Expert and project matching method based on AI big language model

Abstract

The invention discloses an expert and project matching method based on an AI large language model. The method comprises the following steps: S1) data preprocessing: performing data cleaning, denoising, word segmentation and labeling processing on information of projects and experts; s2) text representation and feature extraction: performing vectorization representation on the text information of the project and the expert by using an AI large language model, and extracting key features; s3) a plurality of vector models are used to calculate similarity conditions between the project and the experts, a set of inconsistent data is taken to carry out erank reordering, if the data does not conform to expectation, multi-dimensional calculation of a business similarity radar map is carried out, and the most matched expert is recommended through a weighting mode to carry out project review; and S4) result optimization and feedback: according to user feedback and model self-learning, through iterative training and parameter adjustment, continuously optimizing a matching algorithm and a result, the matching accuracy and efficiency can be improved, and it is ensured that the most matched expert is recommended for project review.

Description

Background Art

[0002] In today's scientific research, education, business management and other fields, project review is an important part of ensuring project quality, promoting academic exchanges and technological innovation. However, with the surge in the number of projects and increasingly stringent review requirements, how to quickly and accurately match the right experts for project review has become an urgent problem to be solved.

[0003] The current project review has the following characteristics:

[0004] 1. The importance of project review

[0005] Project review is a process of comprehensive evaluation of project plans, implementation processes and results, aiming to ensure the scientificity, innovation and practicality of the project. Through review, we can find the highlights and shortcomings of the project and provide direction for subsequent improvement and optimization of the project. At the same time, review is also an important platform for academic exchanges, which helps promote discipline development and technological innovation.

[0006] 2. Scarcity of expert resources

[0007] As the number of projects continues to increase, the demand for expert resources is also growing. However, expert resources are limited and are often distributed in different fields and institutions. How to effectively integrate and utilize expert resources to ensure that each project can be reviewed by appropriate experts has become a challenge.

[0008] 3. Limitations of traditional matching methods

[0009] Traditional methods of matching experts with review project information often rely on manual screening and recommendation, which has problems such as low efficiency, strong subjectivity, and narrow coverage. In addition, with the continuous subdivision and intersection of project fields, traditional methods are increasingly unable to meet the complex and changing review needs.

[0010] As can be seen from the above, traditional expert extraction methods often rely on manual annotation and rule matching, which have problems such as low efficiency, high cost, insufficient coverage, and low accuracy. In addition, with the continuous increase in data volume and the continuous subdivision of fields, traditional methods are increasingly unable to meet actual needs.

[0011] At the same time, with the rapid development of artificial intelligence, the AI ​​big language model, with its powerful natural language processing and knowledge reasoning capabilities, provides a new solution for matching experts with review project information. The advantages of the AI ​​big language model are as follows:

[0012] 1. Natural language processing capabilities

[0013] AI big language models have mastered rich language knowledge and contextual understanding capabilities through pre-training on large-scale text data. They can accurately understand the research content, technical routes, and innovations of the project, and provide precise language support for expert matching.

[0014] 2. Knowledge reasoning ability

[0015] AI large language models can not only understand the surface information of the text, but also perform deep knowledge reasoning. They can identify key concepts, technical difficulties, potential risks and other information in the project, and recommend experts with relevant experience and expertise for review.

[0016] 3. Automation and intelligence

[0017] The expert and review project information matching system based on the AI ​​large language model can realize an automated and intelligent matching process. The system can automatically extract project information and expert information from massive data, and quickly match and recommend through algorithm models. This not only improves matching efficiency, but also reduces labor costs.

[0018] Traditional expert extraction methods often rely on manual labeling and rule matching, which have problems such as low efficiency, high cost, insufficient coverage, and low accuracy. In addition, with the continuous increase in data volume and the continuous subdivision of fields, traditional methods are increasingly unable to meet actual needs. Therefore, it is necessary to provide an expert and project matching method based on an AI large language model to solve the above problems. Summary of the invention

[0019] The technical problem to be solved by the present invention is to provide an expert and project matching method based on an AI large language model, which can continuously optimize the matching algorithm and results, improve the accuracy and efficiency of matching, and ensure that the most matching expert can be recommended for project review.

