Expert matching method and system based on big language model knowledge extension and fusion

Abstract

The invention discloses an expert matching method and system based on large language model knowledge extension and fusion. The method comprises the following steps: firstly, collecting expert basic information, and enhancing through a pre-trained big language model to obtain expert extension information; and then, integrating the basic information and the extended information, and analyzing and fusing the basic information and the extended information by using RoBerta and a knowledge graph pre-training model to form expert fusion representation. According to enterprise technical requirements, requirement details are expanded through a large language model, and technical requirement fusion representation is generated in combination with representation analysis of the technical requirements and enterprise information. And finally, carrying out keyword correlation screening on the experts, narrowing a matching range, comparing fusion representation of the experts and requirements, calculating similarity and sorting, and recommending related matching experts. Through the method and the device, skill matching analysis can be effectively performed on external expert information, and expert matching precision and recommendation effect are improved based on semantic analysis and expert matching of technical requirements.

Description

Technical Field

[0001] The present invention relates to the field of large language model analysis, and more specifically, to an expert matching method and system based on large language model knowledge expansion and fusion. Background Art

[0002] The technology transfer platform is used to promote technology transfer with talent as the core. Through the platform, the technology needs of enterprises and the scientific research talents of universities are linked to achieve effective docking of technological achievements and industrial technology needs. However, in the current platform, due to the large number of industries involved and the complex technical fields in which technical talents are proficient, it is difficult to accurately dock the talents' abilities with the needs of enterprises, and the docking efficiency is low. In addition, since most of the platform data comes from the filling and uploading of registered users, the keyword matching method based on the professional field of talents is often difficult to obtain good matching results due to the wide range of fields. Therefore, how to use external knowledge bases and databases to further improve the matching effect of experts is an important issue that needs to be solved at present.

[0003] In recent years, with the rapid development and popularization of large language models, their rich background knowledge and massive data training models have provided new possibilities for improving the external knowledge base of expert matching. Therefore, this scheme proposes an expert matching method and system based on the expansion and fusion of large language model knowledge. By promoting the matching of enterprise technology needs and experts, the level of scientific and technological innovation in the industry and the efficiency of scientific and technological achievement transformation in universities and research institutes can be improved. Summary of the invention

[0004] The present invention overcomes the defects of the prior art and proposes an expert matching method and system based on large language model knowledge expansion and fusion.

[0005] The first aspect of the present invention provides an expert matching method based on large language model knowledge expansion and fusion, comprising:

[0006] Obtaining basic information of experts in the target field, generating first query information based on a pre-trained large language model through the basic information of the experts, performing information query in the pre-trained large language model, and obtaining expert extended information based on the query results;

[0007] The expert basic information and the expert extended information are integrated into the intermediate information. The intermediate information is represented and analyzed by the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model respectively, and the two representations are spliced ​​and fused to obtain the expert fusion representation information.

[0008] Generate second query information based on the target enterprise's technical requirements and basic enterprise information, perform enterprise demand extension query through the large language model and the second query information to obtain demand extension information, perform representation analysis and fusion splicing on the technical requirements, basic enterprise information and demand extension information to obtain technical requirements fusion representation information;

[0009] Based on the keywords of technical requirements, all experts are analyzed and filtered for relevance, and a list of experts is obtained based on the filtering results;

[0010] Matching analysis is performed based on expert fusion representation information and technical demand fusion representation information, matching results are calculated by representation similarity, sorting is performed based on similarity, and expert recommendations are made based on the sorting results.

[0011] In this solution, the basic information of experts in the target field is obtained, and first query information based on a pre-trained large language model is generated through the basic information of the experts, and information query is performed in the pre-trained large language model, and the expert extended information is obtained based on the query result, specifically:

[0012] Obtain basic information of experts in the target field from the expert database;

[0013] In the expert basic information, determine whether there is extended information, if not, call the preset prompt template, and generate the first query information in combination with the expert basic information;

[0014] The first query information is sent to a pre-trained large language model for information query, the query result is converted into text information and expert extended information is obtained.

[0015] In this solution, the expert basic information and the expert extended information are integrated into intermediate information, and the intermediate information is represented and analyzed by the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model respectively, and the two obtained representations are spliced ​​and fused to obtain expert fusion representation information, including:

[0016] Integrate the expert basic information and the expert extended information into text information to form intermediate information;

[0017] Build a pre-trained representation model based on RoBerta;

[0018] De-noise the intermediate information and perform word segmentation through Roberta's byte pair encoding;

[0019] Encode the intermediate information and import it into the pre-trained representation model, and obtain the representation data through forward propagation and selection of the preset output layer;

[0020] Aggregate all output representation data to obtain the first text-based representation;

[0021] Import the expert basic information and expert extended information into the semantic analysis model to perform semantic analysis in three dimensions: entity, relationship, and attribute, and generate triple data;

