Technical methods for mining and aggregation of scientific and technological terms in professional fields

Abstract

The present application provides a technology for mining and aggregating scientific and technological terms in a professional field, and relates to the technical field of alignment of scientific and technological entities, including: constructing a target scientific and technological graph based on the results of extraction of scientific and technological entity elements; encoding the target scientific and technological graph through a graph encoder, and outputting a graph node vector; performing scientific and technological term recognition and obtaining scientific and technological terms; extracting the contextual information of scientific and technological terms, generating term-related definitions in combination with a large language model, and encoding and outputting term definition text vectors; fine-tuning and optimizing the graph encoder during the text alignment process, and embedding the optimized graph encoder into the large language model framework, and aggregating the terms to be recognized through the large language model that integrates the graph structure information. The present application can solve the technical problem that the prior art often lacks sufficient flexibility and accuracy when dealing with complex relationships and semantic similarities between scientific and technological terms, resulting in poor accuracy and reliability of scientific and technological term alignment.

Description

Technical Field

[0001] The present application relates to the technical field of scientific and technological entity alignment, and in particular to a technology for mining and aggregating scientific and technological terms in a professional field. Background Art

[0002] As the core carrier of scientific research results, the analysis and understanding of scientific and technological literature plays a vital role in scientific progress and technological innovation. In this process, the identification of scientific and technological terms is a key step in achieving deep knowledge mining, because these scientific and technological terms and their relationships constitute the basis of scientific knowledge networks. However, traditional identification methods often rely on simple rule matching, which is often limited by the diversity of professional terms and the complexity of context, making it difficult to fully capture the deep meaning of scientific and technological literature.

[0003] At present, existing technologies often lack sufficient flexibility and accuracy when dealing with complex relationships and semantic similarities between scientific and technological terms, resulting in technical problems such as poor accuracy and reliability of scientific and technological term alignment. Summary of the invention

[0004] The purpose of this application is to provide a technology for mining and aggregating scientific and technological terms in a professional field, so as to solve the technical problem that the existing technology often lacks sufficient flexibility and accuracy when dealing with complex relationships and semantic similarities between scientific and technological terms, resulting in poor accuracy and reliability of scientific and technological term alignment.

[0005] In view of the above problems, the present application provides a technology for mining and aggregating scientific and technological terms in professional fields, including: taking the scope of professional disciplines as constraints, retrieving and acquiring a sample set of scientific and technological documents, preprocessing the sample set of scientific and technological documents, and obtaining a standard set of scientific and technological documents; using a large language model to extract scientific and technological entity elements from the standard set of scientific and technological documents according to a predetermined extraction strategy, and constructing a target scientific and technological graph based on the results of the scientific and technological entity element extraction; encoding the target scientific and technological graph through a graph encoder and outputting a graph node vector; based on predetermined recognition rules, performing scientific and technological term recognition on the results of the scientific and technological entity element extraction to obtain scientific and technological terms, wherein the predetermined recognition rules include entity type level, terminology level, The word length level and the term frequency level; based on the context-aware mechanism, the context information of the scientific and technological terms in the standard scientific and technological literature set is extracted using keywords, and the term-related definitions are generated in combination with the large language model. The term-related definitions are encoded through the text encoder in the large language model, and the term definition text vector is output; the text of the graph node vector and the term definition text vector are gradually aligned, and the graph encoder is fine-tuned and optimized during the text alignment process to obtain an optimized graph encoder; the optimized graph encoder is embedded in the large language model framework to generate a large language model that integrates graph structure information, and the terms to be recognized are aggregated through the large language model that integrates graph structure information.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] By taking the scope of professional disciplines as constraints, a sample scientific and technological literature set is retrieved and obtained, and the sample scientific and technological literature set is preprocessed to obtain a standard scientific and technological literature set; using a large language model, scientific and technological entity elements are extracted from the standard scientific and technological literature set according to a predetermined extraction strategy, and a target scientific and technological graph is constructed based on the results of the scientific and technological entity element extraction; the target scientific and technological graph is encoded by a graph encoder to output a graph node vector; based on predetermined recognition rules, scientific and technological terminology is identified on the results of the scientific and technological entity element extraction to obtain scientific and technological terms, wherein the predetermined recognition rules include entity type level, term length level and term frequency level; based on a context-aware mechanism, keywords are used to extract context information of scientific and technological terms in the standard scientific and technological literature set, and term-related definitions are generated in combination with a large language model, and the term-related definitions are encoded by a text encoder in the large language model to output a term definition text vector; text alignment of the graph node vector and the term definition text vector is performed step by step, and the graph encoder is fine-tuned and optimized during the text alignment process to obtain an optimized graph encoder; the optimized graph encoder is embedded in a large language model framework to generate a large language model that integrates graph structure information, and the terms to be identified are aggregated by the large language model that integrates graph structure information. In other words, by introducing the intelligent knowledge extraction and complex semantic understanding capabilities of a large language model, the accuracy of scientific and technological terminology recognition and the comprehensiveness of terminology knowledge can be improved. At the same time, by comprehensively utilizing graph information and context information, multi-level matching judgments can be carried out, thereby achieving more accurate alignment of scientific and technological terms.