[0020] The technical solution adopted by the present invention to solve the above-mentioned technical problems is to provide an expert and project matching method based on an AI large language model, comprising the following steps: S1) data preprocessing: data cleaning, denoising, word segmentation and annotation processing are performed on the information of projects and experts, so that the input data is deduplicated and consistent; S2) text representation and feature extraction: the text information of projects and experts is vectorized by using the AI ​​large language model, and key features are extracted; S3) similarity calculation and multi-dimensional matching: multiple vector models are used to calculate the similarity between projects and experts, and sort them according to the similarity values. If the sorting results calculated by the multiple vector models are inconsistent, a collection of inconsistent data is taken and rerank is performed. If the reranked results do not meet expectations, a multi-dimensional calculation of a business similarity radar chart is performed for each project and each expert in combination with specific business attributes, and the most matching expert is recommended for project review in a weighted manner; S4) result optimization and feedback: according to user feedback and model self-learning, the matching algorithm and results are continuously optimized through iterative training and parameter adjustment until the response time and matching accuracy reach the preset threshold.

[0021] Furthermore, the data cleaning in step S1 includes: removing duplicate records through a unique identifier, the unique identifier is a project ID or an expert ID; for missing values ​​in key fields, filling in the missing values ​​or directly deleting the record; the key fields are expert research fields and / or project technical details, and the filling method is mean filling, mode filling or specific value filling; identifying and processing outliers, including non-text data appearing in text fields, or numerical data that exceeds a reasonable range.

[0022] Furthermore, the data denoising in step S1 includes: removing HTML tags, special characters and irrelevant punctuation marks in the text, and removing stop words.

[0023] Furthermore, the step S1 performs word segmentation processing on the project description and expert profile according to a Chinese word segmentation tool, and annotates the expert research field and project technical keywords.

[0024] Furthermore, the key features in step S2 include the technical field, industry field, scientific field of the project and the expert, as well as the expert's research direction, research results, and review experience.

[0025] Furthermore, in step S3, the n text embedding models are used as vector models to calculate similarity. When the similarity results calculated by the n vector models are inconsistent, a collection is taken, and the rerank model is used to sort the relevance of the taken collection. The calculation formula is as follows:

[0026] Dataset = model_rerank(model1_result(query, raw_dataset), model2_result(query, raw_dataset), ..........modeln_reset(query, raw_dataset)); Query is the extracted key features of the project to be reviewed, and raw_dataset is the feature data set of multiple candidate experts.

[0027] Furthermore, the value of n is 2, the vector model used in step S3 is the text-ebedding-ada-002 model and the BAAI / bge-large-zh-v1.5 model of OpenAI, and the rerank model used is the BAAI / bge-reranker-large model.

[0028] Furthermore, the step S3 recommends the most suitable expert for project review in the following weighted manner: a k are weights of different dimensions.

[0029] Furthermore, step S3 selects 6 dimensions to calculate the expert radar chart, the value of n is 6, the weight a1 is the expert's technical field, the weight a2 is the expert's industry field, the weight a3 is the expert's scientific field, the weight a4 is the expert's research direction, the weight a5 is the expert's research results, and the weight a6 is the expert's review experience.

[0030] Furthermore, the step S3 calculates the similarity between the two experts and the corresponding projects as follows: first, the feature values ​​of the two candidate experts in the six dimensions are vectorized through the model to obtain a vector matrix embeddings_1, and then the key features of the extracted projects to be reviewed are vectorized to obtain another vector matrix embeddings_2, and finally the vector matrix embeddings_1 and the transposed matrix of the vector matrix embeddings_2 are multiplied to obtain the similarity matrix; each element (i, j) in the similarity matrix represents the similarity or compatibility score between the i-th data point in embeddings_1 and the j-th data point in embeddings_2; the step S4 first assumes that the weights a of the six dimensions are k are all 1, the top-ranked candidate expert is obtained, and then the value of the review experience a6 is increased. If the similarity or compatibility score of the top-ranked candidate expert further increases, the candidate expert is selected.