[0022] Construct an expert knowledge graph through triple data;

[0023] In the expert knowledge graph, the entity and relationship data are vectorized through the TransE algorithm, and the graph vector representation data is obtained;

[0024] In the graph vector representation data, a similarity distance value is set, and knowledge reasoning is performed on the corresponding vectors of entity data and relationship data. The reasoning process is to calculate the distance between the selected vector and other domain vectors based on the cosine similarity method. If the distance is lower than the similarity distance value, it is judged that the selected vector has associated knowledge, and the association information between the corresponding domain vector and the selected vector is recorded. All vectors in the graph vector representation data are analyzed for association, and the obtained association information is used as the knowledge graph reasoning data;

[0025] Based on the knowledge graph reasoning data, the expert knowledge graph is supplemented with knowledge data. At the same time, the triple data in the expert knowledge graph is updated and the knowledge graph structure is updated;

[0026] The knowledge graph pre-trained representation model is used to perform text representation analysis based on entities, relationships and attributes on the expert knowledge graph to generate a second representation based on the knowledge graph.

[0027] In this solution, the expert basic information and the expert extended information are integrated into intermediate information, the intermediate information is represented and analyzed by the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model respectively, and the two obtained representations are spliced ​​and fused to obtain the expert fusion representation information, which also includes:

[0028] Obtaining a first representation and a second representation;

[0029] Based on the splicing and fusion method, the first representation and the second representation are spliced ​​into one representation data, and expert fusion representation information is obtained.

[0030] In this solution, the second query information is generated based on the technical requirements and basic information of the target enterprise, and the enterprise requirements are extended through the large language model and the second query information to obtain the extended requirements information, and the technical requirements, basic enterprise information and extended requirements information are characterized, analyzed and fused to obtain the technical requirements fusion representation information, which is specifically:

[0031] Obtain the target enterprise's technical requirements and basic enterprise information;

[0032] Calling a preset prompt template to generate second query information based on the technical requirements and basic information of the target enterprise;

[0033] Sending the second query information to the pre-trained large language model for information query, converting the query result into text information and obtaining the required expansion information;

[0034] Semantic representation analysis is performed based on technical requirements, basic enterprise information and extended requirements information, and the obtained representations are fused and spliced ​​to obtain technical requirement fusion representation information.

[0035] In this solution, the keywords based on technical requirements are used to perform correlation analysis and filtering on all experts, and a list of experts is obtained based on the filtering results, specifically:

[0036] Perform text conversion and word segmentation operations on technical requirements, and filter out corresponding technical keywords through word frequency calculation;

[0037] In the expert database, the domain keywords of each expert are extracted, and a correlation analysis is performed based on the domain keywords and technical keywords. Experts with a correlation of 0 are filtered out, and an expert list is obtained based on the filtering results.

[0038] In this solution, the matching analysis is performed based on the expert fusion representation information and the technical demand fusion representation information, the matching results are calculated by the similarity of the representations, the ranking is performed based on the similarity, and the expert recommendation is performed based on the ranking results, specifically:

[0039] Obtain the corresponding expert fusion representation information through the expert list and mark it as the current expert fusion representation information;

[0040] Perform matching analysis based on similarity between the current expert fusion representation information and the technical demand fusion representation information;

[0041] The similarity is calculated based on a preset distance calculation method, and the similarity is used as the matching degree to sort the expert list to obtain the sorting result;

[0042] According to the sorting results, a preset number of expert information is sent from the expert list to a preset user terminal for expert recommendation.

[0043] The second aspect of the present invention also provides an expert matching system based on large language model knowledge expansion and fusion, the system comprising: an expert information expansion module, a technical requirement expansion module, a prompt template management module, an expert information knowledge graph representation module, an expert information text representation module, an expert information fusion representation module, a technical requirement fusion representation module, a similarity calculation module, a matching filtering module, and a matching result output module. When the system is running, the following steps are implemented:

[0044] Obtaining basic information of experts in the target field, generating first query information based on a pre-trained large language model through the basic information of the experts, performing information query in the pre-trained large language model, and obtaining expert extended information based on the query results;

[0045] The expert basic information and the expert extended information are integrated into the intermediate information. The intermediate information is represented and analyzed by the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model respectively, and the two representations are spliced ​​and fused to obtain the expert fusion representation information.

[0046] Generate second query information based on the target enterprise's technical requirements and basic enterprise information, perform enterprise demand extension query through the large language model and the second query information to obtain demand extension information, perform representation analysis and fusion splicing on the technical requirements, basic enterprise information and demand extension information to obtain technical requirements fusion representation information;

[0047] Based on the keywords of technical requirements, all experts are analyzed and filtered for relevance, and a list of experts is obtained based on the filtering results;

[0048] Matching analysis is performed based on expert fusion representation information and technical demand fusion representation information, matching results are calculated by representation similarity, sorting is performed based on similarity, and expert recommendations are made based on the sorting results.

