Enterprise technology demand prediction method and system based on large model driving

Abstract

The invention belongs to the technical field of large model application, and relates to an enterprise technology demand prediction method and system based on large model driving, and the method comprises four parts: enterprise problem deep analysis, problem technology essence identification, technology knowledge base intelligent retrieval and multi-granularity technology demand prediction. In the deep analysis of enterprise problems, the problem description is subjected to structured processing, an entity relationship graph is constructed, and domain knowledge is aligned. The problem technology essence identification utilizes large model semantic understanding to analyze core challenges, and classifies technical problems through a multi-label classification model. And the technical knowledge base intelligently retrieves and calculates the similarity between the problem semantic vector and the knowledge graph, fuses multi-source data expansion, and screens candidate technologies according to the technology maturity. And performing multi-granularity technology demand prediction analysis on the core technology field and the subdivision direction to form a multi-level demand result. According to the invention, through systematic analysis and intelligent prediction, enterprise technology demands are accurately identified, and the method is suitable for scenes such as industry-university-research cooperation, technology transfer and innovative resource configuration.

Description

Technical Field

[0001] The present invention belongs to the technical field of large model applications, and particularly relates to a method and system for predicting enterprise technology requirements driven by a large model. Background Art

[0002] With the in-depth development of artificial intelligence and big data technologies, industry-university-research cooperation has become an important model for promoting technological innovation and achievement transformation. In the process of cooperation, enterprises often need to seek technical support from universities and research institutions to solve complex technical problems encountered in production and operation, such as process optimization, product upgrading, intelligent transformation, etc. However, when describing technical requirements, enterprises often start from the problem phenomenon and lack an accurate understanding of the required technology types, professional directions, and cooperation models, resulting in low cooperation docking efficiency and insufficient matching accuracy, which seriously restricts the effectiveness of industry-university-research cooperation. Therefore, developing an intelligent enterprise technology requirement prediction method to accurately identify the essence of technical requirements behind enterprise problems has become one of the key technologies to improve the efficiency of industry-university-research cooperation.

[0003] Currently, the methods for identifying enterprise technology requirements mainly include demand analysis based on keyword matching and demand judgment based on expert experience. The demand analysis based on keyword matching constructs a technical field dictionary, extracts key technical terms from the enterprise problem description, and then infers the required technology types. The demand judgment based on expert experience relies on the knowledge accumulation of domain experts to manually analyze the technical essence of enterprise problems and give cooperation suggestions. However, the inventors have found that the existing technology requirement identification methods have obvious deficiencies in the industry-university-research cooperation scenario. First, there is a lack of in-depth understanding of the complexity and technical span of enterprise problems, resulting in insufficient accuracy and comprehensiveness of demand identification. Second, the granularity division of technical requirements is too rough to accurately match specific disciplines and research directions. In addition, the semantic understanding and knowledge reasoning capabilities of large models are not fully utilized to挖掘隐含的技术需求. Finally, there is a lack of a multi-dimensional technical ability evaluation framework, making it difficult to provide targeted partner recommendations and cooperation model suggestions for enterprises. Summary of the Invention

[0004] To solve at least one of the technical problems in the above background art, the present invention provides an adaptive task decomposition method and system for a large model oriented to achievement transformation, which can significantly improve the decomposition quality of complex tasks, thereby improving the answer accuracy and integrity of large language models and providing better services for users.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The first aspect of the present invention provides a method for predicting enterprise technology requirements driven by a large model, which includes:

[0007] In-depth analysis of enterprise problems. The input enterprise problem description text is preprocessed, including text cleaning, standardization, and noise filtering. Then, multi-level entity recognition is performed, using the BERT model to extract technical entities, business entities, and constraint entities. At the same time, an entity relationship graph is constructed through dependency parsing and semantic role labeling, and the relationship strength score is calculated. Finally, domain knowledge alignment is performed to map the identified entities to a standardized ontology library, forming a structured problem knowledge representation.

[0008] Problem technical essence identification. Based on the semantic understanding capabilities of the large language model, the structured problem representation is input into the model for deep semantic analysis, followed by technical problem classification. A multi-label classifier is used to map the problem to a predefined technical category and calculate the confidence score. At the same time, the problem complexity is evaluated and quantitatively scored through three dimensions: technical complexity, implementation complexity, and integration complexity. Finally, a problem technical essence feature vector is generated, which includes category label, complexity level, and description of technical challenges.

[0009] Intelligent retrieval of technical knowledge base. The process involves semantic vectorization of questions, encoding question features into high-dimensional semantic representations using a pre-trained model and performing L2 normalization. Then, technical graph matching is performed, calculating semantic similarity in a large-scale technical knowledge graph and filtering candidate technical concepts. Simultaneously, multi-source data retrieval is integrated, querying scientific literature, patent databases, and industry-academia-research case libraries in parallel to obtain relevant technical information. Finally, candidate technologies are evaluated, and the candidate technology set is selected and ranked based on technical maturity, domain matching degree, and feasibility indicators.

[0010] Multi-granularity technology demand forecasting. This process identifies core technology areas of demand, uses a multi-label classifier to predict the primary technology areas involved in the company's problems and calculates demand intensity scores. Then, it subdivides professional technology directions, using a conditional random field model to predict secondary professional directions for activated technology areas and analyzes technological complementarity. Simultaneously, it quantifies technology capability requirements, analyzing the capability demand intensity across five dimensions—theoretical research, algorithm design, experimental verification, engineering implementation, and system integration—through a multi-dimensional evaluation framework. Finally, it generates a technology demand forecasting report, outputting complete forecast results including a list of technology areas, a spectrum of professional directions, and a capability demand matrix.

[0011] A second aspect of the present invention provides a large-model-driven enterprise technology demand forecasting system, comprising:

[0012] The deep analysis module for enterprise problems forms a structured understanding of enterprise problems through text preprocessing and multi-level entity recognition, including the identification and relationship extraction of technical entities, business entities and constraint entities, and constructs an entity relationship graph through dependency parsing, and finally generates a standardized problem knowledge representation, providing a structured semantic input foundation for subsequent identification of the technical essence of the problem;

[0013] The problem technical essence identification module combines the deep semantic understanding capabilities of a large language model to determine the core technical challenges and essential characteristics of a problem. This includes running a multi-label classifier to map the problem to a predefined technical category system, assessing the technical complexity and implementation difficulty level, and generating a technical essence feature vector containing category labels and complexity scores, providing accurate query features for technical knowledge base retrieval.