[0031] Compared with the prior art, the present invention has the following beneficial effects: the expert and project matching method based on the AI ​​large language model provided by the present invention uses the AI ​​large language model to vectorize the text information of projects and experts; by calculating the similarity between projects and experts, the expert and review project information are matched; based on user feedback and model self-learning, the matching algorithm and results are continuously optimized to improve the accuracy and efficiency of matching, ensuring that the most matching expert can be recommended for project review. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flow chart of expert and project matching based on the AI ​​large language model of the present invention;

[0033] Figure 2 This is the radar chart of expert No. 7 in the embodiment of the present invention;

[0034] Figure 3 This is the radar chart of expert No. 1 in the embodiment of the present invention. DETAILED DESCRIPTION

[0035] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0036] Figure 1 This is a flowchart of the project review process of selecting experts based on the pre-trained sentence representation model in the present invention.

[0037] See also Figure 1 The expert and project matching method based on the AI ​​large language model provided by the present invention includes the following steps: 1. Data preprocessing

[0038] Objective: To ensure the quality and consistency of input data and provide a clean and accurate data foundation for subsequent processing; the specific steps and methods are as follows.

[0039] 1.1 Data cleaning:

[0040] Deduplication: Remove duplicate records by unique identifiers (such as project ID, expert ID).

[0041] Missing value processing: For missing values ​​in key fields (such as expert research fields and project technical details), fill in the missing values.

[0042] (such as mean filling, mode filling or specific value filling) or delete the records.

[0043] Outlier detection and processing: Identify and process outliers, such as non-text data appearing in text fields, or numerical data that is outside a reasonable range.

[0044] 1.2 Denoising:

[0045] Text denoising: Remove noises such as HTML tags, special characters, and irrelevant punctuation marks in the text.

[0046] Stop word removal: Remove common but meaningless words in the text, such as "de" (of), "shi" (is), etc.

[0047] 1.3 Word segmentation:

[0048] Perform word segmentation on the project description and expert profile according to Chinese word segmentation tools (such as Jieba), for subsequent word frequency statistics and feature extraction.

[0049] 1.4 Annotation:

[0050] Annotate specific information (such as expert research fields, project technical keywords) for subsequent feature extraction and similarity calculation.

[0051] 2. Text representation and feature extraction

[0052] Objective: Convert text information into a numerical form that can be understood by machines and extract key features; the specific steps and methods are as follows.

[0053] 2.1 Text vectorization:

[0054] Use the embedding layer of the AI large language model to convert the text into a high-dimensional vector representation.

[0055] These vectors can capture the semantic information of the text and provide a basis for subsequent similarity calculation.

[0056] 2.2 Feature extraction:

[0057] Domain features: Identify the technical fields, industry fields, etc. of the project and experts through text analysis.

[0058] Expert features: Extract key information such as the research direction of experts, research achievements (such as the number of papers, citation times), and review experience.

[0059] Project features: Include the technical difficulty, budget, time requirements, etc. of the project.

[0060] 2.3 Feature engineering:

[0061] Standardize and normalize the extracted features to ensure comparability between different features.

[0062] Construct feature vectors for subsequent similarity calculation and model training.

[0063] 3. Similarity calculation and multi-dimensional matching

[0064] Objective: Achieve precise matching by calculating the similarity between projects and experts; the specific steps and methods are as follows.

[0065] 3.1 Similarity calculation:

[0066] The similarity between item and expert feature vectors is calculated using a vector model.

[0067] 3.2 Multidimensional matching:

[0068] Rerank is performed based on the results of multiple similarity algorithms, taking into account the matching degree in different dimensions.

[0069] Draw a radar chart to display the multi-dimensional matching of projects and experts, and recommend the most matching experts in a weighted manner.

[0070] The weighting strategy can be derived based on business rules.

[0071] 3.3 Matching result optimization:

[0072] Based on historical matching results and feedback data, the matching algorithm and weight distribution are continuously optimized to improve matching accuracy.

[0073] 4. Result optimization and feedback

[0074] Objective: Continuously optimize matching results through user feedback and model self-learning; the specific steps and methods are as follows.

[0075] 4.1 User feedback collection:

[0076] Design a user feedback mechanism to collect users’ satisfaction evaluation and suggestions on matching results.

[0077] Users are encouraged to submit detailed feedback, including reasons for dissatisfaction and suggestions for improvement.

[0078] 4.2 Model Self-Learning:

[0079] Use machine learning techniques to automatically adjust model parameters and matching algorithms based on user feedback.

[0080] Iterate the training model to gradually optimize the accuracy and efficiency of the matching results.

[0081] 4.3 Performance monitoring and optimization:

[0082] Real-time monitoring of system performance indicators (such as response time, matching accuracy) and user behavior data.