[0049] In this solution, the basic information of experts in the target field is obtained, and first query information based on a pre-trained large language model is generated through the basic information of the experts, and information query is performed in the pre-trained large language model, and the expert extended information is obtained based on the query result, specifically:

[0050] Obtain basic information of experts in the target field from the expert database;

[0051] In the expert basic information, determine whether there is extended information, if not, call the preset prompt template, and generate the first query information in combination with the expert basic information;

[0052] The first query information is sent to a pre-trained large language model for information query, the query result is converted into text information and expert extended information is obtained.

[0053] The third aspect of the present invention also provides a computer-readable storage medium, which includes an expert matching program based on the expansion and fusion of large language model knowledge. When the expert matching program based on the expansion and fusion of large language model knowledge is executed by a processor, the steps of the expert matching method based on the expansion and fusion of large language model knowledge as described in any of the above items are implemented.

[0054] The present invention discloses an expert matching method and system based on large language model knowledge expansion and fusion. First, the basic information of the expert is collected, and the expert extended information is obtained through pre-trained large language model enhancement; then, the basic information and extended information are integrated, and RoBerta and the knowledge graph pre-training model are used to analyze and fuse to form an expert fusion representation. According to the technical needs of the enterprise, the requirements details are expanded through the large language model, and the technical requirements and enterprise information are combined for representation analysis to generate a technical requirements fusion representation. Finally, the experts are screened for keyword relevance to narrow the matching range, the fusion representation of the experts and requirements is compared, the similarity is calculated and sorted, and relevant matching experts are recommended. Through the present invention, it is possible to effectively perform skill matching analysis on external expert information, and match experts based on semantic analysis of technical requirements to improve the expert matching accuracy and recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A flow chart of an expert matching method based on large language model knowledge expansion and fusion according to the present invention is shown;

[0056] Figure 2 The expert fusion representation information acquisition flow chart of the present invention is shown;

[0057] Figure 3 The flowchart of obtaining the information of the fusion characterization of the technical requirements of the present invention is shown;

[0058] Figure 4 A block diagram of an expert matching system based on large language model knowledge expansion and fusion according to the present invention is shown. DETAILED DESCRIPTION

[0059] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0060] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0061] In the problem of matching experts in different fields with industrial technology needs, an expert matching method and system based on large language model knowledge expansion and fusion is proposed to address the technical challenges brought by the fact that both sides of the match are described by multimodal data such as text and pictures.

[0062] Figure 1 A flow chart of an expert matching method based on large language model knowledge expansion and fusion according to the present invention is shown.

[0063] like Figure 1 As shown, the first aspect of the present invention provides an expert matching method based on large language model knowledge expansion and fusion, comprising:

[0064] S102, obtaining basic information of experts in the target field, generating first query information based on a pre-trained large language model through the basic information of the experts, performing information query in the pre-trained large language model, and obtaining expert extended information based on the query result;

[0065] S104, integrating the expert basic information and the expert extended information into intermediate information, performing representation analysis on the intermediate information through the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model respectively, and concatenating and fusing the two obtained representations to obtain expert fusion representation information;

[0066] S106, generating second query information based on the technical requirements and basic information of the target enterprise, performing enterprise requirement extension query through the large language model and the second query information to obtain extended requirement information, performing representation analysis and fusion splicing on the technical requirements, basic information of the enterprise and extended requirement information to obtain fusion representation information of the technical requirements;

[0067] S108, based on the keywords of the technical requirements, performing correlation analysis and filtering on all experts, and obtaining an expert list based on the filtering results;

[0068] S110, performing matching analysis based on the expert fusion representation information and the technical demand fusion representation information, calculating matching results by the similarity of the representations, sorting based on the similarity, and performing expert recommendation based on the sorting results.

[0069] It should be noted that after obtaining extended information from experts, we can obtain more data on the experts' technical capabilities, including but not limited to representative achievements and major industrial application information.

[0070] According to an embodiment of the present invention, the basic information of experts in the target field is obtained, first query information based on a pre-trained large language model is generated through the basic information of the experts, information query is performed in the pre-trained large language model, and expert extended information is obtained based on the query result, specifically:

[0071] Obtain basic information of experts in the target field from the expert database;

[0072] In the expert basic information, determine whether there is extended information, if not, call the preset prompt template, and generate the first query information in combination with the expert basic information;

[0073] The first query information is sent to a pre-trained large language model for information query, the query result is converted into text information and expert extended information is obtained.

[0074] It should be noted that the expert basic information includes information acquisition of multiple experts for subsequent screening and analysis. The pre-trained large language model can be a large language model such as GPT, BERT, etc.