[0008] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented according to the contents of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are specifically cited below. It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the present application or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0010] Figure 1A flowchart of a technology for mining and aggregating scientific and technological terms in a professional field is provided for this application;

[0011] Figure 2 A flowchart of outputting term definition text vectors in a professional field scientific and technological term mining and aggregation technology for this application. DETAILED DESCRIPTION

[0012] This application provides a technology for mining and aggregating scientific and technological terms in professional fields, which solves the technical problem that the existing technology often lacks sufficient flexibility and accuracy when dealing with complex relationships and semantic similarities between scientific and technological terms, resulting in poor accuracy and reliability of scientific and technological term alignment. By introducing the intelligent knowledge extraction and complex semantic understanding capabilities of large language models, the accuracy of scientific and technological term recognition and the comprehensiveness of term knowledge can be improved. At the same time, the comprehensive use of graph information and context information can be carried out to carry out multi-level matching judgments, thereby achieving more accurate scientific and technological term alignment.

[0013] Below, the technical solutions in the present application will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments of the present application. It should be understood that the present application is not limited to the example embodiments described herein. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application. It should also be noted that, for the convenience of description, only the parts related to the present application are shown in the accompanying drawings, rather than all of them.

[0014] For examples, please see the attached Figure 1 , this application provides a professional field scientific and technological term mining and aggregation technology, which specifically includes the following steps:

[0015] Step 1: Retrieve and obtain a sample scientific and technological literature set based on the scope of professional disciplines, and pre-process the sample scientific and technological literature set to obtain a standard scientific and technological literature set.

[0016] Specifically, firstly, with the scope of professional disciplines as constraints, relevant literature retrieval is performed based on big data technology to obtain a sample set of scientific and technological literature, where the scope of professional disciplines can be set by technicians in the field according to actual application scenarios, such as computer science, electronic science, communication technology, etc. Then, the sample set of scientific and technological literature is preprocessed, including text cleaning, word segmentation, and removal of stop words, etc., to optimize the input quality of the model and obtain a standard set of scientific and technological literature.

[0017] Step 2: Using a large language model, extract scientific and technological entity elements from the standard scientific and technological literature set according to a predetermined extraction strategy, and construct a target scientific and technological map based on the results of the scientific and technological entity element extraction.

[0018] Specifically, the big language model is a natural language processing technology based on deep learning, which is specially used to generate and understand human language. By training on a large amount of text data, it learns the grammar, semantics and contextual relationships of the language, so that it can generate coherent text, answer questions and conduct conversations. Representatives of big language models include GPT, ChatGLM and LLaMA, etc., and its application range is wide, including machine translation, text summarization, sentiment analysis and dialogue systems.

[0019] A large language model is used to extract scientific and technological entity elements from the standard scientific and technological literature set according to a predetermined extraction strategy, wherein the scientific and technological entity elements include features such as scientific and technological entities, attributes and relationships in the literature, that is, a large language model is used to perform comprehensive scanning and recognition tasks, and the model is guided to identify scientific and technological entities, attributes and relationships in the literature through carefully designed extraction prompts, wherein entities cover key elements such as scholars, methods, problems, materials, tools, indicators and data sets; relationships include proposed, belonging to, used for, produced and measured, etc., to reveal the interactions between entities; attributes include name abbreviations, data set versions, etc., to provide a richer description for entities, and obtain scientific and technological entity element extraction results, wherein the scientific and technological entity element extraction results include entities, attributes and relationships.

[0020] The technology graph is a highly structured knowledge representation framework that captures the relationships and attributes between technology entities through complex node and edge relationships. The graph covers a range of entity types, such as scholars, methods, problems, materials, tools, indicators, and datasets. Each entity is described in detail through a series of predefined attributes, such as name abbreviations, dataset versions, etc., to ensure its semantic richness. Relationship types include proposed, belonging to, used for, produced, and measured, which reveal the direct connection between entities. Then, based on the results of technology entity element extraction, the entities and attributes are defined as nodes of the graph, and the relationships are used as edges of the graph to construct the target technology graph.

