Industrial innovation question and answer method and system fusing knowledge graph and large model

Abstract

The invention discloses an industrial innovation question and answer method and system fusing a knowledge graph and a large model, relates to the field of natural language processing in artificial intelligence, and solves the problems that existing industrial innovation data is dispersed, so that the utilization rate is low, and accurate decisions cannot be made due to single dependence on a large model lacking knowledge in related fields. According to the method, industrial innovation data combining patent information, academic literature and policy systems and an innovation field knowledge graph are collected, the preprocessed industrial innovation data are converted into a question and answer data set in a question and answer pair format, a sub-graph extraction model is constructed, and a large model is enhanced by utilizing the knowledge graph; and industrial trend analysis and industrial report generation are realized. According to the method, industrial innovation data are efficiently integrated and utilized, the accuracy of functions of the large model for innovation technology insight, industrial trend analysis, industrial report generation and the like is improved, and industrial technology innovation is more accurately assisted.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing in artificial intelligence, and specifically to an industrial innovation question-answering method and system that integrates knowledge graphs and large models. Background Art

[0002] Data resources such as patent information, academic literature, and policies and systems are important sources of information for promoting technological innovation and industrial development, but these data are distributed on different platforms and are difficult to systematically integrate and deeply analyze. Traditional data analysis methods are difficult to effectively tap and utilize the potential value in these data, especially in the face of a complex and ever-changing industrial environment. Therefore, there is an urgent need for a technical solution that can efficiently integrate and utilize various industrial innovation data to provide deep insights and scientific analysis for industrial innovation. Developing a question-and-answer system that combines industrial innovation data such as patent information, academic literature, and policies and systems can effectively solve the problem of data dispersion, provide key support for technological innovation and industrial development, and help related companies and institutions make accurate and forward-looking decisions.

[0003] The large language model (large model) relies on its powerful data-driven capabilities to understand and generate natural language through a large number of parameters, showing excellent language fluency and context perception. Through the large model, key information such as the development trend of the science and technology industry, technological innovation and market demand can be predicted, providing a scientific basis for related research and decision-making, and helping industrial technological innovation. However, due to the lack of a clear knowledge structure, in such tasks that require precise reasoning and professional background knowledge, the large model will produce "hallucination" phenomena due to the lack of relevant domain knowledge. And using domain knowledge to fine-tune the large model requires strong computing power support. By constructing structured entities and relationships and explicitly storing knowledge, the knowledge graph can provide logically rigorous and explainable knowledge support for the question-answering model. The present invention effectively makes up for the knowledge blind spots of the large model through retrieval enhancement, thereby more accurately helping industrial technological innovation. Summary of the invention

[0004] The present invention aims to solve the problems of existing industrial innovation data being dispersed, resulting in low utilization, and single reliance on a large model that lacks relevant field knowledge and is unable to make accurate decisions; and provides an industrial innovation question-answering method and system that integrates knowledge graphs and large models.

[0005] An industrial innovation question-answering method integrating knowledge graph and big model is implemented by the following steps: Step 1: Collect industrial innovation data that combines patent information, academic literature, and policy systems, and then pre-process them; Step 2: Use the pre-processed industrial innovation data to establish a knowledge graph in the field of industrial innovation; Step 3: Convert the preprocessed industrial innovation data into a question-answering dataset in the format of question-answering pairs, and divide the question-answering dataset into a training set, a validation set, and a test set; use the industrial innovation field knowledge graph constructed in step 2 to evaluate the quality of the question-answering pairs; Step 4: construct a subgraph extraction model, use the training set to train the subgraph extraction model to obtain a trained subgraph extraction model, and use the trained subgraph extraction model to extract knowledge subgraphs from the knowledge graph in the field of industrial innovation; convert the extracted knowledge subgraphs into text, and input the text as a prompt word into the large model; Step 5: Use the knowledge graph to enhance the big model to achieve industry trend analysis and industry report generation.