[0075] In calling the preset prompt template, the following is a test implementation example demonstration:

[0076] Question: Look up information on the technical capabilities of physics expert Yang Zhenning and list his major scientific research achievements.

[0077] Answer: Chen-Ning Yang is an outstanding physicist. His major scientific research achievements include but are not limited to the following aspects:

[0078] 1. Mirror Symmetry Violation Theory: Together with Tsung-Dao Lee, he proposed this theory that mirror symmetry may be violated in weak interactions. This theory was proposed in 1957 and won them the Nobel Prize in Physics.

[0079] 2. Polaron theory: Collaborated with Woolf (TD Lee) to propose a theory about polarons, a type of meson.

[0080] 3. Microstructure of matter and high-energy physics: In high-energy physics, Yang Zhenning has made many contributions, including the understanding of particle physics and weak interactions.

[0081] 4. Statistical mechanics and non-equilibrium physics: He also made some important contributions in statistical mechanics and non-equilibrium physics.

[0082] 5. Education and Academic Leadership: Yang Zhenning not only has outstanding achievements in scientific research, but also has a profound influence in the field of education. He has taught at many well-known universities and mentored many students.

[0083] Overall, Yang Zhenning is known for his pioneering work in theoretical physics and particle physics, and for his contributions to physics education.

[0084] Perform information extraction and text conversion on the answer content to obtain the corresponding expert extended information.

[0085] Figure 2 The expert fusion representation information acquisition flow chart of the present invention is shown;

[0086] like Figure 2 The above is a brief flow chart of the analysis and acquisition of expert fusion representation information, and the specific process refers to the embodiment of the present invention;

[0087] According to an embodiment of the present invention, the expert basic information and the expert extended information are integrated into intermediate information, the intermediate information is characterized and analyzed respectively by the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model, and the two obtained representations are spliced ​​and fused to obtain expert fusion representation information, including:

[0088] Integrate the expert basic information and the expert extended information into text information to form intermediate information;

[0089] Build a pre-trained representation model based on RoBerta;

[0090] De-noise the intermediate information and perform word segmentation through Roberta's byte pair encoding;

[0091] Encode the intermediate information and import it into the pre-trained representation model, and obtain the representation data through forward propagation and selection of the preset output layer;

[0092] Aggregate all output representation data to obtain the first text-based representation;

[0093] Import the expert basic information and expert extended information into the semantic analysis model to perform semantic analysis in three dimensions: entity, relationship, and attribute, and generate triple data;

[0094] Construct an expert knowledge graph through triple data;

[0095] In the expert knowledge graph, the entity and relationship data are vectorized through the TransE algorithm, and the graph vector representation data is obtained;

[0096] In the graph vector representation data, a similarity distance value is set, and knowledge reasoning is performed on the corresponding vectors of entity data and relationship data. The reasoning process is to calculate the distance between the selected vector and other domain vectors based on the cosine similarity method. If the distance is lower than the similarity distance value, it is judged that the selected vector has associated knowledge, and the association information between the corresponding domain vector and the selected vector is recorded. All vectors in the graph vector representation data are analyzed for association, and the obtained association information is used as the knowledge graph reasoning data;

[0097] Based on the knowledge graph reasoning data, the expert knowledge graph is supplemented with knowledge data. At the same time, the triple data in the expert knowledge graph is updated and the knowledge graph structure is updated;

[0098] The knowledge graph pre-trained representation model is used to perform text representation analysis based on entities, relationships and attributes on the expert knowledge graph to generate a second representation based on the knowledge graph.

[0099] It should be noted that Roberta (Robustly Optimized BERT Pretraining Approach) is a deep learning model based on the Transformer architecture for natural language processing tasks. The semantic analysis model is a CNN-based semantic analysis model, which is trained using historical expert text information. Knowledge graph reasoning data is generally relational reasoning data. The knowledge graph pretraining representation model is a BERT-based semantic representation analysis model that extracts text semantic representations of triples (entities, relationships, attributes) in the knowledge graph.

[0100] It is worth mentioning that in the process of expert skill analysis and demand matching, traditional technologies are often limited to keyword matching of expert text information, and have poor accuracy and effect in complex demand analysis. Therefore, the present invention integrates two representations (information text and knowledge graph) to achieve multi-dimensional mining and analysis of expert skill information, and uses a distributed representation learning algorithm (TransE) to perform knowledge reasoning in the knowledge graph to mine the expert's related skill information and related technical information, and then performs splicing and fusion of the representations to achieve accurate matching of expert information and demand, thereby improving the recommendation effect.

[0101] According to an embodiment of the present invention, the expert basic information and the expert extended information are integrated into intermediate information, the intermediate information is characterized and analyzed respectively by the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model, and the two obtained representations are spliced ​​and fused to obtain expert fusion representation information, and further includes:

[0102] Obtaining a first representation and a second representation;

[0103] Based on the splicing and fusion method, the first representation and the second representation are spliced ​​into one representation data, and expert fusion representation information is obtained.

