Expert screening and recommending method based on large language model and multi-dimensional matching

Abstract

The invention discloses an expert screening and recommendation method based on a large language model and multi-dimensional matching, and the method comprises the following steps: obtaining enterprise related information, including structured data and unstructured data; integrating the data, and constructing a preliminary description of enterprise technology requirements; presetting a stop word list, and removing stop words in the text; the method comprises the following steps: extracting a standardized keyword set of enterprise technology requirements based on an adaptive knowledge extraction method of a large language model; a target research field is matched according to the keyword set, expert information of the field is obtained, and clustering analysis is carried out; a locally deployed optimized BERT variant is adopted, and matching scores of enterprise requirements and expert research fields are calculated through a multi-level semantic matching strategy, including global semantic matching, keyword weighted matching and an optimal matching strategy; screening first K experts based on a multi-dimensional scoring mechanism of academic influence, domain centrality and social network density; outputting a recommended expert list and a cooperative network atlas; the expert matching precision is improved, and the mismatching rate is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of large language model analysis and intelligent recommendation, and in particular to an expert screening and recommendation method based on a large language model and multi-dimensional matching. Background Art

[0002] In the existing technology, there are usually two mainstream methods for matching enterprises with experts. The first type is the traditional method based on keyword matching, such as TF-IDF (Term Frequency-Inverse Document Frequency) and LDA (Latent Dirichlet Allocation) and other topic modeling technologies. These methods calculate the similarity between the research direction of the enterprise and the research field of the expert by counting the word frequency or topic distribution, and recommend relevant experts [1]. However, since enterprise information often comes from a variety of unstructured data sources such as official websites, news reports, industry literature, patent databases, etc., there are problems such as non-standard format, non-standard description, and inconsistent industry terminology. As a result, the matching method based on word frequency statistics is difficult to accurately identify the core research direction of the enterprise. In addition, the keyword matching method ignores the potential connection between the context and the meaning of the word, resulting in unstable matching results and large errors, which cannot meet the needs of high-precision expert recommendation.

[0003] Another type of method is semantic matching technology based on deep learning, which mainly relies on pre-trained language models (such as BERT, GPT, etc.) to perform deep semantic analysis on text and calculate the similarity between enterprise demand text and expert research field text [2]. These methods can improve the accuracy of matching to a certain extent and overcome the limitations of traditional keyword matching methods. However, the current deep learning matching methods still have some problems. For example, most pre-trained models need to call cloud APIs online for reasoning, which makes it difficult to ensure data privacy and may expose enterprise demand information to the risk of leakage. In addition, although the general BERT model can perform deep semantic analysis on text, it may not be suitable for specific industries or academic fields, resulting in semantic bias when the model processes industry terms. Moreover, most existing semantic matching methods only consider a single text similarity calculation, and fail to combine multi-level information fusion technologies such as keyword weighted matching and optimal matching strategies, resulting in a lack of fine-grained adjustment of matching results, making it difficult to meet the needs of enterprises for accurate expert recommendation.

[0004] In terms of expert screening, existing recommendation systems often use a single academic influence indicator (such as h-index, number of papers published, number of citations, etc.) to measure the authority of experts. Although these methods can reflect the academic contributions of experts to a certain extent, they cannot select the most suitable experts. For example, screening methods based on academic influence tend to favor high-output experts, while ignoring some experts who have high influence in the field but less output. In addition, traditional methods fail to consider the status of experts in academic social networks, that is, whether the experts are in the core area of ​​field research and whether they play a key role in the dissemination and development of knowledge in the field. At present, some studies have attempted to use graph neural networks (GNNs) to model academic social networks to calculate the centrality and influence of experts in the field, but there is still a lack of personalized optimization screening mechanisms for different research fields, resulting in the screened experts not necessarily representing the cutting-edge research in the field [3].

[0005] Existing methods for matching companies with experts still have significant shortcomings in terms of company information parsing, semantic matching accuracy, and representative expert screening. The unstructured nature of company information and the diversity of industry terminology make it difficult for traditional keyword matching methods to accurately parse company research directions. While existing semantic matching technology has improved, there is still room for improvement in areas such as privacy protection, industry adaptability, and the diversity of matching strategies. Furthermore, expert screening methods overly rely on a single metric of academic influence and fail to comprehensively consider the expert's status and influence within the field's social network, resulting in the possibility that recommended experts lack representativeness. Summary of the Invention

[0006] Purpose of the invention: The purpose of the present invention is to provide an expert screening and recommendation method based on a large language model and multi-dimensional matching, so as to address the limitations of the existing technology in terms of accurate matching and expert screening.