[0083] Regularly evaluate and optimize system performance to ensure system stability and efficiency.

[0084] A specific embodiment is given below. The AI ​​large language model used in the present invention is mainly a Text Embedding model. Text Embedding is a core technology in natural language processing (NLP). It refers to converting text (such as words, phrases, sentences or paragraphs) into fixed-size real number vectors. These vectors can capture the semantic information in the text, so that semantically similar texts have similar vector representations in the embedding space. This conversion enables text data to be understood and processed by computers, so that it can be used in various NLP tasks.

[0085] To determine which text embedding models work better, an evaluation metric is usually needed for comparison. MTEB (Massive Text Embedding Benchmark) is an evaluation benchmark for massive text embedding models. C_MTEB is the most comprehensive Chinese semantic vector evaluation benchmark recognized by the industry. Through this benchmark, the present invention found that no specific text embedding method dominates all tasks, which shows that the field has not yet focused on a universal text embedding method and expanded it to provide state-of-the-art results on all embedding tasks. In the Retrieval field, the text embedding models are ranked as follows (data as of July 31, 2024, for the complete list, please refer to

[0086] https: / / github.com / FlagOpen / FlagEmbedding / tree / master / C_MTEB#leaderboard):

[0087]

[0088] At present, the reranking dataset for general scenarios is mainly T2Reranking. It can be seen that OpenAI's text-ebedding-ada-002 model has the highest score for general scenarios. In terms of comprehensive average score, the bge-large-zh model has the highest score. Therefore, the basic Text Embedding models selected in this paper mainly include OpenAI's text-ebedding-ada-002 model and BAAI / bge-large-zh-v1.5 model.

[0089] The text-ebedding-ada-002 model is an embedding model provided by OpenAI in December 2022. Its main features are:

[0090] Unified capabilities: By merging five separate models (text similarity, text search-query, text search-document, code search-text, and code search-code) into a new model, this single representation outperforms previous embedding models on a range of different text search, sentence similarity, and code search benchmarks;

[0091] Context: The context length is 8192, which makes it more convenient when processing long documents;

[0092] Embedding Size: The embedding size of this model is only 1536 dimensions.

[0093] The BAAI / bge-large-zh-v1.5 model is a Chinese text embedding model developed by the Zhiyuan Research Institute. The model performs well in tasks such as semantic retrieval, classification, clustering, and semantic matching. It supports a context length of 512 and a vector dimension of 1024. This makes the new embedding more cost-effective when processing vector databases.

[0094] Its main features are: 1. Efficient pre-training and large-scale text fine-tuning; 2. The RetroMAE pre-training algorithm is used on two large-scale corpora to further enhance the semantic representation ability of the model; 3. The discriminative power of semantic vectors is enhanced through negative sampling and hard negative example mining; 4. Drawing on the strategy of Instruction Tuning, the general ability in multi-task scenarios is enhanced.

[0095] The present invention takes the following expert feature data set for testing:

[0096] sentences_1=[

[0097] '1, Urban and rural planning design and theory; Smart city service support technology, rural planning and construction, national land space planning, planning implementation, planning regulations, planning management',

[0098] '2, Piezoelectric and ferroelectric ceramic materials, functional ceramic composite materials, crystalline solid material chemistry, fuel cells, functional ceramic preparation technology, energy-saving and new energy material preparation technology, fuel cell technology, material chemistry, fuel cells, electronic ceramics, defect chemistry',

[0099] '3,"Magnetic resonance structural imaging and disease diagnosis, fMRI and brain and spinal cord dysfunction detection, X-ray and CT, electron and ion beam, radiological diagnosis and quality control, magnetic resonance imaging technology and contrast agents, interventional medicine and engineering", medical imaging diagnosis technology, general hospital,"imaging diagnosis,"',

[0100] '4,"Tumor recurrence and metastasis, Tumor diagnosis, Comprehensive treatment of tumors, Digestive system tumors, Tumor physical therapy",,,"Application of new radiotherapy technology, Comprehensive treatment of digestive tract tumors, Individualized treatment of tumors,"',