[0104] It should be noted that the splicing and fusion method can retain the original representation information to the greatest extent possible, and achieve more comprehensive and accurate retrieval analysis in the subsequent process.

[0105] Figure 3 The flowchart of obtaining the information of the fusion characterization of the technical requirements of the present invention is shown;

[0106] like Figure 3 The above is a brief flowchart of analyzing and obtaining the technical requirements fusion representation information, and the specific process refers to the embodiment of the present invention;

[0107] According to an embodiment of the present invention, the second query information is generated based on the technical requirements and basic information of the target enterprise, and the enterprise requirements extension query is performed through the large language model and the second query information to obtain the extended requirements information, and the technical requirements, the basic information of the enterprise and the extended requirements information are characterized, analyzed and fused to obtain the technical requirements fusion representation information, which is specifically:

[0108] Obtain the target enterprise's technical requirements and basic enterprise information;

[0109] Calling a preset prompt template to generate second query information based on the technical requirements and basic information of the target enterprise;

[0110] Sending the second query information to the pre-trained large language model for information query, converting the query result into text information and obtaining the required expansion information;

[0111] Semantic representation analysis is performed based on technical requirements, basic enterprise information and extended requirements information, and the obtained representations are fused and spliced ​​to obtain technical requirement fusion representation information.

[0112] It should be noted that, in the present invention, text representation can be extracted by word embedding or sentence vector etc. The demand extension information includes text information such as technical features and talent capability requirements in the enterprise demand.

[0113] After calling the preset template, in the process of obtaining the demand extension information through the second query information, the following is a test embodiment demonstration:

[0114] Q: There is a technical requirement to develop an expert recommendation system. Please help me find the technical information that needs to be used and the technical requirements of the talents needed.

[0115] A: Developing an expert recommendation system involves multiple technical fields and talent requirements. The following are the main technical information and talent technical requirements that need to be considered:

[0116] Technical information:

[0117] 1. Data collection and processing:

[0118] -Web scraping and data cleaning: Obtain data from various data sources (such as academic paper databases, expert social networks, internal enterprise systems, etc.), and perform cleaning, deduplication and standardization.

[0119] 2. Data storage and management:

[0120] -Database design and management: Select appropriate database technology (such as relational database or NoSQL database), design data models to store expert and related data, and ensure efficient retrieval and update of data.

[0121] 3. Information retrieval and recommendation algorithms:

[0122] - Natural Language Processing (NLP): used to extract key information from texts such as experts’ papers and project descriptions.

[0123] -Information retrieval technology: Building search engines or indexes for retrieving expert information.

[0124] - Recommendation system algorithm: Use machine learning algorithms (such as collaborative filtering, content recommendation, deep learning models, etc.) to recommend suitable experts to users.

[0125] 4. Front-end development:

[0126] -User Interface Design and Development: Design a user-friendly interface that enables users to easily enter queries, view recommended expert lists and detailed information.

[0127] 5. Backend development:

[0128] -Application logic development: implement the core functions of the system, handle user requests, call data processing and recommendation algorithms, and provide API interfaces for front-end calls.

[0129] 6. Security and Privacy:

[0130] -Data security and privacy protection: Ensure the security of user data and expert information, and comply with relevant laws, regulations and privacy policies.

[0131] Talent and technical requirements:

[0132] 1. Data Engineer:

[0133] - Proficient in data collection, cleaning and processing techniques.

[0134] -Familiar with various data storage and management systems, able to design and optimize database structures.

[0135] 2. Machine Learning Engineer:

[0136] -Have a deep understanding and practice of recommendation system algorithms and techniques.

[0137] - Proficiency in programming languages ​​such as Python and machine learning frameworks (e.g. TensorFlow, PyTorch).

[0138] 3. Software Engineer:

[0139] - Proficient in front-end and back-end development technologies, including front-end technologies such as HTML / CSS / JavaScript, and back-end frameworks such as Django, Flask, etc.

[0140] -Ability to design and implement scalable, stable applications.

[0141] 4. Information Retrieval Expert:

[0142] - Understand the principles and implementation of information retrieval technology and search engines.

[0143] -Familiarity with SEO related techniques is a plus.

[0144] 5. Security Expert:

[0145] - Familiar with best practices in cybersecurity and data privacy protection.

[0146] -Ability to evaluate and improve the security and privacy protection capabilities of the system.

[0147] 6. Project management and teamwork skills:

[0148] -Able to effectively manage project schedule and team collaboration to ensure projects are delivered on time and meet requirements.