[0007] Technical solution: The present invention provides an expert screening and recommendation method based on a large language model and multi-dimensional matching, comprising the following steps:

[0008] Step 1: Obtain enterprise-related information, including structured data and unstructured data;

[0009] Step 2: Integrate the data and construct a preliminary description of the enterprise's technical needs;

[0010] Step 3: Preset a stop word list and remove stop words from the text;

[0011] Step 4: Adaptive knowledge extraction based on the Large Language Model (LLM) to extract a standardized set of keywords for enterprise technical requirements.

[0012] Step 5: Match the target research field based on the keyword set, obtain expert information in the field and perform cluster analysis;

[0013] Step 6: Using the locally deployed optimized BERT variant, we calculate the matching scores between enterprise needs and expert research areas through a multi-layered semantic matching strategy, including global semantic matching, keyword-weighted matching, and optimal matching strategy.

[0014] Step 7: Select the top K experts based on a multi-dimensional scoring mechanism based on academic influence, field centrality, and social network density;

[0015] Step 8: Output the recommended expert list and cooperation network map.

[0016] Furthermore, in step 1, the data acquisition method includes a combination of enterprise self-provided data and web crawling, and the data sources include enterprise official websites, patent databases, academic papers and market reports.

[0017] Furthermore, in step 3, the stop word list includes: English stop words, academic general words, quantity measurement words, connecting preposition words, general verb words, time-related words and field-specific stop words.

[0018] Furthermore, in step 4, the standardized keyword extraction method includes: text preprocessing including removal of special characters, sentence processing, and text normalization; generating high-dimensional semantic vectors (Formula 1) through a domain-optimized large language model; using topic modeling technology combining BERTopic+TF-IDF; and keyword context optimization based on the Transformer attention mechanism.

[0019] Furthermore, in step 6, the optimized BERT variant satisfies the following requirements: Domain-Adaptive Fine-Tuning technology is used, and the training dataset contains academic papers and technical documents; the semantic matching score is calculated using a weighted fusion strategy.

[0020] Furthermore, in step 7, logarithmic scaling and square root scaling are introduced into the calculation of academic influence.

[0021] Furthermore, in step 7, the domain centrality calculation includes: calculating the Euclidean distance from the expert to the cluster center through t-SNE dimensionality reduction; and mapping it into a centrality score using a normalization formula.

[0022] Furthermore, in step 7, the social network density calculation includes: constructing an expert collaboration graph based on the co-authorship network; using the graph neural network (GNN) to calculate the average shortest path distance; and combining the PageRank algorithm to evaluate academic communication capabilities.

[0023] Furthermore, in step 7, the final scoring formula is:

[0024] S final(E) = αS academic (E)+βS center (E)+γS density (E)

[0025] Among them, α, β, and γ represent the weights of academic influence, field centrality, and social network density, respectively, and α+β+γ=1.

[0026] Furthermore, in step 8, the output content includes: name, gender, work unit, job title, educational background, professional field, research direction, research results, expert honors, awards, published monographs and papers.

[0027] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: the present invention uses a large language model to analyze the technical intentions behind corporate needs, and combines optimized BERT variants to perform multi-level semantic matching, which can identify the same technical direction behind different expressions, improve the accuracy of expert matching, and reduce the mismatch rate; the present invention adopts multi-dimensional screening mechanisms such as academic influence, field centrality, and cooperation network density to ensure that the recommended experts have more advantages in academic authority and team collaboration capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0029] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0030] like Figure 1 As shown, an embodiment of the present invention provides an expert screening and recommendation method based on a large language model and multi-dimensional matching, comprising the following steps:

[0031] Step 1: Obtain enterprise-related information, including structured data and unstructured data; combine data provided by the enterprise with independent crawling, and data sources include official website of the enterprise, patent database, academic papers and market reports, etc.