[0101] '5,"Basic methods for building secure and reliable intelligent systems, new intelligent technologies and applications, intelligent human-computer interaction, social media processing and cross-media analysis, multimodal perception and scenario computing","Network and communication security technology, security evaluation technology, system and software security technology, energy-saving monitoring technology, Internet of Things equipment, components and networking technology",,"Artificial intelligence, new energy, intelligent human-computer interaction, Internet of Things, integrated circuits,"',

[0102] '6,"Digestive system tumors, Endocrine system tumors, Tumorigenesis",,,"Pathology, Oncology,"',

[0103] '7,"Active / passive fiber optic devices, optical and fiber optic sensing materials, devices and technologies, microwave photon transmission chips and technologies, laser technology and applications, ultrafast, ultra-strong and ultra-narrow linewidth laser technology","Integrated optoelectronic device design, manufacturing and process technology, optical transmission system technology, Internet of Things equipment, components and networking technology, sensitive components and sensors, new sensors",,"Fiber optic sensing, fiber optic devices, time-frequency transmission, laser technology,"',

[0104] '8,"Cytoskeleton, Cellular Material Transport, Cell Membrane and Organelles, Cell Movement, Extracellular Matrix and Microenvironment",,,"Intermediate Filaments, Microfilaments, Pathogen-Host Interaction, Cell Movement, Cytoskeleton,"',

[0105] '9,"Computer and Mathematics Intersection, Electronic Communication and Mathematics Intersection","Cryptography, Authentication and Authorization Technology, Security and Confidentiality Technology, Application Security Technology",,"Information Security, Blockchain, Cryptography, Authentication, Privacy Protection,"',

[0106] '10,"Building energy-saving technology, high-temperature heat pump technology, thermal storage combustion technology, energy recovery technology, energy system management, optimization and control technology",,"Building energy-saving, air conditioning system design, energy station design, ventilation design,"' ]

[0108] Query = ["computer related"]

[0109] The results of similarity calculation of the above data using the BAAI / bge-large-zh-v1.5 model are as follows: [[0.494764] [ 0.42190862] [0.46524864] [ 0.38506556] [ 0.52790684] [ 0.406992] [0.5105162] [ 0.40613496] [0.5464188] [0.43746695]]

[0120] It can be seen that the 9th item has the highest relevance. The ranking order by relevance is:

[0121] [9,5,7,3,1,10,2,8,4,6]

[0122] The similarity results calculated by the text-ebedding-ada-002 model are ranked in order of relevance:

[0123] [9,5,1,7,10,3,8,6,4,2]

[0124] If the business requirement of the present invention is to find two experts related to computers, then, without a doubt, the 9th and 5th can be taken. However, if the business requirement is to take the top three experts related to computers, then the two models will conflict on who should be ranked third. The BAAI / bge-large-zh-v1.5 model believes that it should be expert No. 7, while the text-ebedding-ada-002 model believes that it should be expert No. 1. At this time, the collection can be taken and reranked, that is, the 7th and 1st experts can be reranked. According to the authoritative rerank model list, BAAI / bge-reranker-large performs best, so the model is used to sort the above collection by relevance, and the result is: [7,1], so the present invention selects expert No. 7 as the third relevant expert.

[0125] The above matching algorithm can be expressed by this formula:

[0126] Dataset=model_rerank(model1_result(query,raw_dataset),model2_result(query,raw_dataset))

[0127] The general formula is:

[0128] Dataset=model_rerank(model1_result(query,raw_dataset),model2_result(query,raw_dataset),.........model n_reset(query,raw_dataset))

[0129] In the above example, n = 2, query = "computer related", raw_dataset = the feature data set of experts No. 1 to No. 10.

[0130] But look at the characteristics of expert No. 7: '7,"Active / passive fiber optic devices, optical and fiber optic sensing materials, devices and technologies, microwave photon transmission chips and technologies, laser technology and applications, ultrafast, ultra-strong and ultra-narrow linewidth laser technology","Integrated optoelectronic device design, manufacturing and process technology, optical transmission system technology, Internet of Things equipment, components and networking technology, sensitive components and sensors, new sensors","Fiber optic sensing, fiber optic devices, time-frequency transmission, laser technology"', from common sense, it is not highly relevant to computers. At this time, in order to optimize the expert matching results, the present invention uses a radar chart to display the multi-dimensional matching of projects and experts, and recommends the most matching experts in a weighted manner.