[0006] The present invention also provides an industrial innovation question-answering system integrating knowledge graph and big model, the system includes a data integration and preprocessing module, a knowledge graph creation module, a question-answering data construction module, a knowledge graph and big model fusion module and a top-level application design module; The data integration and preprocessing module is used to collect industrial innovation data that combines patent information, academic literature, and policy systems, and then perform preprocessing; The knowledge graph creation module uses the preprocessed data to establish a knowledge graph in the field of industrial innovation; The question-answer data construction module converts the preprocessed industrial innovation data into a question-answer format and inputs the data into the knowledge graph and large model fusion module as a training set, a validation set, and a test set; The knowledge graph and large model fusion module is used to train the subgraph extraction model and extract knowledge subgraphs from the domain knowledge graph; convert the extracted knowledge subgraphs into text and input the text into the large model as prompt words; The top-level application design module is used to implement industry trend analysis and industry report generation.

[0007] Beneficial effects of the present invention: The method of the present invention constructs an industrial innovation knowledge map by collecting relevant CNKI papers, patents and industry report information. The map not only covers a wide range of technical fields and industry dynamics, but also reflects the cross-influence and development trends between various fields. Through natural language processing technology and machine learning algorithms, the system can automatically identify and extract key concepts, technical innovations and potential market opportunities in the literature, thereby providing intuitive and detailed data support for researchers and decision makers.

[0008] The method described in the present invention adopts retrieval enhancement technology based on knowledge graph, and uses graph attention neural network to accurately locate and evaluate the importance of entities between subgraphs. Through this advanced algorithm, the system can intelligently parse and understand the complex relationship structure in the knowledge graph, not only identifying the association between directly connected entities, but also capturing the implicit connection in multi-hop relationships. Specifically, the method emphasizes those entities that are critical to specific queries or application scenarios by assigning different weights to nodes in the knowledge graph. The graph attention mechanism allows the model to focus on the most relevant parts, thereby improving the relevance and accuracy of the retrieval results. At the same time, this method can also effectively process large-scale knowledge graph data, ensuring that an efficient response speed can be maintained when faced with massive amounts of information.

[0009] The system described in the present invention efficiently integrates and utilizes industrial innovation data, improves the accuracy of the big model for innovative technology insights, industrial trend analysis, industrial report generation and other functions, and more accurately assists industrial technological innovation. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a flow chart of an industrial innovation question-answering method integrating knowledge graph and big model described in the present invention; Figure 2 This is a principle block diagram of an industrial innovation question-answering system integrating knowledge graph and big model described in the present invention; Figure 3 Create a module diagram for the knowledge graph; Figure 4 Schematic diagram of building a module for question-answering data; Figure 5 for Figure 1 Schematic diagram of the knowledge graph enhanced large model module in ; Figure 6 Design module diagrams for top-level applications. DETAILED DESCRIPTION

[0011] Specific implementation method 1. Combination Figures 1 to 5 This embodiment describes an industrial innovation question-answering method that integrates knowledge graph and big model. The method is implemented by the following steps: Step S1: collect industrial innovation data combining patent information, academic literature, and policy systems, and then pre-process them; In this implementation, open source data is collected using crawlers, and some closed source data is acquired through commercial cooperation, thereby achieving the aggregation of industry technology related data.

[0012] Data preprocessing generally includes the conversion from multiple formats to plain text, such as pdf, csv, doc, html, etc., as well as the cleaning and text normalization of erroneous data, redundant data, and duplicate data to ensure data quality. This includes removing special characters, extra spaces, tags and other redundant content in the text; identifying and correcting erroneous data in the text; using data processing tools and functions provided by programming languages ​​to find and delete duplicate data; and converting data such as date formats and currency units into a consistent format. The processed data will be used for knowledge graph construction and knowledge graph question and answer.

[0013] Step S2: Use the pre-processed industrial innovation data to establish a knowledge graph in the field of industrial innovation; the specific process is as follows: Step S21, extracting structured information from industrial innovation data; Extract structured information from the preprocessed industrial innovation data, including entity extraction, relationship extraction and attribute extraction. Entity extraction refers to the use of natural language processing technology to identify key entities from text data. These entities may include names of people, places, institutions, patent numbers, technical terms, etc. Relationship extraction refers to the use of text analysis, syntactic analysis, machine learning and other technical means to identify the relationship types between entities, such as "inventor", "owner", "citation", etc. Attribute extraction focuses on obtaining attribute values ​​that describe entity characteristics, such as patent application date, inventor's nationality, etc. This information will be organized into a structured form for subsequent processing.