[0032] Step 2: Integrate the data to construct a preliminary description of the enterprise's technical needs; use the big prediction model to construct a preliminary description of the enterprise's research direction, ensuring that as much information as possible that can be used for matching is extracted

[0033] Step 3: Preset a stop word list to remove stop words from the text. This includes common English stop words like 'a', 'an', 'and', and 'are'; common academic words like 'study', 'research', and 'paper'; quantity and measurement related words like 'one', 'first', and 'significant'; common conjunctions and prepositions like 'therefore', 'however', and 'to'; general verbs like 'found', 'showed', and 'demonstrated'; time-related words like 'time', 'day', and 'period'; other common academic vocabulary; and more domain-specific stop words. Subsequent analysis will avoid these stop words, significantly improving the efficiency of subsequent processing such as keyword extraction and similarity matching.

[0034] Step 4: Based on the Large Language Model (LLM) adaptive knowledge extraction method, a standardized set of keywords for enterprise technical requirements is extracted; this includes the following steps:

[0035] S41: Text preprocessing. This involves cleaning and formatting the text provided by the enterprise to remove any redundant content and non-standard expressions. This includes removing special characters, splitting sentences, deleting stop words, and normalizing the text to improve the accuracy of subsequent semantic analysis.

[0036] S42: Large language model embedding representation. For the cleaned text data, the present invention uses a locally deployed domain-optimized large language model (such as Fine-tuned SciBERT, LLaMA) to perform deep semantic analysis on the text and obtain a high-dimensional embedding vector representation. This model can understand industry terminology and infer the technical direction of the enterprise based on the context. For the input enterprise text T, its embedding calculation formula is as follows:

[0037] E T =LLM(T) (1)

[0038] Among them, E T It is the semantic vector representation of the enterprise text, which is used for subsequent keyword extraction and matching calculations.

[0039] S43: Topic modeling and keyword extraction. After completing the text vectorization, the present invention uses a method combining BERTopic and TF-IDF to automatically identify the keywords of the company's core research direction from a semantic level. The importance of keywords in the text is calculated based on TF-IDF:

[0040]

[0041] in, For word wi The number of occurrences in the text, N is the total number of all documents, is the number of documents containing the word.

[0042] Subsequently, the present invention uses BERTopic for topic modeling. BERTopic uses HDBSCAN (Hierarchical Density-Based Spatial Clustering) for unsupervised clustering, mapping text embedding vectors to different research topics:

[0043] Cluster(E T )=HDBSCAN(E T ) (3)

[0044] Among them, Cluster(E T ) indicates the subject category to which the text belongs.

[0045] The present invention extracts high-frequency keywords of research directions based on the topic modeling method, and screens out the most representative research direction keywords through a keyword weight calculation method.

[0046] S44: Keyword Optimization. Since some keywords may be redundant or ambiguous, the present invention further adopts a contextual keyword refinement (CKR) strategy to optimize the extracted keywords. This step uses the Transformer-based Attention mechanism to analyze the contextual information of the keywords and remove keywords with low relevance. Finally, the keyword importance weight is calculated as follows:

[0047]

[0048] Among them, α i Calculated by the Attention mechanism, it reflects the importance of keywords in different contexts.

[0049] Finally, the keyword set of enterprise research direction extracted by the present invention can be expressed as:

[0050] K={w1,w2,...,w n} (5)

[0051] Among them, w i These are standardized keywords for the company's research direction. This keyword set will be used in subsequent expert matching calculations. Finally, the company's technical needs will be combined from these keyword sets in a hierarchical structure of keywords and technical fields, and will be retained in order of score.

[0052] Step 5: Match the target research field to the keyword set, obtain expert information in that field, and perform cluster analysis. Specifically, retrieve papers in the target research field from authoritative academic databases and crawl research information from relevant experts. Then, use clustering methods to categorize experts to more clearly define the academic community structure and provide optimized input for subsequent matching steps. Select the three most distinctive subject terms from the research direction classification (e.g., materials science, physics, chemistry, etc.) and the three most representative and distinctive keywords from the current cluster to form the final group name.