[0131] The dimensions selected by the present invention are 6 dimensions: the expert's technical field (weight a1), industry field (weight a2), scientific field (weight a3), and the expert's research direction (weight a4), research results (weight a5), and review experience (weight a6). Take expert No. 7 and expert No. 1 as examples, and extract the characteristic values ​​of these two experts in these 6 fields:

[0132] The characteristics of expert No. 7 in these 6 areas are as follows:

[0133] Scientific field: Optics and fiber optic sensing materials

[0134] Technical field: IoT devices, components and networking technologies, new sensors

[0135] Industry fields: Fiber optic sensing, fiber optic devices, time-frequency transmission, laser technology

[0136] Research direction: Fiber optic gyroscope and resonant fiber optic gyroscope

[0137] Research results: Fiber Optic Sensor Networks for the Internet of Things

[0138] Review experience: Reviewed more than 10 projects related to new optical fiber sensing materials and devices

[0139] The characteristics of the No. 1 expert in these 6 areas are as follows:

[0140] Scientific field: Urban and rural planning design and theory

[0141] Technical field: Smart city service support technology

[0142] Industry: Rural planning and construction

[0143] Research direction: National land space planning, planning implementation

[0144] Research results: AI-driven planning management approach

[0145] Review experience: Reviewed more than 16 computer system planning and design related projects

[0146] The feature of the project to be reviewed is extracted as "computer related".

[0147] The BAAI / bge-large-zh-v1.5 model is used to calculate the correlation between the corresponding projects of the two experts. First, the eigenvalues ​​of expert No. 7 and expert No. 1 in the six fields are vectorized through the model to obtain a vector matrix embeddings_1. Then, "computer related" is vectorized to obtain another vector matrix embeddings_2. Similarity = embeddings_1@embeddings_2.T is used to calculate the similarity.

[0148] The shape of embeddings_1 is (m=12, d=1024), where m is the number of data points in the first set and d is the dimension of the embedding vector.

[0149] The shape of embeddings_2 is (n=1, d=1024), where n is the number of data points in the second set and d also represents the dimension of the embedding vector.

[0150] embeddings_2.T is the transposed matrix of embeddings_2, and its shape becomes (d=1024,n=1).

[0151] When embeddings_1@embeddings_2.T is executed, this is actually performing matrix multiplication. In Python's math library NumPy or PyTorch, the @ symbol is used to represent matrix multiplication. The result of this operation is a matrix with a shape of (m=12,n=1), called the similarity matrix.

[0152] Each element (i, j) in the similarity matrix represents the similarity or compatibility score between the i-th data point in embeddings_1 and the j-th data point in embeddings_2. This score is usually calculated by dot product, which is the sum of the corresponding elements of the two vectors multiplied together. Dot product is a common method to measure the similarity of two vectors. The larger the value, the more similar the two vectors are (in direction).

[0153] The similarity matrix results of expert No. 7 and expert No. 1 are as follows: [ 0.35086593] [0.47247428] [0.46333694] [ 0.33570868] [0.32512718] [ 0.39716867] [0.3634815] [0.3618956] [0.36044583] [ 0.38608974] [0.37323153] [0.5609167]

[0166] According to this matrix, assuming that the weights of the six dimensions a1, a2, a3, a4, a5, and a6 are equal, we can get the six-dimensional radar charts of expert No. 7 and expert No. 1, as shown in Figure 2 and Figure 3 shown.

[0167] Assuming that the weights of the six dimensions are all 1, the total score of expert No. 7 is 2.34468168, and the total score of expert No. 1 is 2.4060609, so expert No. 1 is better than expert No. 7. In actual review work, review experience is often more important. When a6>1 is given, the advantage of expert No. 1 is more obvious; therefore, after optimizing the sorting, expert No. 1 is ranked higher.

[0168] The above matching algorithm can be expressed by this formula:

[0169]

[0170] Finally, collect user feedback and adjust the business parameter weights. For example, if the user feedback shows that the result of a certain business attribute is inaccurate, the algorithm will reduce the corresponding weight and rerun the process until the matching result meets expectations, such as the response time and matching accuracy reaching the preset thresholds.

[0171] Although the present invention has been disclosed as above with preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art may make some modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the definition of the claims.