[0014] The intelligent retrieval module of the technology knowledge base performs semantic matching and association expansion on a large-scale technology knowledge graph according to the essential characteristics of the problem technology. This ensures that the retrieved candidate technologies not only meet the requirements of semantic similarity, but also discover potential technology associations through multi-source data fusion and graph structure analysis. Furthermore, it selects a high-quality set of candidate technologies based on technology maturity, domain matching degree, and feasibility indicators.

[0015] The multi-granularity technology demand forecasting module uses a hierarchical classifier and a conditional random field model to perform multi-level analysis on candidate technologies, identify the primary technology areas and secondary professional directions required by enterprises, including predicting the intensity of demand in the technology areas, analyzing the complementarity of professional directions, and quantifying the five dimensions of technology capability requirements. Through multi-granularity analysis, a complete technology demand forecasting report is generated to ensure that the forecasting results meet the actual technology cooperation needs of enterprises.

[0016] A third aspect of the present invention provides a computer-readable storage medium.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the adaptive task decomposition method for a large-scale achievement transformation model as described above.

[0018] A fourth aspect of the present invention provides a computer device.

[0019] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the adaptive task decomposition method for a large-scale achievement transformation model as described above.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] (1) The enterprise technology demand prediction method based on large model driven proposed in this invention can significantly improve the accuracy and comprehensiveness of technology demand identification, thereby improving the success rate and cooperation effect of industry-university-research cooperation, and providing more accurate decision support for enterprise technology innovation and university scientific research results transformation.

[0022] (2) This invention establishes a systematic multi-granularity technology requirements analysis system. This system deeply analyzes enterprise problems from three levels: technology field, professional direction, and capability requirements. At the technology field level, a multi-label classifier is used to identify the demand intensity of primary technology fields. At the professional direction level, a conditional random field model is used to predict the matching degree and complementarity of secondary professional technology directions. At the capability requirements level, a five-dimensional evaluation framework is used to quantify the specific capability requirements of theoretical research, algorithm design, experimental verification, engineering implementation, and system integration.

[0023] (3) This invention realizes an intelligent technology retrieval and matching mechanism based on semantic understanding. This method differs from traditional keyword matching methods, and instead designs a complete semantic vectorization and graph association analysis process. It uses a pre-trained language model to encode enterprise questions into high-dimensional semantic vectors, and then performs semantic similarity calculation and association expansion in a large-scale technology knowledge graph. At the same time, it integrates multi-source data such as scientific and technological literature, patent databases, and industry-university-research case libraries. Finally, it screens and ranks candidate technologies based on multi-dimensional indicators such as technology maturity, domain matching degree, and implementation feasibility.

[0024] (4) This invention innovatively combines the deep semantic understanding capabilities of large language models with structured knowledge representation. By converting natural language problem descriptions into structured knowledge graph representations, it achieves accurate mapping from unstructured text to standardized technical concepts. It utilizes pre-trained models such as BERT for entity recognition and relation extraction, combines dependency parsing to construct the semantic structure of the problem, and finally generates standardized problem representations through domain ontology alignment, providing high-quality input features for technology demand prediction.

[0025] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0026] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0027] Figure 1 This is a flowchart of an enterprise technology demand forecasting method based on a large model driven by an embodiment of the present invention;

[0028] Figure 2 This is a flowchart illustrating the in-depth analysis of enterprise problems according to an embodiment of the present invention;

[0029] Figure 3 This is a flowchart illustrating the identification of the technical essence of the problem in an embodiment of the present invention;

[0030] Figure 4This is a flowchart of the intelligent retrieval of the technical knowledge base according to an embodiment of the present invention;

[0031] Figure 5 This is a flowchart of multi-granularity technology demand prediction according to an embodiment of the present invention. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0034] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0035] Example 1

[0036] Reference Figure 1 This embodiment provides a method for predicting enterprise technology demand based on a large model, which includes:

[0037] S101: Deep Analysis of Enterprise Problems. It receives enterprise problem description text, performs text preprocessing and standardization cleaning, runs a BERT-based named entity recognition model to extract multi-level entities (including technical entities, business entities, and constraint entities), uses dependency parsing and semantic role labeling to extract semantic relationships between entities, performs knowledge alignment and standardization mapping through a domain ontology library, and finally constructs a structured problem knowledge graph containing entities, relationships, and attributes, providing standardized semantic representation input for subsequent identification of the technical essence of the problem.

[0038] S102: Problem Technical Essence Identification. Based on the structured problem representation output by S101, it utilizes the deep semantic understanding capabilities of a large language model to analyze the core challenges and abstract the technical essence of the problem. It runs a multi-label classification model based on the Transformer architecture to map the problem to a predefined system of 30 technical problem categories. It quantifies the difficulty level of the problem through a multi-dimensional complexity evaluation algorithm (technical complexity, implementation complexity, integration complexity) and outputs a problem technical essence feature vector containing problem category labels, complexity scores, and technical challenge descriptions, which serves as the query basis for S103 technical retrieval.

[0039] S103: Intelligent retrieval of technical knowledge base. It receives the essential technical features of the problem in S102, encodes the problem features into high-dimensional semantic vectors using a pre-trained language model, performs semantic similarity matching and graph structure association analysis in a pre-constructed large-scale technical knowledge graph, integrates multi-source technical data such as scientific and technological literature, patent databases, and industry-academia-research case libraries to expand knowledge and discover associations, and filters and sorts candidate technology sets based on technology maturity assessment models, application domain matching algorithms, and implementation feasibility analysis indicators to provide a high-quality technical solution candidate pool for S104.

[0040] S104: Multi-granularity technology demand forecasting. Based on the candidate technology set provided in S103, it runs a hierarchical classifier based on large model fine-tuning to predict the primary technology fields required by enterprises (such as artificial intelligence, new materials, biotechnology, and 15 other major fields). For each activated technology field, a conditional random field model is used to map the corresponding secondary professional technology direction. The intensity of capability demand in five dimensions, including theoretical research, algorithm design, experimental verification, engineering implementation, and system integration, is evaluated through a technology capability demand quantitative analysis framework. Finally, a complete multi-granularity technology demand forecasting report is output, which includes a list of technology fields, a spectrum of professional directions, and a capability demand matrix.