[0053] Step 6: Using the locally deployed optimized BERT variant, calculate the matching score between enterprise needs and expert research areas through a multi-level semantic matching strategy, including global semantic matching, keyword weighted matching, and best matching strategy. This includes the following steps:

[0054] S61: Locally deploy an optimized BERT variant. This method uses a lightweight, efficient BERT variant and performs domain-adaptive fine-tuning on it. This fine-tuning process is based on a large-scale dataset of academic papers and technical documents, making the model more adaptable to matching tasks in academic contexts and enterprise technical needs. This step uses local deployment to ensure the security of enterprise data and optimize inference efficiency, allowing large-scale expert matching tasks to be completed within an acceptable timeframe.

[0055] S62: Multi-level semantic matching calculation. To ensure the accuracy of the matching, the present invention adopts a three-level matching strategy, which calculates the matching scores between the enterprise's technical requirements and the expert's research field from three levels: global semantic matching, keyword weighted matching, and optimal matching strategy.

[0056] First layer: global semantic matching

[0057] The core goal of global semantic matching is to measure the overall similarity between enterprise technical requirements text and expert research text in the semantic space. This step uses an optimized BERT variant to calculate text embeddings and uses cosine similarity to measure semantic similarity:

[0058]

[0059] Among them, T represents the technical requirements text of the enterprise; R represents the research field text of the expert; E T and E R The text semantic embedding vectors calculated by BERT respectively.

[0060] This score measures the degree of match between a company's overall needs and the expert's research areas, capturing macro-level similarities. However, due to differences in how company descriptions and expert research areas are expressed, relying solely on global semantic matching can result in some experts with precise matches being overlooked. Therefore, this paper introduces a more refined keyword matching mechanism to enhance fine-grained control over matching.

[0061] Second level: keyword weighted matching

[0062] In order to further improve the matching accuracy, the present invention adopts TF-IDF weighted keyword matching to calculate the similarity between enterprise needs and expert research directions in key terms.

[0063] First, the weights of important keywords in enterprise demand texts and expert research field texts are calculated based on TF-IDF:

[0064]

[0065] Where: w i Represents the i-th keyword in the text; α i Calculated by BERT's Attention mechanism, it indicates the importance of the keyword in the text; TF-IDF(w i )Calculate the weight of keywords to ensure that high-frequency and distinctive words have higher influence.

[0066] Next, the keyword similarity between the enterprise's technical needs and the expert's research direction is calculated based on the keyword weighted matching strategy:

[0067]

[0068] Among them, S BERT (w i ,w j ) indicates the keyword w i (keywords in enterprise needs) and w j The semantic similarity of (keywords in the expert’s research field) is calculated by BERT embedding.

[0069] The third layer: best matching strategy

[0070] Because some enterprise needs may involve multiple technical directions, a single global match or keyword match may lead to underestimation of some research directions. Therefore, the present invention further adopts the best matching strategy, traversing all keywords to find the highest matching score between each enterprise technical requirement word and the expert research direction word:

[0071]

[0072] Among them, SBERT (w i ,w j ) represents the enterprise demand word w i Research field with experts j The semantic similarity between .

[0073] This step ensures that each enterprise demand keyword can find the most matching expert research keyword, which helps to improve the matching stability when the description of some enterprise demand is vague.

[0074] S63: Final matching score calculation

[0075] In order to integrate the results of global semantic matching, keyword weighted matching and best matching strategy, the present invention adopts weighted fusion strategy to calculate the final matching score:

[0076] S final =λ1S global +λ2S keyword +λ3S best (10)

[0077] Where λ1, λ2, and λ3 represent the weights of global semantic matching, keyword weighted matching, and the best matching strategy, respectively, and λ1+λ2+λ3=1. Extensive testing by the present invention has shown that when λ1 is set to 0.4, λ2 is set to 0.3, and λ3 is set to 0.4, the system achieves optimal matching results, and the predicted fields are most closely aligned with the technical needs of enterprises.

[0078] Step 7: Select the top K experts based on a multi-dimensional scoring mechanism based on academic influence, field centrality, and social network density. This includes the following steps:

[0079] S71: The present invention introduces log-scaling and sqrt-scaling based on academic influence to reduce the impact of extreme values ​​and make the scoring more stable.