[0021] Step 3: Encode the target technology graph through a graph encoder and output a graph node vector.

[0022] Specifically, graph structure encoding is a technology that represents graph data in a form that can be processed by neural networks. By converting the nodes in the graph and their interconnected relationships (edges) into feature vectors, neural networks can analyze and process graph data. Common methods include graph convolutional networks (GCN), graph attention networks (GAT), and graph autoencoders. Graph structure encoding extracts features from the local neighborhood of nodes and the global graph structure through multi-layer convolution or attention mechanisms, thereby generating high-dimensional node embedding representations.

[0023] First, a graph encoder is constructed based on a graph neural network; then the target technology graph is encoded through the graph encoder, that is, the graph structure information in the technology graph is encoded using a graph neural network (GNN), and the text information is integrated into the encoding process of the graph structure through a comparative learning method. With the help of graph convolutional network technology, the deep fusion of text information and graph structure encoding of the technology graph is realized. The core idea of ​​GCN is to update the representation of each node by aggregating neighborhood information so that it can better capture the semantic features in the global graph structure. The representation of the node is updated and aggregated through multiple graph convolutional layers, so that the information of the node not only contains its own features, but also the features of its neighboring nodes, realizing the effective transmission and fusion of information. Each node in the graph corresponds to an entity, and the edge corresponds to the relationship between entities. Through this method, a vector representation of each entity corresponding to the graph node is generated to obtain the graph node vector.

[0024] Step 4: Based on predetermined recognition rules, scientific and technological terminology is identified on the scientific and technological entity element extraction results to obtain scientific and technological terms, wherein the predetermined recognition rules include entity type level, term length level and term frequency level.

[0025] Specifically, first, the predetermined recognition rules are obtained, among which the predetermined recognition rules include the entity type level, the term length level and the term frequency level. The entity type level refers to selecting only specific entities, such as methods, materials, tools, etc., and excluding abstract concepts, such as problems, values, etc.; the term length level refers to the number of characters in the term. Usually, scientific and technological terms are shorter and can be limited to 3 to 5 words. Terms that are too long may not be specific; the term frequency level refers to the frequency of the term appearing in the relevant literature. For example, terms that appear more frequently in the relevant literature are more likely to be scientific and technological terms. The frequency can be set according to actual needs.

[0026] Then, based on the predetermined recognition rules, the scientific and technological terminology recognition is performed on the scientific and technological entity element extraction results, wherein scientific and technological terminology refers to the professional terminology used to describe specific entities in the field of science and technology. These entities include methods, materials, tools, etc., which are the core components of scientific and technological activities. Scientific and technological terminology is obtained based on the terminology recognition results.

[0027] Step 5: Based on the context-aware mechanism, use keywords to extract the context information of scientific and technological terms in the standard scientific and technological literature set, combine with the large language model to generate term-related definitions, encode the term-related definitions through the text encoder in the large language model, and output the term definition text vector.

[0028] Specifically, context awareness means that when processing text information, not only the meaning of a single word or phrase itself is considered, but also its environment and contextual relationship in the text. Context awareness technology usually involves advanced deep learning technologies such as attention mechanism and Transformer architecture to achieve in-depth understanding and analysis of text context.

[0029] Based on the context-aware mechanism, the context information of the scientific and technological terms in the standard scientific and technological literature set is extracted, and the term-related definitions are generated in combination with the large language model. First, the large language model is used to deeply analyze the scientific and technological literature. Through pre-training and fine-tuning, the model can understand the complex semantics and professional terms in the scientific and technological literature. Using the context-aware mechanism, the model can extract context fragments related to scientific and technological terms. These fragments include key information such as usage context, functional description, and attribute characteristics. Then, based on the extracted context fragments, the model generates detailed term-related definitions. In this way, the term-related definitions not only contain basic information, but also reflect their specific applications and attribute characteristics in different research fields, enhancing the semantic expression ability and practicality of the graph.

[0030] Finally, the term-related definitions are encoded by a text encoder in the large language model to output a term definition text vector, wherein the text encoder is embedded in the large language model.

[0031] Step 6: Perform text alignment between the graph node vector and the term definition text vector step by step, and fine-tune and optimize the graph encoder during the text alignment process to obtain an optimized graph encoder.