[0041] Reference Figure 2 This embodiment provides a flowchart of the steps for in-depth analysis of enterprise problems, and provides a detailed explanation of the process.

[0042] Step 200: Enterprise Problem Text Preprocessing. The input enterprise problem description is standardized to form a standardized text suitable for subsequent analysis. This provides standardized text input for Step 201. Step 200 can be further refined as follows:

[0043] Step 200-1: Receive the enterprise problem description text input, obtain the original text string, and record the text length and encoding information;

[0044] Step 200-2: Text cleaning and word segmentation. Regular expressions are used to clean HTML tags, special characters, and redundant whitespace, and the text is uniformly converted to UTF-8 encoding. The cleaned text is then input into the pre-trained BERT-Chinese model for word segmentation, generating a list of token sequences [token1, token2, ..., tokenN]. Each token contains the original text, its position index, and pre-labeled part-of-speech tags.

[0045] Step 200-3: Syntactic Structure Analysis and Normalization Output. The lexical sequence from Step 200-2 is input into the part-of-speech tagging model and dependency parser to identify key nouns, verbs, adjectives, and other grammatical components, constructing a dependency syntax tree structure. Subject-verb-object structures and modification relations are extracted to generate a grammatical structure graph, forming a structured data package containing {normalized text, lexical sequence, part-of-speech tagging, and syntax tree}. This data package will serve as the direct input to the BERT-based entity recognition model in Step 201.

[0046] Step 201: Multi-layer entity recognition and classification. Receive the standardized text and grammatical structure from step 200, and identify and classify different types of key entity information from the enterprise question text. This provides entity information support for step 202. Step 201 can be further refined as follows:

[0047] Step 201-1: Business Entity Identification. Input the normalized text and part-of-speech tagging information into the BERT-BiLSTM-CRF sequence labeling model to extract business-related entities such as product names, business processes, and performance metrics, generating a business entity list [{entity text, start position, end position, confidence level}].

[0048] Step 201-2: Technical entity recognition. Use the pre-trained RoBERTa model to process the same normalized text, and identify technical terms such as algorithm name, technical standard, and equipment model. Output a list of technical entities in the same format as business entities.

[0049] Step 201-3: Constraint entity recognition. Combining the syntactic structure information from step 200, a hybrid model of regular expressions and deep learning is used to identify constraints such as time limits, cost budgets, and accuracy requirements, and the constraints are converted into key-value pairs in the format {constraint type: constraint value}.

[0050] Step 201-4: Entity quality control and classification. Confidence assessment is performed on all identified entities. Results below the 0.85 threshold are filtered out. Valid entities are classified and labeled as business entities, technical entities, and constraint entities. Overlapping entities across categories are handled, and the classification result with the highest confidence is retained. The confidence filtering formula is:

[0051] E valid ={e i / Conf(e i )≥θ,e i ∈E all}

[0052] Among them, E valid Denotes the set of valid entities, e i Represents the i-th entity, Conf(e) i ) represents entity ei The confidence score, θ = 0.85 is the preset confidence threshold, E all This represents the set of all identified entities;

[0053] Step 201-5: Entity list construction and output preparation. Record the {type, text content, document location, confidence, context window} of each entity and generate a structured entity list in JSON format. This entity list will be used as the node input of the relation extraction model in Step 202, and each entity will become a candidate node in the relation graph.

[0054] Step 202: Entity Relationship Extraction and Graph Construction. Receive the entity list from Step 201, establish semantic relationships between entities, construct a directed relation graph and perform graph structure analysis, providing the entity relation graph for Step 203. Step 202 can be further refined as follows:

[0055] Step 202-1: Entity Relationship Identification. Read the entity list from Step 201, construct a candidate set of entity pairs [(entity1, entity2),...]. Input each entity pair and its context text into the BERT-RE attention mechanism model. The model outputs the relationship type (dependent on, cause, contain, restrict) and confidence score. Relationship triples (entity1, relationship type, entity2, confidence) with a confidence score greater than a preset threshold are retained.

[0056] Step 202-2: Construct a directed graph. Based on the entities in Step 201 as nodes and the relationships in Step 202-1 as edges, construct a NetworkX directed graph object, use the relationship confidence as the weight attribute of the edges, and add attribute information such as entity type and text content to each node.

[0057] Step 202-3: Calculate graph structure complexity. Use the NetworkX algorithm to calculate the edge connectivity and vertex connectivity of the graph, calculate the global and local clustering coefficient distributions, and calculate the average shortest path length and graph diameter. The formula for calculating graph structure complexity is:

[0058]

[0059] Where |E| represents the number of edges in the graph, and |V| represents the number of nodes in the graph. For normalized edge density, CC global represents the global clustering coefficient, APL represents the average shortest path length, D represents the graph diameter, α, β, and γ are weighting coefficients, and the score range is [0, 10].

[0060] Step 202-4: Graph vector embedding generation, using the Node2Vec algorithm, setting the embedding dimension to 128, the walk length to 80, and the window size to 10, performing random walk sampling based on the constructed relation graph. The objective function of Node2Vec is:

[0061]

[0062] Where, f:V→R d Let N be the mapping function, mapping node u to a d-dimensional feature space (d = 128). S (u) represents the set of neighboring nodes generated by node u through a random walk strategy S, and V represents the set of all nodes in the graph. Node embedding vectors are trained to generate a 128-dimensional vector representation for each entity node, which is used for subsequent semantic similarity calculations.

[0063] Step 202-5: Integrate the graph data package. Collect the adjacency matrix, node degree distribution statistics, and graph embedding vector dictionary. Serialize the graph structure data into JSON format, save the vector data as a NumPy array, and output the complete entity relationship graph data package {graph object, adjacency matrix, embedding vector, complexity score}. This data package will be passed to step 203 for knowledge alignment processing along with the entity list from step 201.