[0080] Academic score calculation formula:

[0081]

[0082] Among them, P E represents the number of papers published by expert E; C E represents the total number of citations of expert EE; C max The most cited expert in the current research field; p and w c are the weights of the number of published papers and the number of citations, which are set to 0.3 and 0.7 respectively in the present invention.

[0083] S702: Domain Centrality Calculation. Domain centrality measures the importance of an expert within their research field, specifically, whether the expert is at the core of academic research. This method uses t-Distributed Stochastic Neighbor Embedding (TSNE) for dimensionality reduction, calculates the Euclidean distance from the expert to the cluster center, and uses a normalized formula to map this distance to a centrality score.

[0084]

[0085] Among them, d E is the distance from the expert EE to the cluster center; d max is the maximum distance from all experts in the research field to the cluster center.

[0086] S703: Social Network Density Calculation

[0087] Social network density measures the academic collaboration of experts in their research fields. This paper constructs an expert collaboration graph based on the co-authorship network of academic papers and calculates the academic communication capabilities of experts in social networks through graph neural networks (GNN).

[0088] The present invention uses the inverse weighted summation method to calculate the expert's density score:

[0089]

[0090] Among them, d E It represents the average shortest path distance of experts in the collaboration network, that is, the degree of connection between experts and other researchers in the field.

[0091] In addition, the present invention further combines PageRank influence calculation to measure the importance of experts in academic communication:

[0092]

[0093] Among them, N(E) is the set of cooperative neighbors of expert EE, d j is the number of cooperative connections of neighbor jj.

[0094] S704: Final expert score calculation and screening. The present invention weightedly integrates the scores of academic influence, field centrality, and social network density to calculate the final comprehensive score of the expert:

[0095] S final (E) = αS academic (E)+βS center (E)+γS density (E) (15)

[0096] Here, α, β, and γ represent the weights of academic influence, field centrality, and social network density, respectively, and α + β + γ = 1. After extensive testing, the present invention found that when α, β, and γ were 0.5, 0.2, and 0.3, respectively, the results were better, and the selected authors had high academic level and were more in line with the technical needs of enterprises.

[0097] Finally, according to the calculated S final (E) value sorting, select the top K experts with the highest scores as the final recommendation experts.

[0098] Step 8: Output the recommended expert list and collaborative network map. The output expert information includes: name, gender, work unit, job title, educational background, professional field, research direction, research achievements, expert honors, awards, published monographs and papers, and academic circles, showing information such as fellow students, classmates, and collaborative relationships.

[0099] Example 1: Recommendation for cooperation projects with high-tech enterprises

[0100] When high-tech companies explore new technology R&D collaborations, finding the right team of experts is crucial. However, traditional expert search methods often rely on a company's own connections or simple keyword searches, which cannot guarantee the recommended experts' high technical compatibility or practical value in collaboration.

[0101] This method extracts standardized enterprise technology requirements and uses a large language model to analyze the company's research directions and technical keywords, providing a deep understanding of the company's innovation needs. Subsequently, using an optimized BERT variant model and a multi-level semantic matching strategy, it conducts a multi-dimensional match between enterprise technology requirements and expert research areas, screening candidate expert groups.

[0102] To further ensure the representativeness and authority of experts, this paper introduces a multi-dimensional screening mechanism that comprehensively evaluates candidate experts' academic influence (number of paper citations, journal publication ranking), field centrality (closeness of collaboration with core teams in the target field), and social network density (the breadth and depth of connections in academic and industry networks). Ultimately, the top K experts with the greatest academic authority and industry influence are selected from the candidate pool and provided to companies as a recommended list of potential partners.

[0103] Example 2: Industry Alliance and Standard Setting Expert Recommendations

[0104] Amid the rapid development of emerging industries and the transformation and upgrading of traditional industries, industry alliances and standards-setting organizations play a crucial role in guiding industry direction and establishing technical specifications. However, the selection of experts to participate in alliance formation and standard-setting directly impacts the alliance's credibility, the standards' technological foresight, and market acceptance.

[0105] This method can analyze the core technology directions in the alliance's target areas during the initial stages of alliance preparation, generating standardized technical descriptions and keyword sets. Subsequently, an optimized multi-level semantic matching strategy is used to match expert groups highly aligned with the target technology directions from global academic and industry databases.