[0014] Step S22: logically attribute and filter redundant data through knowledge fusion; The structured information extracted in step S21 is logically classified and redundantly filtered through knowledge fusion technology. In this process, the problem of entity disambiguation needs to be solved, that is, when multiple entities have the same name, further analysis is performed to determine whether they refer to the same entity. In addition, reference resolution is also required, that is, identifying pronouns or other indicators in the text and clarifying the specific entities they refer to. The logical consistency of the extracted information is ensured, and unnecessary redundant information is removed.

[0015] Step S23: Fusing information from different sources to construct a complete knowledge representation; The information processed in step S22 is integrated to construct a complete knowledge representation. This information will be organized into triples (i.e., a set consisting of a head entity, the relationship between entities, and a tail entity) and stored in the graph database Neo4j. In this way, entities and their relationships can be effectively managed to provide reliable data support for subsequent applications. In this embodiment, the constructed knowledge graph G is represented as: ; Among them, E and R represent entity set and relationship set respectively. Represents the head entity, Represents the tail entity, and r represents the relationship between entities. Each set of triples consisting of the head entity, the relationship between entities, and the tail entity together constitutes a basic unit of the recognition graph. These units are interconnected to form a complete knowledge network for storing and expressing complex relationships and information in industrial innovation data.

[0016] Step S3, converting the industrial innovation data preprocessed in step S2 into a question-answer format, and using it as a training set, a validation set, and a test set to train a subgraph extraction model; Figure 3 As shown in the figure, the pre-processed industrial innovation data is used and converted into a question-answering dataset with the help of a large model to facilitate the subsequent subgraph extraction model training and performance evaluation. The specific process is as follows: Step S31, data conversion; The industrial innovation data preprocessed in part of step S1 is manually converted into a question-answer pair format. A question-answer pair usually consists of a question and a corresponding answer. This format is conducive to the subgraph extraction model to learn how to derive the correct answer from a given question. The manually converted question-answer pairs will be used as prompts to clarify the tasks of the large model and standardize the output format of the large model.

[0017] Step S32: construct a question-answering dataset using a large model; The data in the question-answer pair format converted in step S31 is used as a sample, and the question-answer pair data is generated using the few-sample learning of the large model. The large model has powerful language understanding and generation capabilities, and can generate contextual question-answer pairs by understanding the original data and the provided question-answer pair samples. Specifically, the large model reads the input text and extracts meaningful questions and their corresponding answers from it. The generated question-answer pairs will be organized into a structured question-answer data set for subsequent subgraph extraction model training. These question-answer data sets are usually divided into training sets, validation sets, and test sets, where the training set is used to train the subgraph model, the validation set is used to evaluate the performance of the subgraph model during the training process, prevent the model from overfitting, and the test set is used to finally evaluate the performance of the subgraph model. The question-answer data set can be used to test the performance of the subgraph model when answering similar questions, help optimize the parameter settings of the model, and improve its accuracy and robustness in actual application scenarios.

[0018] Step S33, evaluate rationality: In order to ensure the rationality and accuracy of the question-answer pairs generated by the large model, the knowledge graph constructed in step S2 is used to evaluate the quality of the question-answer pairs. By comparing the logical relationship between the question and the answer in the generated question-answer pairs, and the support of the answer to the question in the knowledge graph, it is evaluated whether the generated question-answer pairs meet the expected standards. If unreasonable situations are found, the generated question-answer pairs need to be corrected or excluded.

[0019] In this implementation, the pre-processed industrial innovation data is not only converted into a question-and-answer format that is easy to understand and process, but also ensures the quality and accuracy of these question-and-answer pairs, providing high-quality data support for subsequent model training. This question-and-answer data set can not only improve the learning efficiency of the model, but also promote the in-depth understanding and effective use of industrial innovation data. For the processed question-and-answer data, 80% is extracted as a training set, 10% as a validation set, and 10% as a test set.

[0020] Step S4, construct a subgraph extraction model, use the training set to train the subgraph extraction model to obtain the trained subgraph extraction model, use the trained subgraph extraction model to extract knowledge subgraphs from the knowledge graph in the field of industrial innovation, and remove irrelevant information; convert the extracted knowledge subgraphs into text, and input the converted text into the large model as a prompt word; the specific process is: Retrieval enhancement is a technical method that combines information retrieval and generative models to improve the knowledge and accuracy of generative models, especially when dealing with problems involving a wide range of domain knowledge. The core idea of ​​retrieval enhancement is to retrieve relevant information from external data sources (knowledge bases, document bases, and Internet information) before generating answers to help generate large models to answer questions more accurately and thoroughly. Figure 4 As shown, the retrieval enhancement includes two parts: Step S41, knowledge subgraph extraction; In the knowledge graph of the industrial innovation field described in step S2, relevant partial knowledge subgraphs are extracted according to the problem entity. In this embodiment, a graph neural network is used to encode the knowledge graph.