[0064] Step 203: Domain Knowledge Alignment and Semantic Disambiguation. Receive the entity relationship graph from Step 202 and the entity list from Step 201. Align the identified entities with the technical domain knowledge base to determine the precise technical meaning of the entities. This step outputs a structured question representation, which serves as input for the question intent analysis in Step 300. Step 203 can be further refined as follows:

[0065] Step 203-1: Knowledge base query and preliminary matching. Read the technical entities and business entities in the entity list in Step 201, and perform text matching in the pre-built technical technology ontology library (CAS Technology Ontology, IEEE Computer Society Dictionary). Generate multiple candidate concepts for each entity, including concept ID, standard name, and definition description.

[0066] Step 203-2: Semantic similarity calculation and ambiguity resolution. Combine the 128-dimensional graph embedding vector from step 202 with the entity context text, use Sentence-BERT to generate semantic vectors, calculate the cosine similarity between the entity semantic vector and the candidate concept vectors in the knowledge base, and select the concept with the highest similarity and exceeding the 0.75 threshold as the standardized representation of the entity.

[0067] Step 203-3: Entity standardization and unknown entity processing. Replace successfully matched entities with standard concepts from the knowledge base, retain the original text as aliases, mark unmatched entities as "unknown entities", retain the original recognition results, verify the consistency of entity references in the relationship graph, and update node identifiers.

[0068] Step 203-4: Domain Coverage Assessment. Based on the relationship graph structure in Step 202, calculate the proportion of entities that successfully match the knowledge base. Domain coverage is equal to the number of matched entities divided by the total number of entities, reflecting the breadth of the technical field of the problem. According to the distribution of matched knowledge base concepts, automatically generate the main technical field tags of the problem.

[0069] Step 203-5: Final Question Representation Generation and Output. Summarize the disambiguated entity list, updated relationship graph, domain labels, and coverage scores. Based on the standardized entity and relationship structure, generate an overall question semantic vector representation. Construct a complete data package containing {standardized entity list, relationship graph, domain labels, coverage score, question vector}. This standardized data package will serve as the direct input for the question intent analysis stage in Step 300, where the question vector is used for intent classification, the relationship graph is used for complexity assessment, and the entity list is used for technical essence identification.

[0070] Reference Figure 3 This embodiment provides a flowchart of the steps for identifying the technical essence of a problem, and provides a detailed explanation of the process for identifying the technical essence of a problem.

[0071] Step 300: Deep Intent Analysis of the Problem. Receive the standardized data packet {standardized entity list, relationship graph, domain labels, coverage score, question vector} output from Step 203. Utilize the semantic understanding capabilities of the large model to analyze the core challenges and implicit technical needs of the enterprise's problem, outputting an intent analysis result packet to provide intent feature input for the technology classification model in Step 301. Step 300 can be further refined as follows:

[0072] Step 300-1: Data reception and preprocessing. Receive the structured question representation data packet output from step 203, extract the question semantic vector and entity relationship graph, verify data integrity and standardize the format, and use the question vector as the main input feature of the intent analysis model.

[0073] Step 300-2: Question Intent Classification and Recognition. Input the question semantic vector into the pre-trained RoBERTa-Large intent classifier (which has been fine-tuned on the industry-academia-research collaborative corpus). The model outputs the probability distribution of core intent types, such as performance improvement, function expansion, cost reduction, and security reinforcement. Select the intent with the highest probability that exceeds the threshold as the main intent type.

[0074] Step 300-3: Core challenge description generation. Input the problem semantic vector and the identified intent type into the sequence-to-sequence model based on the BART architecture (pre-trained on the technical problem summary dataset). The model automatically generates core challenge description text, which summarizes the technical difficulties and key bottlenecks of the problem.

[0075] Step 300-4: Keyword weight analysis. Use multi-head self-attention mechanism to analyze the keywords and phrases in the original question description, calculate the contribution weight of each word to the identification of question intent, and generate a word-weight mapping table {word: weight score}, with the weight score ranging from [0,1], for subsequent extraction of technical points;

[0076] Step 300-5: Integrate the intent analysis results, record and structure the output {main intent type, intent confidence, core challenge description, keyword weight distribution}, and generate a JSON format intent analysis result package. This result package will serve as the auxiliary feature input for the multi-label classification model in Step 301, where intent type is used to guide technology classification and keyword weight is used to enhance classification accuracy.

[0077] Step 301: Multi-label classification of technical issues. Receive the issue vector from step 203 and the intent analysis result package from step 300. Map the enterprise issues to a predefined technical issue classification system and output the multi-label classification results, providing technical dimension input for complexity calculation in step 302. Step 301 can be further refined as follows:

[0078] Step 301-1: Construction of the technology classification system. Establish a hierarchical classification system containing 5 primary categories: performance optimization (including 8 secondary subcategories such as algorithm optimization, hardware upgrade, and architecture improvement), function implementation (including 6 secondary subcategories such as new function development, interface integration, and module expansion), quality assurance (including 7 secondary subcategories such as test optimization, defect repair, and stability improvement), cost control (including 5 secondary subcategories such as resource optimization and process improvement), and intelligent upgrade (including 4 secondary subcategories such as AI integration and automation transformation), totaling 30 secondary technology categories.

[0079] Step 301-2: Multi-label classification model construction. BERT is used as the base encoder, and 30 binary classification heads are added at the top layer. The model is trained using the binary cross-entropy loss function. The model supports a single question belonging to multiple technology categories at the same time, and each classification head outputs the probability score of that category.

[0080] Step 301-3: Classification prediction execution. The question vector from step 203 is concatenated with the intent feature from step 300 to form an extended feature vector. This vector is then input into the multi-label classification model. The sigmoid activation function is used to output the probability distribution of 30 technical categories. A classification threshold is set, and categories with probabilities greater than the threshold are retained as prediction results.

[0081] Step 301-4: Post-processing optimization of classification results. Make consistency adjustments based on predefined category hierarchy relationships and mutual exclusion rules. For example, if a second-level category is activated, the corresponding first-level category must be activated. Handle logically conflicting classification combinations, such as conflict detection and adjustment between "cost reduction" and "significant functional expansion".

[0082] Step 301-5: Integrate and output the classification results. Record the final list of technical category labels, the confidence score dictionary for each category, and the 30×30 category association strength matrix. Generate a multi-label classification result data package {list of activated categories, category confidence, association matrix}. This data package will be input into step 302, where the number of activated categories is used to calculate the technical cross-domain degree, and the category confidence is used for complexity weight adjustment.