[0106] During the screening process, this paper introduces a multi-dimensional scoring mechanism that comprehensively considers experts' academic influence (including published papers, H-index, and patent achievements), their domain centrality (collaborative networks with other well-known experts in the field), and their experience and voice in the development of industry standards. Furthermore, combined with social network density analysis, this ensures that recommended experts possess industry influence and cross-disciplinary collaboration capabilities, enabling them to serve as both a guide and a bridge within the alliance.

[0107] The final list of recommended experts ensures that the alliance can absorb a diverse team of academic authorities and industry experts, provide technical endorsement for the formulation of industry standards, enhance the authority and market acceptance of the standards, and thus promote the healthy development of the industry ecosystem.

[0108] Example 3: Selection of experts for government scientific research project review

[0109] During the review process for government science and technology projects, ensuring that review experts possess the necessary expertise and authority while avoiding potential conflicts of interest is crucial for ensuring quality and fairness. Traditional expert database matching methods typically only consider experts' expertise, neglecting their academic standing, social networks, and even their connections to the project applicants. This can easily lead to unfair reviews or a lack of authoritative opinions.

[0110] During the application phase of government science and technology projects, this method can deeply analyze project materials, extract technical keywords and innovation directions, and identify the sub-sectors and interdisciplinary areas involved in the application based on a large language model. Combined with the optimized BERT multi-level semantic matching model, it can screen out expert groups that meet the field requirements from the expert database.

[0111] On this basis, a multi-dimensional screening and scoring mechanism is introduced to further evaluate the academic influence of candidate experts (including paper achievements, scientific research project experience, and review history), field centrality (closeness of cooperation with other top experts in the field), and combined with social network density analysis to ensure the comprehensive influence of recommended experts in academia and industry. At the same time, the present invention can also automatically identify possible conflicts of interest by combining the academic cooperation records between experts and the applicant units, giving priority to recommending experts with matching academic backgrounds and no direct cooperation relationship with the applicant units, thus ensuring the fairness and impartiality of the review process.

[0112] Example 4: Recommendations from experts in clinical research and innovation cooperation in the medical field

[0113] In clinical research and medical innovation projects, finding the right experts to guide research, tackle technical challenges, and translate research findings into practical applications is crucial for increasing research success and impact. However, due to the numerous sub-fields and complex interdisciplinary approaches involved in medicine, traditional expert recommendation methods struggle to accurately match teams with both deep expertise and cross-disciplinary backgrounds.

[0114] The application of this invention in medical scenarios can first conduct an in-depth analysis of the background information of clinical research projects to extract core diseases, research directions, and innovative technology keywords. Through a large language model, the types of diseases involved, medical technology paths, and possible interdisciplinary fields involved (such as bioinformatics, drug development, artificial intelligence-assisted diagnosis, etc.) are identified. Subsequently, using the optimized BERT multi-level semantic matching strategy, expert groups that meet the needs of the field are screened from medical academic databases and scientific research institution databases, including clinicians, medical researchers, and interdisciplinary technology experts.

[0115] During the screening phase, experts are comprehensively evaluated based on their academic influence (publications in top medical journals, clinical research experience, and medical patent contributions), field centrality (research collaboration networks in related medical fields), and social network density (connections with peers and cross-disciplinary experts within the medical field). Ultimately, a list of representative experts with both clinical experience and academic innovation capabilities, as well as experience collaborating in multidisciplinary fields, is selected and recommended to the clinical research team.

[0116] Example 5: Recommendations from FinTech Experts

[0117] Amidst the rapid development of FinTech, emerging technologies such as blockchain, big data, and artificial intelligence are deeply integrated with traditional financial services, creating both innovative opportunities and regulatory challenges. Financial institutions often require multidisciplinary teams of experts, including those in finance, computer science, data science, and legal compliance, to implement innovative FinTech projects, optimize risk modeling, and design financial products. However, traditional expert selection methods struggle to capture this diverse talent pool, leading to one-sided technical solutions and inadequate compliance risk assessments.

[0118] This method can be used to provide expert recommendations in the fintech field. It first analyzes project descriptions provided by financial institutions, extracting keywords related to financial business logic and technical direction, as well as regulatory compliance clauses. Using a large language model, it performs semantic structured analysis on this information to identify the multi-dimensional technical directions required by fintech projects, such as innovative elements such as intelligent risk control, financial big data, and blockchain evidence storage.