[0149] According to an embodiment of the present invention, the keywords based on the technical requirements perform correlation analysis and filtering on all experts, and obtain an expert list based on the filtering results, specifically:

[0150] Perform text conversion and word segmentation operations on technical requirements, and filter out corresponding technical keywords through word frequency calculation;

[0151] In the expert database, the domain keywords of each expert are extracted, and a correlation analysis is performed based on the domain keywords and technical keywords. Experts with a correlation of 0 are filtered out, and an expert list is obtained based on the filtering results.

[0152] It should be noted that the relevance is 0, which means that in the semantic analysis of the corresponding expert information, no expert with a technical keyword or related technical terms appears in the field keywords.

[0153] According to an embodiment of the present invention, the matching analysis is performed based on the expert fusion representation information and the technical demand fusion representation information, the matching results are calculated by the similarity of the representations, the matching results are sorted based on the similarity, and the expert recommendation is performed based on the sorting results, specifically:

[0154] Obtain the corresponding expert fusion representation information through the expert list and mark it as the current expert fusion representation information;

[0155] Perform matching analysis based on similarity between the current expert fusion representation information and the technical demand fusion representation information;

[0156] The similarity is calculated based on a preset distance calculation method, and the similarity is used as the matching degree to sort the expert list to obtain the sorting result;

[0157] According to the sorting results, a preset number of expert information is sent from the expert list to a preset user terminal for expert recommendation.

[0158] It should be noted that the expert recommendation is to recommend a preset number of experts with high rankings. The preset distance calculation method includes Euclidean distance, cosine similarity, etc.

[0159] Figure 4 A block diagram of an expert matching system based on large language model knowledge expansion and fusion according to the present invention is shown.

[0160] The second aspect of the present invention also provides an expert matching system based on large language model knowledge expansion and fusion, the system comprising: an expert information expansion module, a technical requirement expansion module, a prompt template management module, an expert information knowledge graph representation module, an expert information text representation module, an expert information fusion representation module, a technical requirement fusion representation module, a similarity calculation module, a matching filtering module, and a matching result output module. When the system is running, the following steps are implemented:

[0161] Obtaining basic information of experts in the target field, generating first query information based on a pre-trained large language model through the basic information of the experts, performing information query in the pre-trained large language model, and obtaining expert extended information based on the query results;

[0162] The expert basic information and the expert extended information are integrated into the intermediate information. The intermediate information is represented and analyzed by the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model respectively, and the two representations are spliced ​​and fused to obtain the expert fusion representation information.

[0163] Generate second query information based on the target enterprise's technical requirements and basic enterprise information, perform enterprise demand extension query through the large language model and the second query information to obtain demand extension information, perform representation analysis and fusion splicing on the technical requirements, basic enterprise information and demand extension information to obtain technical requirements fusion representation information;

[0164] Based on the keywords of technical requirements, all experts are analyzed and filtered for relevance, and a list of experts is obtained based on the filtering results;

[0165] Matching analysis is performed based on expert fusion representation information and technical demand fusion representation information, matching results are calculated by representation similarity, sorting is performed based on similarity, and expert recommendations are made based on the sorting results.

[0166] The expert matching system based on large language model knowledge expansion and fusion specifically includes:

[0167] The expert information extension module is responsible for calling the pre-trained large language model (such as GPT, BERT, etc.) interface through the prompt template to obtain the corresponding expert extended information, and store the expert extended information in the database as the expert's extended information;

[0168] The technical requirement extension module uses the prompt module to call the pre-trained large language model interface to obtain the extended information of the technical requirements and store it in the database as the extended information of the technical requirements. For non-confidential technical requirements, the basic information of the enterprise can be added to the information of the prompt template;

[0169] The prompt template management module is responsible for managing the information templates that call the pre-trained large language model interface, including the addition, deletion, modification and query operations of template information;

[0170] The expert information knowledge graph representation module is responsible for extracting keywords from expert information and constructing a knowledge graph. It generates the expert's knowledge graph representation (i.e., the second representation) through the knowledge graph pre-training model.

[0171] The expert information text representation module is responsible for integrating the expert basic information with the expert extended information, and generating the text representation of the expert information (i.e., the first representation) through the pre-trained representation model;

[0172] The expert information fusion representation module is responsible for fusing the expert’s knowledge graph representation with the text representation;

[0173] The technology demand fusion representation module is responsible for combining the enterprise's technology demand, basic enterprise information and expanded information to generate the enterprise's technology demand fusion representation information;

[0174] The similarity calculation module is responsible for calculating the similarity between the expert representation and the requirement representation (the similarity calculation can use Euclidean distance or cosine similarity) to obtain the expert similarity ranking based on specific technical requirements;

[0175] The matching and filtering module, based on the keywords of technical requirements, filters out experts whose domain keywords do not contain the keywords of technical requirements at all, ensuring that at least one of the matched expert information contains the keywords of technical requirements, as a correction mechanism for the model;

[0176] The matching result output module outputs the expert matching results based on the requirements, and uses the similarity as the matching value for recommendation by users or downstream tasks. A preset number of recommended experts can be output.