[0083] Step 302: Quantitative Assessment of Problem Complexity. Input the entity list from Step 201, the classification results from Step 301, and the relationship graph from Step 202. Quantitatively calculate the comprehensive technical complexity of the enterprise problem from three dimensions: entity complexity, technical cross-domain complexity, and constraint complexity. This step outputs a problem complexity assessment report, including scores for each dimension and the comprehensive complexity level. Step 302 can be further refined as follows:

[0084] Step 302-1: Entity complexity calculation. Read the entity list identified in step 201 and count the total number of entities E. total Number of entity types T types (Business, Technology, and Constraint categories), the number of relationships between entities in step 202, R. count The entity complexity is calculated using the weighted summation formula: Entity Complexity =a×E total +b×T types +c×R count Then, normalize the values ​​in the [0,1] interval. Where a, b, and c are weighting coefficients;

[0085] Step 302-2: Calculate the cross-domain technology degree. Based on the multi-label classification results of step 301, count the number N of activated first-level technology fields. domains and the number of secondary technology categories N subcats Calculate the probability distribution P = [p1, p2, ..., p5] for each domain, and use the Shannon entropy formula to calculate the cross-domain degree: Cross Domain= -∑(pi×log(pi)), and adjusted by the number of activated categories: Final CrossDomain =Cross Domain ×(N subcats / 30);

[0086] Step 302-3: Constraint complexity calculation. Analyze the constraint entities identified in step 201 and classify them by constraint type: number of time constraints T. time Cost constraint quantity T cost Number of performance constraints T perf Number of safety constraints T secu The strength weights of various constraints are calculated based on the constraint strength (derived from entity confidence). The constraint complexity formula is: Constraint Complexity = Σ(Number of constraints × Constraint strength weight) / Maximum possible number of constraints;

[0087] Step 302-4: Determine the weights of complexity dimensions. Use the Analytic Hierarchy Process (AHP) to determine the weight coefficients of each complexity dimension. Calculate the following using the expert evaluation matrix: Entity complexity weight W1 = 0.4 (reflecting problem size), Technical cross-domain weight W2 = 0.35 (reflecting breadth of scope), Constraint complexity weight W3 = 0.25 (reflecting implementation difficulty). Weight sum verification: W1 + W2 + W3 = 1.0;

[0088] Step 302-5: Comprehensive Complexity Calculation and Ranking Mapping. Calculate the comprehensive complexity score of the problem using a weighted average: Final Complexity =W1×Entity Complexity +W2×Cross Domain +W3×Constraint Complexity Level mapping is performed based on a preset threshold: low complexity (Final) Complexity <0.3), medium complexity (0.3≤Final) Complexity <0.7), high complexity (Final) Complexity (≥0.7), output a complete complexity assessment report {scores for each dimension, overall score, complexity level, main complexity factors}, which will serve as an important basis for intelligent retrieval of the technical knowledge base.

[0089] Reference Figure 4 This embodiment provides a flowchart of the steps for intelligent retrieval of technical knowledge base, and provides a detailed explanation of the intelligent retrieval process of technical knowledge base.

[0090] Step 400: Problem semantic vectorization processing. Inputting the complexity assessment report from step 302, the classification results from step 301, the graph embedding vector from step 202, and the original problem text, this step fuses the multi-dimensional problem representation information into a unified high-dimensional semantic vector, constructing a standardized retrieval feature representation. This step outputs a normalized problem semantic vector, which is directly used for the technical concept similarity calculation in step 401. Step 400 can be further refined as follows:

[0091] Step 400-1: Multi-source data reception and preprocessing. Receive the complexity assessment report {scores of each dimension, comprehensive score, complexity level, main complexity factors} output from step 302, and the classification results {activation category list, category confidence, association matrix} from step 301. Extract key features: complexity score, activation technique category, and category confidence distribution.

[0092] Step 400-2: Semantic encoding of the question text. The original question description text is input into the pre-trained Sentence-BERT model for encoding processing to generate a question semantic embedding vector. This vector captures the deep semantic information and contextual association of the question, serving as the core feature basis for semantic retrieval.

[0093] Step 400-3: 3D feature fusion and splicing. Extract the entity relationship graph embedding vector from step 202, convert the 30 technical categories from step 301 into 30-dimensional one-hot encoded vectors (activation category is 1, the rest are 0), and splice them with the complexity feature vector to form a multimodal feature vector.

[0094] Step 400-4: Unified semantic space mapping. The fused feature vector is input into a three-layer fully connected neural network for dimensionality reduction mapping. The ReLU activation function and dropout (0.3) are used to prevent overfitting, and the heterogeneous features are mapped to a unified semantic space to ensure the effective fusion of information from different modalities.

[0095] Step 400-5: Vector normalization and standardized output. The generated question semantic vector is subjected to L2 normalization to normalize the vector magnitude to 1, generating the final normalized question semantic vector. This vector will be used as the query vector for technical concept matching in step 401 to ensure the accuracy and consistency of subsequent cosine similarity calculation.

[0096] Step 401: Semantic matching of the technical knowledge graph. Input the normalized question semantic vector from Step 400, and perform efficient semantic retrieval in the pre-constructed technical knowledge graph to find the candidate set of technical concepts most relevant to the question. This step outputs a sorted list of candidate technical concepts, providing a keyword basis for multi-source retrieval in Step 402. Step 401 can be further refined as follows:

[0097] Step 401-1: Loading and initializing the technology knowledge graph. Load the pre-built large-scale technology knowledge graph, which covers major technology fields such as artificial intelligence, biotechnology, new materials, new energy, and advanced manufacturing. Each technology node contains attribute information such as {concept name, technology description, domain tag, maturity level}.

[0098] Step 401-2: Construction of the technical concept vector library. Using the same Sentence-BERT model as in Step 400-2, the standardized descriptive text of technical concepts is batch encoded to generate a semantic embedding matrix of technical concepts. Principal component analysis (PCA) is then used to reduce the dimensionality to be consistent with the question vector, thus constructing an efficient vector retrieval library.

[0099] Step 401-3: Fast similarity calculation and retrieval. Embed the semantic vector of the problem and the technical concept from Step 400-5 into the vector library for batch cosine similarity calculation. Use the Faiss approximate nearest neighbor search algorithm to accelerate the calculation process.