[0032] Specifically, the text of the graph node vector and the term definition text vector is further gradually aligned, and the graph encoder is fine-tuned and optimized during the text alignment process, wherein a two-stage graph instruction fine-tuning strategy is adopted to fine-tune and optimize the graph encoder during the text alignment process to obtain an optimized graph encoder. First, in the first stage, a structure-aware graph matching task is constructed, and the language model is guided by natural language labels to match the vector representation of each node in the graph with the text description of the term corresponding to the node. Through contrast learning fine-tuning, the parameters of the large language model and the graph encoder are fixed, and only the parameters of the structure-aware projector on the vector representation of each node are optimized; in the second stage, the large language model is fine-tuned using task-specific graph instructions, based on the parameters of the structure-aware projector trained in the first stage, the parameters of the graph encoder are kept unchanged, and the parameters of the projector and the large language model are further optimized to generate a response that is more suitable for the current graph learning task, and enhance its adaptability and performance when processing scientific and technological graph data; in this way, the graph encoder is fine-tuned and optimized to obtain an optimized graph encoder.

[0033] Step 7: embed the optimized graph encoder into a large language model framework to generate a large language model that integrates graph structure information, and aggregate the terms to be recognized through the large language model that integrates graph structure information.

[0034] Specifically, the optimized graph encoder is finally embedded in the large language model framework to generate a large language model that integrates graph structure information; then, the large language model that integrates graph structure information is used to align and aggregate the terms to be recognized, that is, the large language model that integrates graph structure information is used to perform pairwise alignment and judgment on the terms to be recognized in turn, using terms, attributes and context information as the basis for judgment, and adopting multi-level matching strategies such as name matching, attribute feature matching and relationship matching to improve the accuracy of term alignment judgment.

[0035] By providing a technology for mining and aggregating scientific and technological terms in professional fields, we can solve the technical problem that existing technologies often lack sufficient flexibility and accuracy when dealing with complex relationships and semantic similarities between scientific and technological terms, resulting in poor accuracy and reliability of scientific and technological term alignment. By introducing the intelligent knowledge extraction and complex semantic understanding capabilities of large language models, we can improve the accuracy of scientific and technological term recognition and the comprehensiveness of term knowledge. At the same time, we can make comprehensive use of graph information and context information to carry out multi-level matching judgments, thereby achieving more accurate scientific and technological term alignment.

[0036] Furthermore, based on the extraction results of scientific and technological entity elements, the target scientific and technological map is constructed. Step 2 of this application includes:

[0037] Configure a predetermined extraction strategy, wherein the predetermined extraction strategy includes entity extraction, relationship extraction and attribute extraction; utilize a large language model to extract scientific and technological entity elements from the standard scientific and technological literature set according to the predetermined extraction strategy to obtain element extraction results; based on graph computing principles, set entities and attributes as nodes in the graph, set entity relationships as edges in the graph, perform graph simulation based on the element extraction results, and generate the target scientific and technological graph.

[0038] Specifically, first, a predetermined extraction strategy is configured, wherein the predetermined extraction strategy includes entity extraction, relationship extraction, and attribute extraction. Then, a large language model is used to extract scientific and technological entity elements from the standard scientific and technological literature set according to the predetermined extraction strategy. In entity recognition, the model identifies key elements including scholars, methods, problems, materials, tools, indicators, and data sets, and uses these entities as nodes in the graph; in relationship extraction, the interactions between entities are identified, and these relationships are represented in the graph in the form of edges; in attribute enrichment, the name abbreviation of the entity, the data set version, and other attributes are extracted to provide a richer description for the entity, enhance the semantic expression ability of the graph, and obtain the element extraction result.

[0039] Further based on the graph computing principle, entities and attributes are set as nodes in the graph, entity relationships are set as edges in the graph, and graph simulation is performed according to the element extraction results to generate a target technology graph, which contains accurate entity recognition, clear relationship definitions, and detailed attribute descriptions. In this way, not only can a technology graph containing rich entities and relationships be constructed, but it can also ensure that each entity has a detailed attribute description, thereby providing a solid foundation for in-depth analysis and knowledge discovery of scientific and technological literature.

[0040] Further, the graph node vector is outputted. Step three of this application includes:

[0041] An initial graph encoder is constructed based on a graph convolutional network, wherein the initial graph encoder includes multiple graph convolutional layers; based on a predetermined loss function, the initial graph encoder is supervised trained through a sample data set with the goal of minimizing cross entropy loss to obtain a graph encoder that meets expected convergence indicators; the target technology graph is encoded through the graph encoder, and the graph node vector is output.

[0042] Specifically, first, an initial graph encoder is constructed based on a graph convolutional network, wherein the initial graph encoder includes multiple graph convolutional layers, that is, with the help of graph convolutional network technology, the deep fusion of graph structure encoding of text information and technology graphs is realized. The core idea of ​​GCN is to update the representation of each node by aggregating neighborhood information so that it can better capture the semantic features in the global graph structure. The representation of the node is updated and aggregated through multiple graph convolutional layers, so that the information of the node not only contains its own features, but also the features of its neighboring nodes, thereby realizing the effective transmission and fusion of information.