[0177] The third aspect of the present invention also provides a computer-readable storage medium, which includes an expert matching program based on the expansion and fusion of large language model knowledge. When the expert matching program based on the expansion and fusion of large language model knowledge is executed by a processor, the steps of the expert matching method based on the expansion and fusion of large language model knowledge as described in any of the above items are implemented.

[0178] The present invention discloses an expert matching method and system based on large language model knowledge expansion and fusion. First, the basic information of the expert is collected, and the expert extended information is obtained through pre-trained large language model enhancement; then, the basic information and extended information are integrated, and RoBerta and the knowledge graph pre-training model are used to analyze and fuse to form an expert fusion representation. According to the technical needs of the enterprise, the requirements details are expanded through the large language model, and the technical requirements and enterprise information are combined for representation analysis to generate a technical requirements fusion representation. Finally, the experts are screened for keyword relevance to narrow the matching range, the fusion representation of the experts and requirements is compared, the similarity is calculated and sorted, and relevant matching experts are recommended. Through the present invention, it is possible to effectively perform skill matching analysis on external expert information, and match experts based on semantic analysis of technical requirements to improve the expert matching accuracy and recommendation effect.

[0179] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0180] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0181] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0182] Those skilled in the art can understand that: all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.

[0183] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0184] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An expert matching method based on large language model knowledge expansion and fusion, characterized in that: include: Obtaining basic information of experts in the target field, generating first query information based on a pre-trained large language model through the basic information of the experts, performing information query in the pre-trained large language model, and obtaining expert extended information based on the query results; The expert basic information and the expert extended information are integrated into the intermediate information. The intermediate information is represented and analyzed by the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model respectively, and the two representations are spliced ​​and fused to obtain the expert fusion representation information. Generate second query information based on the target enterprise's technical requirements and basic enterprise information, perform enterprise demand extension query through the large language model and the second query information to obtain demand extension information, perform representation analysis and fusion splicing on the technical requirements, basic enterprise information and demand extension information to obtain technical requirements fusion representation information; Based on the keywords of technical requirements, all experts are analyzed and filtered for relevance, and a list of experts is obtained based on the filtering results; Matching analysis is performed based on expert fusion representation information and technical demand fusion representation information, matching results are calculated by representation similarity, sorting is performed based on similarity, and expert recommendations are made based on the sorting results.

2. According to claim 1, an expert matching method based on large language model knowledge expansion and fusion is characterized in that: The method of obtaining basic information of experts in the target field, generating first query information based on a pre-trained large language model through the basic information of the experts, performing information query in the pre-trained large language model, and obtaining expert extended information based on the query result is specifically as follows: Obtain basic information of experts in the target field from the expert database; In the expert basic information, determine whether there is extended information, if not, call the preset prompt template, and generate the first query information in combination with the expert basic information; The first query information is sent to a pre-trained large language model for information query, the query result is converted into text information and expert extended information is obtained.

3. The expert matching method based on large language model knowledge expansion and fusion according to claim 2 is characterized in that: The expert basic information and the expert extended information are integrated into intermediate information, the intermediate information is represented and analyzed by the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model respectively, and the two obtained representations are spliced ​​and fused to obtain expert fusion representation information, including: Integrate the expert basic information and the expert extended information into text information to form intermediate information; Build a pre-trained representation model based on RoBerta; De-noise the intermediate information and perform word segmentation through Roberta's byte pair encoding; Encode the intermediate information and import it into the pre-trained representation model, and obtain the representation data through forward propagation and selection of the preset output layer; Aggregate all output representation data to obtain the first text-based representation; Import the expert basic information and expert extended information into the semantic analysis model to perform semantic analysis in three dimensions: entity, relationship, and attribute, and generate triple data; Construct an expert knowledge graph through triple data; In the expert knowledge graph, the entity and relationship data are vectorized through the TransE algorithm, and the graph vector representation data is obtained; In the graph vector representation data, a similarity distance value is set, and knowledge reasoning is performed on the corresponding vectors of entity data and relationship data. The reasoning process is to calculate the distance between the selected vector and other domain vectors based on the cosine similarity method. If the distance is lower than the similarity distance value, it is judged that the selected vector has associated knowledge, and the association information between the corresponding domain vector and the selected vector is recorded. All vectors in the graph vector representation data are analyzed for association, and the obtained association information is used as the knowledge graph reasoning data; Based on the knowledge graph reasoning data, the expert knowledge graph is supplemented with knowledge data. At the same time, the triple data in the expert knowledge graph is updated and the knowledge graph structure is updated; The knowledge graph pre-trained representation model is used to perform text representation analysis based on entities, relationships and attributes on the expert knowledge graph to generate a second representation based on the knowledge graph.