[0021] Step S411, initial sub-graph extraction; First, according to the entity set of question q , the entity set of answer a , all and Entities in the entity set are in the knowledge graph G The two-hop neighbor entities and relations in the graph are extracted, and the cosine similarity algorithm is used to calculate the semantic similarity. The top 100 semantically relevant nodes are retained to form the initial subgraph. .

[0022] Step S412, sub-image encoding; The knowledge graph uses a graph structure as its representation, that is, each entity is represented as a node in the graph, and each relationship is represented as an edge in the graph. The graph attention algorithm (GAT) in the graph neural network is used to encode the initial subgraph, and the representation of each entity e is updated through the three-layer GAT. , the GAT algorithm is expressed as: ; in, Represents the number of layers of the graph neural network, represents all neighbor nodes of node e, represents the attention coefficient between node s and node e, Represents the information transmitted between node s and node e, is a 2-layer MLP, where n represents the number of layers of GAT, and the sum is obtained by Get and express The same graph node vector. The initial subgraph node embedding after graph neural network encoding is represented as : ; in, It is an element in the union of the question entity and the answer entity.

[0023] Step S413: entity embedding; For question entity set and answer entity set and , use the pre-trained model Bert to embed each entity to extract the entity's features. The entity embedding process is expressed as: ; In the formula, is the question entity embedding, is the answer entity embedding and PML is the pre-trained model.

[0024] Step S414, calculating the weight of the problem entity; Since different question entities have different importance to the answer, in this implementation, a bidirectional attention mechanism is designed to weight the answer entity using the representation of the question entity, thereby obtaining the importance of these question entities to the answer. Let the question entity be the query Q and the answer entity be the key K, and the attention score is used to measure the influence of the question entity on the answer entity: ; in, is the embedding dimension, softmax is the normalization function, is the problem entity weight.

[0025] Step S415, attention coefficient calculation; In order to identify the intermediate nodes most relevant to the question in the initial subgraph, the information irrelevant to the question is removed and the noise is reduced to improve the accuracy of the answer.

[0026] First, since GAT only aggregates information about neighboring nodes each time and has poor ability to capture global information, the self-attention layer of the transformer is introduced before pruning to dynamically identify the importance of nodes in the subgraph by using the internal features of the graph and the implicit interactions between nodes. After the self-attention layer, the initial subgraph node embedding is expressed as : ; Then, using the question entity embedding As query Q, As the key K, the attention mechanism is used to calculate the cross attention of the two modalities of entity embedding and graph node embedding to obtain the attention coefficient of the problem entity to other entities (neighboring entities) : ; Among them, d k is the embedding dimension, softmax is the normalization function, and a similarity result with a value range of [0,1] is obtained, which is the attention coefficient.

[0027] Step S416, pruning the knowledge subgraph; For the obtained attention coefficient , using the question entity weight Weight it to get the weighted attention coefficient of the problem entity to the surrounding entities (other entities in the entire subgraph) : ; because Only for the problem entity but not for the intermediate entity, after calculating the weighted attention of the problem entity to the first-order neighbor node, for the extended neighbor node, the attention to its neighbor node is only expressed as Entities with low attention coefficients are considered less relevant to the question and will be removed from the initial subgraph.

[0028] During the training process, each selection or The maximum number of neighbor nodes with the highest value is expanded, and the expansion depth is defined as 4. After reaching the expansion depth, it will no longer be expanded. The attention coefficient is optimized through the cross entropy loss. and , the cross entropy loss is expressed as: ; Where N represents the number of nodes in the initial subgraph, For the The labels of the nodes, For the real distribution predicted by the large model, hit@1 is used as the evaluation indicator. Finally, the entities with high relevance to the question are obtained, and the relationship between the related entities is found to form a triplet as the prompt triplet.

