: The lightweight BERT model is used for vector encoding. The steps are as follows: BERT-bas is compressed through soft target distillation and intermediate layer supervision to generate a distilled lightweight BERT model, which is then trained using historical business data. Use the trained lightweight BERT model to encode user natural language questions and generate business semantic embedding vectors.

3. The method for constructing an intelligent question-answering system based on a lightweight large model according to claim 2, characterized in that: The steps of calculating the cosine similarity between the natural language question and the business indicator vector to generate a structured semantic graph are as follows: Extract business indicator vectors from the enterprise structured data table, calculate the cosine similarity between the natural language question and the business indicator vector, and select the top L business indicator vectors with the highest similarity as high correlation indicators; The semantic association rules between highly correlated indicators are generated through gradient boosting decision tree fitting, and a structured semantic graph is constructed.

4. The method for constructing an intelligent question-answering system based on a lightweight large model according to claim 1, characterized in that: The steps of converting the structured semantic graph into a scene feature vector and injecting the adapter parameter block to build a lightweight scene adaptation model are as follows: A scene vector encoder is built based on a three-layer Transformer, which converts the structured semantic graph into a scene feature vector. Extract training samples from historical business data and initialize the adapter parameter block structure; According to the training samples, the contrastive learning loss function is used to train the initialized adapter parameter block structure to generate the adapter parameter block; Based on the scene feature vector, the adapter parameter block is integrated with the LiteBERT architecture through the parameter plug-in interface to generate a lightweight scene adaptation model.

5. The method for constructing an intelligent question-answering system based on a lightweight large model according to claim 1, characterized in that: The steps of obtaining structured multimodal response are as follows: Separate the spatial dimension weights and temporal dimension identities in the scene feature vector; Set up the RPC protocol interface based on the spatial dimension weight and time dimension identifier; The lightweight scene adaptation model receives business data matching the current scene feature vector in real time through the RPC protocol interface, and obtains structured multimodal responses in combination with semantic association rules.

6. The method for constructing an intelligent question-answering system based on a lightweight large model according to claim 5, characterized in that: The construction of the indicator tracking tree refers to taking the text conclusion in the structured multimodal response as the root node and expanding it layer by layer according to the business indicator association path.

7. The method for constructing an intelligent question-answering system based on a lightweight large model according to claim 1, characterized in that: The indicator tracks the unexpanded nodes of the tree, and the identification process is as follows: Use the depth-first search algorithm to traverse the indicator tracking tree and identify the coverage of the business indicator association path in the leaf node; Define coverage thresholds and compare them with the coverage of paths associated with business indicators to identify unexpanded nodes in the indicator tracking tree.

8. The method for constructing an intelligent question-answering system based on a lightweight large model according to claim 7, characterized in that: The following steps are used to actively initiate scenario migration questions through the questioning strategy engine, obtain the target scenario feature vector and perform in-depth analysis to generate an in-depth question and answer analysis report: Through the questioning strategy engine, we proactively initiate scenario-transfer questions to the unexpanded nodes of the indicator tracking tree, obtain natural language question suggestions with target scenario annotations, re-encode them, and generate target scenario feature vectors. Based on the target scenario feature vector, the lightweight scenario adaptation model loads the adapter parameter block and generates an in-depth question and answer analysis report in combination with the business data that matches the target scenario feature vector.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the processor implements the steps of the method for constructing an intelligent question-answering system based on a lightweight large model as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for constructing an intelligent question-answering system based on a lightweight large model as described in any one of claims 1 to 8 are implemented.

Images