4. The expert matching method based on large language model knowledge expansion and fusion according to claim 3 is characterized in that: The expert basic information and the expert extended information are integrated into intermediate information, the intermediate information is represented and analyzed by the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model respectively, and the two obtained representations are spliced ​​and fused to obtain the expert fusion representation information, and also includes: Obtaining a first representation and a second representation; Based on the splicing and fusion method, the first representation and the second representation are spliced ​​into one representation data, and expert fusion representation information is obtained.

5. The expert matching method based on large language model knowledge expansion and fusion according to claim 4 is characterized in that: The second query information is generated based on the technical requirements and basic information of the target enterprise, and the enterprise requirements extension query is performed through the large language model and the second query information to obtain the extended requirements information, and the technical requirements, the basic information of the enterprise and the extended requirements information are characterized, analyzed and fused to obtain the technical requirements fusion representation information, specifically: Obtain the target enterprise's technical requirements and basic enterprise information; Calling a preset prompt template to generate second query information based on the technical requirements and basic information of the target enterprise; Sending the second query information to the pre-trained large language model for information query, converting the query result into text information and obtaining the required expansion information; Semantic representation analysis is performed based on technical requirements, basic enterprise information and extended requirements information, and the obtained representations are fused and spliced ​​to obtain technical requirement fusion representation information.

6. The expert matching method based on large language model knowledge expansion and fusion according to claim 5, characterized in that: Based on the keywords of the technical requirements, all experts are subjected to correlation analysis and filtering, and an expert list is obtained based on the filtering results, specifically: Perform text conversion and word segmentation operations on technical requirements, and filter out corresponding technical keywords through word frequency calculation; In the expert database, the domain keywords of each expert are extracted, and a correlation analysis is performed based on the domain keywords and technical keywords. Experts with a correlation of 0 are filtered out, and an expert list is obtained based on the filtering results.

7. The expert matching method based on large language model knowledge expansion and fusion according to claim 6, characterized in that: The matching analysis is performed based on the expert fusion representation information and the technical demand fusion representation information, the matching results are calculated by the similarity of the representations, the ranking is performed based on the similarity, and the expert recommendation is performed based on the ranking results, specifically: Obtain the corresponding expert fusion representation information through the expert list and mark it as the current expert fusion representation information; Perform matching analysis based on similarity between the current expert fusion representation information and the technical demand fusion representation information; The similarity is calculated based on a preset distance calculation method, and the similarity is used as the matching degree to sort the expert list to obtain the sorting result; According to the sorting results, a preset number of expert information is sent from the expert list to a preset user terminal for expert recommendation.

8. An expert matching system based on large language model knowledge expansion and fusion, characterized in that: The system includes: an expert information expansion module, a technical requirement expansion module, a prompt template management module, an expert information knowledge graph representation module, an expert information text representation module, an expert information fusion representation module, a technical requirement fusion representation module, a similarity calculation module, a matching filtering module, and a matching result output module. The system implements the following steps when it is running: Obtaining basic information of experts in the target field, generating first query information based on a pre-trained large language model through the basic information of the experts, performing information query in the pre-trained large language model, and obtaining expert extended information based on the query results; The expert basic information and the expert extended information are integrated into the intermediate information. The intermediate information is represented and analyzed by the RoBerta pre-trained representation model and the knowledge graph pre-trained representation model respectively, and the two representations are spliced ​​and fused to obtain the expert fusion representation information. Generate second query information based on the target enterprise's technical requirements and basic enterprise information, perform enterprise demand extension query through the large language model and the second query information to obtain demand extension information, perform representation analysis and fusion splicing on the technical requirements, basic enterprise information and demand extension information to obtain technical requirements fusion representation information; Based on the keywords of technical requirements, all experts are analyzed and filtered for relevance, and a list of experts is obtained based on the filtering results; Matching analysis is performed based on expert fusion representation information and technical demand fusion representation information, matching results are calculated by representation similarity, sorting is performed based on similarity, and expert recommendations are made based on the sorting results.

9. The expert matching system based on large language model knowledge expansion and fusion according to claim 8, characterized in that: The method of obtaining basic information of experts in the target field, generating first query information based on a pre-trained large language model through the basic information of the experts, performing information query in the pre-trained large language model, and obtaining expert extended information based on the query result is specifically as follows: Obtain basic information of experts in the target field from the expert database; In the expert basic information, determine whether there is extended information, if not, call the preset prompt template, and generate the first query information in combination with the expert basic information; The first query information is sent to a pre-trained large language model for information query, the query result is converted into text information and expert extended information is obtained.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes an expert matching program based on the expansion and fusion of large language model knowledge. When the expert matching program based on the expansion and fusion of large language model knowledge is executed by the processor, the steps of the expert matching method based on the expansion and fusion of large language model knowledge as described in any one of claims 1 to 7 are implemented.

Images