[0029] Step S42, prompting the learning process; The prompt triple is input to the large model as a prompt or guide, and the context learning ability of the large model is used to generate corresponding output content. In this embodiment, the extracted triple is converted into natural language as a prompt word (prompt), and is input into the large model together with the question to obtain the output answer.

[0030] Step S421: The prompt template of the triplet converted into natural language is: Prompt1 = { Please generate a natural language sentence based on the following triples: 1. (subject: "[subject1]", relation: "[relation1]", object: "[object1]"); 2. (subject: "[subject2]", relation: "[relation2]", object: "[object2]"); ...}; Step S422: After converting the triples into natural language, they are input into the big model together with the question, and the answer to the question is output. The prompt template for prompt learning is: Prompt2={ Read the following text: [Insert given text]. Answer the following questions based on the text: [Insert question]. Make sure to base your answers on the information in the text.}; Step S5: Utilize the large model and the sub-graph extraction model trained in step S4 to implement industry trend analysis and industry report generation.

[0031] In this implementation, by inputting custom questions into the system, the subgraph extraction model will retrieve the knowledge subgraph in the knowledge graph, convert the knowledge subgraph into natural language as prompts, and then input the questions and prompts into the big model, and use the question-answering function of the big model to achieve innovative technology insights, industry trend analysis, and industry report generation.

[0032] Specific implementation method 2: Combination Figure 6 This embodiment is described as a system for an industrial innovation question-answering method integrating knowledge graph and big model as described in the first embodiment. The system includes a data integration and preprocessing module, a knowledge graph creation module, a question-answering data construction module, a knowledge graph and big model fusion module, and a top-level application design module. The data integration and preprocessing module is used to collect industrial innovation data that combines patent information, academic literature, and policy systems, and then perform preprocessing; The knowledge graph creation module uses the preprocessed data to establish a knowledge graph in the field of industrial innovation; The question-answer data construction module converts the preprocessed industrial innovation data into a question-answer format and inputs the data into the knowledge graph and large model fusion module as a training set, a validation set, and a test set; The knowledge graph and large model fusion module is used to train the subgraph extraction model and extract knowledge subgraphs from the domain knowledge graph; convert the extracted knowledge subgraphs into text, and input the converted text into the large model as prompt words; The top-level application design module is used to achieve innovative technology insights, industry trend analysis, and industry report generation.

[0033] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0034] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. An industrial innovation question-answering method integrating knowledge graph and big model, characterized by: The method is implemented by the following steps: Step 1: Collect industrial innovation data that combines patent information, academic literature, and policy systems, and then pre-process them; Step 2: Use the pre-processed industrial innovation data to establish a knowledge graph in the field of industrial innovation; Step 3: Convert the preprocessed industrial innovation data into a question-answering dataset in the format of question-answering pairs, and divide the question-answering dataset into a training set, a validation set, and a test set; use the industrial innovation field knowledge graph constructed in step 2 to evaluate the quality of the question-answering pairs; Step 4: construct a subgraph extraction model, use the training set to train the subgraph extraction model to obtain a trained subgraph extraction model, and use the trained subgraph extraction model to extract knowledge subgraphs from the knowledge graph in the field of industrial innovation; Convert the extracted knowledge subgraph into text, and input the text into the big model as prompt words; Step 5: Use the knowledge graph to enhance the big model to achieve industry trend analysis and industry report generation.

2. According to claim 1, the industrial innovation question-answering method integrating knowledge graph and big model is characterized by: In step one, open source data is collected using crawlers, and closed source data is acquired through commercial cooperation to achieve the aggregation of industry technology-related data.

3. According to claim 1, the industrial innovation question-answering method integrating knowledge graph and big model is characterized by: In step 1, data preprocessing includes cleaning redundant data, erroneous data, and duplicate data, as well as data normalization; Data cleaning includes removing special characters, extra spaces, and redundant content of tags in the text; identifying and correcting erroneous data in the text; using data processing tools and functions provided by programming languages ​​to find and delete duplicate data; and converting data in date format and currency unit to a consistent format.

4. According to claim 1, the industrial innovation question-answering method integrating knowledge graph and big model is characterized by: The specific process of step 2 is: Step 21: Extract structured information from industrial innovation data, including entity extraction, relationship extraction and attribute extraction; Step 22: Logically attribute and filter redundant data through knowledge fusion, including entity disambiguation and reference resolution; Step 2: Integrate information from different sources, build a complete knowledge representation, and store it in the graph database.