[0100] Step 401-4: Filter technical concepts with similarity higher than a preset threshold (e.g., 0.65), sort them in descending order of similarity, and select the Top-K (e.g., K=100) candidate technical concepts;

[0101] Steps 401-5: Structured output of candidate results, recording the complete information of each candidate technical concept {concept name, similarity score, technical field tag, knowledge graph node ID, technical description summary}, generating a structured list of candidate technical concepts. This list will be passed to step 402 as the source of query keywords for multi-source retrieval, where the concept name is used to construct the search query, and the similarity score is used for subsequent result weighting.

[0102] Step 402: Multi-source technical data fusion retrieval. Input the candidate technical concept list from step 401, and retrieve relevant technical documents in parallel from multiple authoritative data sources based on the candidate technical concepts to obtain comprehensive technical information support. This step outputs a deduplicated set of multi-source technical documents, providing a document foundation for the association and expansion in step 403. Step 402 can be further refined as follows:

[0103] Step 402-1: Construct a multi-source technology data retrieval interface, including data sources such as CNKI academic literature database, State Intellectual Property Office patent database, and technology transfer case database;

[0104] Step 402-2: Based on the candidate technology concepts in step 401, generate query keywords for each data source, and enhance query coverage using synonym expansion and word form changes;

[0105] Step 402-3: Parallel retrieval of each data source, using TF-IDF and BM25 algorithms to calculate the relevance score between the document and the query, and returning the Top-K relevant documents for each data source;

[0106] Step 402-4: Deduplicate the search results and use document fingerprinting technology (MinHash) to identify duplicate or highly similar documents, retaining the version with the highest relevance score;

[0107] Step 402-5: Document Metadata Extraction and Structuring. Extract standardized metadata from the selected documents: {Document ID, Title, Abstract, Keyword List, Author Information, Publication Time, Journal / Conference Name, Citation Count, Technical Field Tags}. Use the BERT model to encode the abstract to generate document semantic vectors, and construct a structured multi-source technical information set. This set will serve as the document base library for the technical association analysis in Step 403.

[0108] Step 403: Technology Association Network Expansion. Input the candidate technology concepts from Step 401, the technology document set from Step 402, and the technology knowledge graph. Based on the graph structure relationships, discover more related technologies, including precursor technologies, derivative technologies, and cross-application technologies. This step outputs an expanded set of technology candidates, providing comprehensive technology options for the technology evaluation in Step 404. Step 403 can be further refined as follows:

[0109] Step 403-1: Using the candidate technology concepts retrieved in step 401 as the starting node, perform multi-hop neighbor expansion in the technology knowledge graph, with the expansion depth set to 2 hops;

[0110] Step 403-2: Perform differentiated expansion based on the relationship type. For "dependency relationship", expand the prerequisite technology; for "development relationship", expand the derivative technology; for "application relationship", expand the relevant application field.

[0111] Step 403-3: Use the Random Walk with Restart algorithm to discover related technologies on the technology knowledge graph. The number of walk steps is set to 10, and the restart probability is set to 0.2.

[0112] Step 403-4: Calculate the semantic relevance between the extended technology nodes and the original problem, use a graph neural network (GraphSAGE) to learn the representation of the nodes in the graph structure, and combine semantic similarity to make a comprehensive score;

[0113] Step 403-5: Association strength assessment and candidate set screening. Taking into account three dimensions, namely graph structure association strength, random walk access frequency and semantic similarity, the final association score is calculated using a weighted fusion formula. An association strength threshold is set to screen highly relevant technologies, and finally an extended candidate set of technologies {technology ID, technology name, association score, extension path} is formed. This set will be input to step 404 for comprehensive evaluation and final screening.

[0114] Step 404: Candidate Technology Filtering and Evaluation. Input the expanded candidate technology set from Step 403 and the technical document set from Step 402. Perform a comprehensive evaluation and screening of the candidate technologies based on multi-dimensional evaluation indicators. This step outputs the final ranked candidate technology set, provided for enterprise decision-making. Step 404 can be further refined as follows:

[0115] Step 404-1: Construct a technology evaluation index system, including five dimensions: technology maturity (TRL level 1-9), application domain matching degree, implementation complexity, cost-effectiveness ratio, and degree of innovation;

[0116] Step 404-2: Use a machine learning-based technology maturity assessment model to predict the TRL level based on features such as the number of published documents, patent applications, and industrialization cases of the technology;

[0117] Step 404-3: Calculate the domain matching degree between the technology and the enterprise problem, and use the hierarchical semantic matching algorithm to compare the similarity between the application domain of the technology and the industry to which the problem belongs;

[0118] Step 404-4: Assess the feasibility of the technology implementation, taking into account factors such as technical complexity, required resources, and implementation cycle, and use the multi-objective decision analysis method (TOPSIS) to conduct a comprehensive score;

[0119] Step 404-5: Apply threshold filtering and diversity constraints to retain technologies with a comprehensive score higher than the set threshold, while ensuring that candidate technologies cover different technical paths, and finally output the sorted candidate technology set.

[0120] Reference Figure 5 This embodiment provides a flowchart of the multi-granularity technology demand forecasting process, and provides a detailed explanation of the multi-granularity technology demand forecasting process.

[0121] Step 500: Core Technology Area Needs Identification. Inputting the final candidate technology set from Step 404, the problem complexity assessment from Step 302, and the original problem description, this step intelligently identifies the main technology areas required to solve the problem from the enterprise's problem characteristics, constructing a top-level architecture for the technology requirements. This step outputs a list of technology area needs, providing domain scope constraints for the professional direction prediction in Step 501. Step 500 can be further refined as follows:

[0122] Step 500-1: Multi-source input data integration. Receive the candidate technology set {technology name, comprehensive score, scores of each dimension, reasons for recommendation, implementation suggestions} output from step 404. Extract the problem complexity features {scores of each dimension, comprehensive score, complexity level, main complexity factors} from step 302. Combine the original problem description text to construct a comprehensive feature input for technical field identification.

[0123] Step 500-2: Construct a classification system for the technology fields of industry-academia-research cooperation, including primary technology fields such as artificial intelligence, big data and cloud computing;

[0124] Step 500-3: Use a multi-label technology classifier (such as BERT). The classifier is pre-trained on data from industry-academia-research collaborative projects to learn the mapping relationship between enterprise problems and technology fields.