[0119] Subsequently, combined with an optimized BERT multi-level matching strategy, a matching cross-disciplinary expert group was screened from fintech databases, academic databases, and legal and regulatory databases. A multi-dimensional screening and scoring mechanism further assessed the authority and relevance of candidate experts in the financial field.

[0120] The expert team finally recommended can simultaneously cover the three key dimensions of financial business logic, technological innovation path, and legal compliance review, helping financial institutions to balance technological advancement and compliance security in innovative projects, reduce business innovation risks, and accelerate the implementation of financial technology.

Claims

1. An expert screening and recommendation method based on a large language model and multi-dimensional matching, characterized in that: The following steps are involved: Step 1: Obtain enterprise-related information, including structured data and unstructured data; Step 2: Integrate the data and construct a preliminary description of the enterprise's technical needs; Step 3: Preset a stop word list and remove stop words from the text; Step 4: Adaptive knowledge extraction based on a large language model to extract a standardized set of keywords for enterprise technical requirements; Step 5: Match the target research field based on the keyword set, obtain expert information in the field and perform cluster analysis; Step 6: Using the locally deployed optimized BERT variant, we calculate the matching scores between enterprise needs and expert research areas through a multi-layered semantic matching strategy, including global semantic matching, keyword-weighted matching, and optimal matching strategy. Step 7: Select the top K experts based on a multi-dimensional scoring mechanism based on academic influence, field centrality, and social network density; Step 8: Output the recommended expert list and cooperation network map.

2. The expert screening and recommendation method based on a large language model and multi-dimensional matching according to claim 1, characterized in that: In step 1, data acquisition methods include a combination of enterprise self-provided data and web crawling, and data sources include enterprise official websites, patent databases, academic papers and market reports.

3. The expert screening and recommendation method based on a large language model and multi-dimensional matching according to claim 1, characterized in that: In step 3, the stop word list includes: English stop words, academic general words, quantity measurement words, connecting preposition words, general verb words, time-related words and domain-specific stop words.

4. The expert screening and recommendation method based on a large language model and multi-dimensional matching according to claim 1, characterized in that: In step 4, the standardized keyword extraction method includes: text preprocessing including removing special characters, sentence processing, and text normalization; generating high-dimensional semantic vectors through a domain-optimized large language model; using topic modeling technology combining BERTopic+TF-IDF; and keyword context optimization based on the Transformer attention mechanism.

5. The expert screening and recommendation method based on a large language model and multi-dimensional matching according to claim 1, characterized in that: In step 6, we optimize the BERT variant to meet the following requirements: We use Domain-Adaptive Fine-Tuning (DoFTU) with a training dataset consisting of academic papers and technical documentation. The semantic matching score is calculated using a weighted fusion strategy.

6. The expert screening and recommendation method based on a large language model and multi-dimensional matching according to claim 1, characterized in that: In step 7, logarithmic scaling and square root scaling are introduced into the calculation of academic influence.

7. The expert screening and recommendation method based on a large language model and multi-dimensional matching according to claim 1, characterized in that: In step 7, the domain centrality calculation includes: calculating the Euclidean distance from the expert to the cluster center through t-SNE dimensionality reduction; and mapping it into a centrality score using a normalized formula.

8. The expert screening and recommendation method based on a large language model and multi-dimensional matching according to claim 1, characterized in that: In step 7, the calculation of social network density includes: constructing an expert collaboration graph based on the co-authorship network; using graph neural network (GNN) to calculate the average shortest path distance; and combining the PageRank algorithm to evaluate academic communication capabilities.

9. The expert screening and recommendation method based on a large language model and multi-dimensional matching according to claim 1, characterized in that: In step 7, the final scoring formula is: S final (E)=αS academic (E)+βS center (E)+γS density (AND) Among them, α, β, and γ represent the weights of academic influence, field centrality, and social network density, respectively, and α+β+γ=1.

10. The expert screening and recommendation method based on a large language model and multi-dimensional matching according to claim 1, characterized in that: In step 8, the output content includes: name, gender, work unit, job title, educational background, professional field, research direction, research results, expert honors, awards, published monographs and papers.

Images