5. According to claim 1, the industrial innovation question-answering method integrating knowledge graph and big model is characterized by: The specific process of step three is: Step 31: Manually convert part of the original data into question-answer format for sample learning; Step 32: Use the big model to convert the processed industrial innovation data into question-answer pair format, build a question-answer dataset, and divide it into training set, validation set, and test set; Step 33: Use the constructed knowledge graph to evaluate the rationality of the question and answer pairs.

6. According to claim 1, the industrial innovation question-answering method integrating knowledge graph and big model is characterized by: In step 4, the subgraph extraction model uses the retrieval enhancement method to extract the knowledge subgraph. The specific process is as follows: Step 41: For the entity in the question, extract the corresponding knowledge subgraph in the knowledge graph; Step 4: Extract the question entity and the answer entity, as well as all their two-hop neighbor entities, calculate the semantic similarity using the word vector algorithm, and retain the first 100 semantically related nodes to form the initial subgraph; Step 412: Encode the initial subgraph using a graph attention algorithm; Step 413: Use the pre-trained model Bert to obtain the embedding of the question entity and the answer entity; Step 414: Calculate the attention between the question entity and the answer entity through the bidirectional attention mechanism to obtain the weight of the question entity; Step 4-5: Using graph node encoding and question entity embedding, the attention coefficient of the question entity on all entities in the initial subgraph is calculated through the attention mechanism; Step 416: prune the knowledge subgraph, retain the entity with the highest attention coefficient each time, extract the relationship between the entities with the highest attention coefficient to obtain the knowledge subgraph, and represent the knowledge subgraph with triples; Step 42: Use the big model to convert the extracted triples into natural language; use the generated text as a prompt to enhance the output of the big model.

7. The industrial innovation question-answering method integrating knowledge graph and big model according to claim 6 is characterized by: In step 414, a bidirectional attention mechanism is designed to weight the answer entity using the representation of the question entity to obtain the importance of the question entity to the answer entity; let the question entity be the query Q and the answer entity be the key K, and measure the influence of the question entity on the answer entity through the attention score: ; In the formula, is the embedding dimension, is the question entity embedding, is the answer entity embedding, softmax is the normalization function, is the problem entity weight.

8. The industrial innovation question-answering method integrating knowledge graph and big model according to claim 6 is characterized by: The specific process of step 416 is as follows: The attention coefficient of the problem entity obtained in step 4-5 to all entities in the initial subgraph , using the question entity weight Weight it to obtain the weighted attention coefficient of the problem entity to the surrounding entities , each time you select or The three neighbor nodes with the highest value are expanded at most, the expansion depth is set to 4, and the attention coefficient is optimized by cross entropy loss; The attention coefficient , which can be expressed as: ; in, is the initial subgraph node embedding, expressed as follows: ; in, is the initial subgraph node embedding after graph neural network encoding, is the self-attention layer.

9. The industrial innovation question-answering method integrating knowledge graph and big model according to claim 1 is characterized by: In step five, the knowledge graph is used to enhance the big model. By extracting problem entities and knowledge subgraphs and using contextual learning, the big model can realize the functions of industry trend analysis and industry report generation.

10. An industrial innovation question-answering system integrating knowledge graph and big model, characterized by: The system is used to implement an industrial innovation question-answering method integrating knowledge graph and large model as described in any one of claims 1 to 9; the system includes a data integration and preprocessing module, a knowledge graph creation module, a question-answering data construction module, a knowledge graph and large model fusion module and a top-level application design module; The data integration and preprocessing module is used to collect industrial innovation data that combines patent information, academic literature, and policy systems, and then perform preprocessing; The knowledge graph creation module uses the preprocessed data to establish a knowledge graph in the field of industrial innovation; The question-answer data construction module converts the preprocessed industrial innovation data into a question-answer format and inputs the data into the knowledge graph and large model fusion module as a training set, a validation set, and a test set; The knowledge graph and large model fusion module is used to train the subgraph extraction model and extract knowledge subgraphs from the domain knowledge graph; Convert the extracted knowledge subgraph into text, and input the text into the big model as prompt words; The top-level application design module is used to implement industry trend analysis and industry report generation.

Images