[0125] Step 500-4: Calculate the matching probability between the problem description and each technical field. The sigmoid activation function can be used to output the multi-label classification results. Set the domain relevance threshold (e.g., 0.4).

[0126] Step 500-5: Structured output of technical field requirements results, generating a list of technical fields involved in the enterprise's problems {field code, field name, requirement intensity score, confidence level}. The requirement intensity score is calculated using the following formula:

[0127]

[0128] Among them, Intensity i P(y) represents the demand intensity of the i-th technical field. i =1 / x) is the predicted probability, N represents the number of candidate technologies in this field. total This represents the total number of candidate technologies. The average score for candidate technologies in this field is given, with α, β, and γ as weighting coefficients. A cross-domain dependency matrix is ​​constructed (through cross-domain distribution statistics of candidate technologies), and a structured technical field demand analysis report is output. This report will serve as the domain constraint for the professional direction prediction in step 501.

[0129] Step 501: Detailed Prediction of Professional Technical Directions. Input the technical field demand list from Step 500 and the candidate technology set from Step 404. For the identified technical fields, deeply predict the specific professional technical directions the company needs to seek cooperation in, constructing a refined technical demand spectrum. This step outputs a multi-level technical demand spectrum diagram, providing a detailed technical path for the capability analysis in Step 502. Step 501 can be further refined as follows:

[0130] Step 501-1: Construct a secondary professional direction classification system for each primary technical field, such as machine learning theory, deep neural networks, natural language processing, computer vision, knowledge graphs, reinforcement learning, etc. in the field of artificial intelligence;

[0131] Step 501-2: Use a hierarchical demand forecasting model and employ Conditional Random Fields (CRF) to capture dependencies and co-occurrence patterns among professional directions;

[0132] Step 501-3: Combining the characteristics of the enterprise's industry and the assessment of problem complexity, use an attention mechanism to weightedly integrate industry knowledge and problem characteristics to improve the accuracy of professional direction prediction;

[0133] Step 501-4: For each technology field, predict the Top-K most relevant professional technology directions and calculate the direction matching degree and technology complementarity index;

[0134] Step 501-5: Generation and evaluation of the technology requirement spectrum. Construct a multi-level enterprise technology requirement spectrum and record the key attributes of each professional direction: urgency of need (0-10 points, based on problem urgency and technology maturity), technology threshold level (basic / intermediate / advanced / expert level), and cooperation closeness score (0-1, based on cooperation depth and cycle). Generate a three-layer structure of the technology requirement spectrum in JSON format: {domain → professional direction → capability requirements}. This spectrum will be passed to step 502 for quantitative analysis of technical capabilities.

[0135] Step 502: Quantitative Analysis of Technical Capability Requirements. Inputting the technical requirement spectrum diagram from Step 501 and the complexity analysis from Step 302, this step quantitatively analyzes the specific types of technical capabilities and depth of professional knowledge required for the enterprise to solve its problems, forming a precise profile of cooperation needs. This step outputs a list of technical capability requirements, directly guiding the selection and matching of industry-academia-research partners. Step 502 can be further refined into the following steps:

[0136] Step 502-1: Construct a technical capability requirements assessment framework, which includes five dimensions: theoretical research capability, algorithm design capability, experimental verification capability, engineering implementation capability, and system integration capability.

[0137] Step 502-2: Transformer technology capability prediction model training. Construct a technology capability requirement prediction model based on the Transformer architecture. The model is pre-trained on the database of university research projects to learn the deep correlation pattern between problem characteristics and technology capability requirements.

[0138] Step 502-3: Intelligent identification of technical challenge types. Deeply analyze the technical challenge nature of the enterprise's problems, and use text classification and keyword extraction technology to identify challenge types: basic theoretical breakthrough (requiring original theoretical innovation), algorithm innovation (requiring core algorithm breakthrough), technology verification (requiring experimental and testing verification), engineering optimization (requiring system performance optimization), and integrated innovation (requiring the integration of multiple technologies). Match corresponding capability requirement weights to each challenge type.

[0139] Step 502-4: Quantitative assessment of capability requirements intensity. A 0-10 scale is used to quantitatively assess each technical capability dimension. The assessment algorithm combines the problem complexity from Step 302 (higher complexity indicates stronger capability requirements), the technical threshold level from Step 501 (higher threshold indicates higher professional depth requirements), and the technology maturity level (lower maturity indicates stronger research capability requirements). The calculation formula is as follows:

[0140] Capability Score =α×Complexity+β×Threshold+γ×(1-Maturity)

[0141] Where α, β and γ are weighting coefficients;

[0142] Step 502-5: Comprehensive Output of Technical Capability Requirements List. A detailed technical capability requirements list will be generated, containing complete information for each capability dimension: {Capability Type, Requirement Intensity Score, Professional Depth Level, Urgency Level, and Suggested Cooperation Forms}. The suggested cooperation forms include four types: technical consultation (short-term guidance), joint R&D (medium- to long-term cooperation), talent training (long-term capability building), and platform co-construction (strategic cooperation). Finally, a structured technical capability requirements analysis report will be output, which will serve as the core basis for matching industry-academia-research partners and project planning.

[0143] Example 2

[0144] This embodiment provides an enterprise technology demand forecasting system based on a large model, which includes:

[0145] The Enterprise Problem Deep Analysis Module extracts business entities, technical entities, and constraint entities from enterprise problem descriptions through text standardization preprocessing and multi-layer entity recognition technology. It constructs an entity relationship graph and performs domain knowledge alignment to form a structured problem representation that includes a disambiguated entity list, relationship graph, domain labels, and problem semantic vectors, providing a standardized input basis for subsequent problem essence identification.

[0146] The Problem Technology Essence Identification Module is based on structured problem representation. It utilizes the semantic understanding capabilities of large models to perform in-depth analysis of problem intent, runs a multi-label technology problem classification model to map enterprise problems to a predefined technology category system, and quantitatively evaluates problem complexity from dimensions such as entity complexity, technology cross-domain degree, and constraint complexity. Finally, it forms a technology essence feature that includes problem intent, technology classification labels, and complexity level, providing a semantic foundation and strategy guidance for technology knowledge base retrieval.