Claims

1. A method for matching experts and projects based on an AI large language model, characterized in that: The steps include: S1) Data preprocessing: Data cleaning, denoising, word segmentation and labeling of project and expert information to remove duplication and maintain consistency of input data; S2) Text representation and feature extraction: Use AI big language model to vectorize the text information of projects and experts and extract key features; S3) Similarity calculation and multi-dimensional matching: Use multiple vector models to calculate the similarity between projects and experts, and sort them according to the similarity values. If the sorting results calculated by multiple vector models are inconsistent, take the collection of inconsistent data and rerank them. If the reranking result does not meet expectations, perform multi-dimensional calculation of the business similarity radar chart between each project and each expert based on specific business attributes, and recommend the most matching expert for project review in a weighted manner; S4) Result optimization and feedback: Based on user feedback and model self-learning, the matching algorithm and results are continuously optimized through iterative training and parameter adjustment until the response time and matching accuracy reach the preset threshold.

2. The expert and project matching method based on AI large language model according to claim 1, characterized in that: The data cleaning in step S1 includes: Remove duplicate records by unique identifier, which is project ID or expert ID; For missing values ​​in key fields, fill in the missing values ​​or directly delete the record; the key fields are expert research fields and / or project technical details, and the filling methods are mean filling, mode filling or specific value filling; Identify and handle outliers, including non-text data appearing in text fields, or numeric data outside of a reasonable range.

3. The expert and project matching method based on AI large language model as claimed in claim 1, characterized in that: The data denoising in step S1 includes: removing HTML tags, special characters and irrelevant punctuation marks in the text, and removing stop words.

4. The expert and project matching method based on AI large language model according to claim 1, characterized in that: The step S1 performs word segmentation processing on the project description and expert profile according to the Chinese word segmentation tool, and marks the expert research field and project technical keywords.

5. The expert and project matching method based on AI large language model according to claim 1, characterized in that: The key features in step S2 include the technical field, industry field, scientific field of the project and the expert, as well as the expert's research direction, research results, and review experience.

6. The expert and project matching method based on AI large language model according to claim 1, characterized in that: The step S3 uses n text embedding large models as vector models to calculate similarity. When the similarity results calculated by the n vector models are inconsistent, a collection is taken and the rerank model is used to sort the relevance of the taken collection. The calculation formula is as follows: Dataset=model_rerank(model1_result(query,raw_dataset),model2_result(query,raw_dataset),.........modeln_reset(query,raw_dataset)); Query is the key features of the project to be reviewed, and raw_dataset is the feature data set of multiple candidate experts.

7. The expert and project matching method based on AI large language model according to claim 6, characterized in that: The value of n is 2. The vector model used in step S3 is the text-ebedding-ada-002 model and the BAAI / bge-large-zh-v1.5 model of OpenAI, and the rerank model used is the BAAI / bge-reranker-large model.

8. The method for matching experts and projects based on an AI large language model as claimed in claim 6, wherein said step S3 recommends the most matching expert for project review in the following weighted manner: Dataset=model_improve( model_rerank(query,raw_dataset)); are weights of different dimensions.

9. The expert and project matching method based on AI large language model as claimed in claim 8, characterized in that: The step S3 selects 6 dimensions to calculate the expert radar chart, the value of n is 6, the weight a1 is the expert's technical field, the weight a2 is the expert's industry field, the weight a3 is the expert's scientific field, the weight a4 is the expert's research direction, the weight a5 is the expert's research results, and the weight a6 is the expert's review experience.

10. The expert and project matching method based on AI large language model according to claim 9, characterized in that: The step S3 calculates the similarity between the two experts and the corresponding projects as follows: first, the feature values ​​of the two candidate experts in the six dimensions are vectorized through the model to obtain a vector matrix embeddings_1, then the key features of the extracted project to be reviewed are vectorized to obtain another vector matrix embeddings_2, and finally the vector matrix embeddings_1 and the transposed matrix of the vector matrix embeddings_2 are multiplied to obtain a similarity matrix; each element (i, j) in the similarity matrix represents the similarity or compatibility score between the i-th data point in embeddings_1 and the j-th data point in embeddings_2; The step S4 first assumes that the weights of the six dimensions are are all 1, the top-ranked candidate expert is obtained, and then the value of the review experience a6 is increased. If the similarity or compatibility score of the top-ranked candidate expert further increases, the candidate expert is selected.

Images