[0147] The intelligent retrieval module of the technology knowledge base receives the essential characteristics of the problem technology, performs semantic vectorization of the problem and semantic matching of the technology knowledge graph, integrates multi-source technology data such as scientific and technological literature, patent databases, and industry-university-research case databases for correlation and expansion, and filters and evaluates candidate technologies based on indicators such as technology maturity, application field matching degree and feasibility of implementation, forming a high-quality candidate technology set after ranking, providing a rich technology selection basis for multi-granularity technology demand prediction.

[0148] The multi-granularity technology demand prediction module, based on a candidate technology set, runs a large model-driven multi-label classifier to identify the demand in core technology areas, uses a hierarchical technology demand prediction model to subdivide professional technology directions, and constructs a technology capability demand assessment framework to conduct quantitative analysis of technology capability requirements. Finally, it generates a complete multi-granularity technology demand analysis result that includes a list of technology areas, a professional direction genealogy diagram, and a list of technology capability requirements, providing accurate basis for partner matching and solution formulation.

[0149] Example 3

[0150] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the adaptive task decomposition method for a large-scale achievement transformation model as described above.

[0151] Example 4

[0152] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the above-described method for predicting enterprise technology demand based on a large model.

[0153] Those skilled in the art will recognize that embodiments of the present invention can take the form of a method, system, or computer program product. Therefore, the invention can be implemented in hardware, software, or a combination of both. Furthermore, the invention can take the form of a computer program product on a computer-usable storage medium (including but not limited to disk storage and optical storage), containing one or more computer-usable program codes.

[0154] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting enterprise technology demand based on a large model, characterized in that, Breaking through the limitations of traditional keyword matching and expert experience judgment, this approach adopts a four-stage progressive technology demand mining concept: "unstructured problem → structured representation → technology essence identification → intelligent demand prediction." It includes four core modules: in-depth analysis of enterprise problems, identification of the technical essence of problems, intelligent retrieval of technical knowledge base, and multi-granularity technology demand prediction. This enables the intelligent transformation from enterprise problem phenomena to precise technical demands, solving the technical problems of inaccurate demand identification and coarse matching granularity in industry-academia-research cooperation.

2. The enterprise problem in-depth analysis process according to claim 1, characterized in that, This paper proposes a four-step structured parsing concept: "text preprocessing → multi-layer entity recognition → relation graph construction → domain knowledge alignment". Unlike traditional single entity extraction methods, this paper constructs a structured problem representation of semantic vectors by using a three-dimensional entity recognition system of business entities, technical entities, and constraint entities, combined with relation extraction based on attention mechanism and graph embedding technology. This lays a precise semantic foundation for subsequent technical requirement identification.

3. The problem-solving technical essence identification process according to claim 1, characterized in that, Breaking through the bottleneck of the lack of deep semantic understanding in existing technologies, this paper innovatively integrates the semantic understanding capabilities of large models with multi-dimensional complex quantification and evaluation technology. Through a three-layer progressive recognition concept of "deep intent analysis → multi-label technology classification → complex quantification and evaluation", it uses a fine-tuned large model for intent recognition and adopts the hierarchical analysis method to quantify the three dimensions of entity complexity, technology cross-domain degree, and constraint complexity, forming deep technical essential characteristics that are different from traditional surface feature analysis.

4. The intelligent retrieval process of the technical knowledge base according to claim 1, characterized in that, Breaking through the limitations of single-source data retrieval and static matching algorithms, this paper innovatively proposes a four-stage retrieval concept: "semantic vector fusion → knowledge graph matching → multi-source data expansion → intelligent evaluation and screening." By using a fully connected neural network, the question semantic vector, graph embedding vector, and technology category encoding are fused and mapped to a unified semantic space. Approximate nearest neighbor search and random walk algorithms are combined for association expansion, and multi-objective decision analysis methods are used to comprehensively evaluate the maturity and feasibility of the technology.

5. The multi-granularity technology demand forecasting process according to claim 1, characterized in that, This innovative approach constructs a three-tiered, progressive prediction architecture: "Technology Area Identification → Specialty Area Subdivision → Technical Capability Quantification." Unlike traditional coarse-grained classification methods, it employs a multi-label classifier for first-level technology area prediction, uses conditional random fields to capture dependencies between specialties for second-level subdivision, and utilizes a Transformer-based technical capability requirement prediction model to conduct quantitative evaluation across five dimensions: theoretical research, algorithm design, experimental verification, engineering implementation, and system integration. This achieves comprehensive technical requirement identification from macro-level domains to micro-level capabilities.

6. The structured parsing process according to claim 2, characterized in that, The multi-layer entity recognition adopts a differentiated model strategy. Business entities use a BERT-based BiLSTM-CRF sequence labeling model, technical entities use a pre-trained model, and constraint entities combine regular expressions and deep learning models. Quality control is performed through confidence thresholds to ensure the accuracy and completeness of entity recognition, providing a reliable foundation for building a high-quality entity relationship graph.

7. The technical essence identification process according to claim 3, characterized in that, The complexity quantification assessment adopts a weighted fusion strategy, with entity complexity weight, technology cross-domain degree weight, and constraint complexity weight. The cross-domain degree score is calculated using the Shannon entropy formula, and the complexity is mapped to three levels—low, medium, and high—using a preset threshold, providing quantitative guidance for subsequent retrieval strategies and prediction depth.

8. The intelligent retrieval process according to claim 4, characterized in that, The multi-source data fusion adopts a parallel retrieval and intelligent deduplication strategy, simultaneously searching data sources such as CNKI, the State Intellectual Property Office patent database, and the technology transfer case database. Document fingerprinting technology is used for deduplication, and a multi-dimensional screening mechanism, including TRL level assessment, domain matching degree calculation, and TOPSIS comprehensive scoring, is used to ensure the high quality and diversity of the candidate technology set.

9. The multi-granularity prediction process according to claim 5, characterized in that, The quantitative analysis of technical capabilities adopts a 0-10 scoring system, and combines problem complexity level and technology maturity information for comprehensive evaluation. It generates complete multi-granular analysis results including a list of technical fields, a professional direction spectrum, and a list of technical capability requirements, providing accurate decision support for matching industry-academia-research partners and selecting cooperation models.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the large model-driven enterprise technology demand forecasting method as described in any one of claims 1-9.

Images