[0043] Then, based on the predetermined loss function, the initial graph encoder is supervised and trained through a sample data set with the goal of minimizing the cross entropy loss. During the training process, the model parameters are optimized by minimizing the cross entropy loss function, so that the vector representation of the nodes in the graph contains various known associations, thereby embedding the association information of each node in the graph and obtaining a graph encoder that meets the expected convergence index. The use of graph convolutional network technology can effectively integrate text information with the graph structure of the science and technology map, and better capture the complex relationships between nodes through multi-level encoding and representation learning, so as to achieve accurate understanding and efficient processing of entity information in the graph.

[0044] Finally, the target technology graph is encoded by the graph encoder to output the graph node vector.

[0045] Further, the output term definition text vector is as follows: Figure 2 As shown, step five of this application includes:

[0046] The context information is comprehensively analyzed by using context fusion technology through a large language model to generate term-related definitions; the term-related definitions are optimized and corrected according to a predetermined verification scheme to obtain optimized term definitions; the optimized term definitions are encoded by the text encoder to obtain the term definition text vector, wherein the text encoder is embedded in the large language model.

[0047] Specifically, this application has achieved intelligent generation of term-related definitions by introducing advanced context-aware mechanisms. It uses large language models and context fusion technology to deeply analyze scientific and technological documents. Through pre-training and fine-tuning, the model can understand the complex semantics and professional terms in scientific and technological documents.

[0048] First, through a large language model, the context information is comprehensively analyzed using context fusion technology to generate term-related definitions. Context fusion technology enhances the accuracy and depth of term definitions by integrating relevant fragments and information in scientific and technological literature. In this way, term-related definitions not only contain basic information, but also reflect their specific applications and attribute characteristics in different research fields, thereby enhancing the semantic expression ability and practicality of the graph.

[0049] Then, the relevant definitions of the terms are optimized and corrected according to a predetermined verification scheme, that is, the definitions are further optimized and verified. The contextual information in more literature and the manual review of experts can be combined to ensure the accuracy and comprehensiveness of the definitions and obtain optimized term definitions.

[0050] Then, with the help of a large language model, the optimized term text is encoded through a text encoder to obtain a term definition text vector, that is, the vector representation of the node in the graph is used as the vector representation of the entity, and the node vector representation corresponding to the entity is concatenated with the vector representation containing the term definition text generated by the large language model, effectively associating the graph information with its related term text description information as a representation of the scientific and technological term.

[0051] Further, to obtain an optimized graph encoder, step six of the present application includes:

[0052] Perform text alignment of the graph node vector and the term definition text vector; in the process of gradual text alignment, perform first-stage fine-tuning optimization and second-stage fine-tuning optimization on the graph encoder based on the graph instruction fine-tuning strategy to obtain the optimized graph encoder.

[0053] Specifically, first, the text alignment of the graph node vector and the term definition text vector is performed. The above model framework directly integrates the graph encoder with pre-trained parameters, thereby seamlessly integrating the function of the graph encoder and ensuring that the graph structure information is highly aligned with the text information. Then, in the text step-by-step alignment process, a two-stage graph instruction fine-tuning strategy is adopted, that is, in the text step-by-step alignment process, the graph encoder is fine-tuned and optimized in the first stage and the second stage based on the graph instruction fine-tuning strategy, and the optimized graph encoder is output.

[0054] Further, the first stage of fine-tuning optimization is performed, and the present application also includes the following steps:

[0055] The large language model is guided by natural language labels to gradually align the graph node vector and the term definition text vector; based on the comparative learning fine-tuning of the graph node vector and the term definition text vector, the parameters of the large language model and the graph encoder are fixed, and the parameters of the structure-aware projector are optimized to obtain optimized projection parameters.

[0056] Specifically, first, in the first stage, a structure-aware graph matching task is constructed, and the language model is guided by natural language labels to match the vector representation of each node in the graph with the text description of the scientific and technological terms corresponding to the node, so as to achieve the gradual alignment of the graph node vector and the term definition text vector; then, the graph node vector and the term definition text vector are subjected to contrastive learning and fine-tuning, and further through contrastive learning and fine-tuning, the parameters of the large language model and the graph encoder are fixed, and only the parameters of the structure-aware projector on the vector representation of each node are optimized to obtain the optimized projection parameters.

[0057] Further, the second stage of fine-tuning optimization is performed, and the present application also includes the following steps:

[0058] After the first stage fine-tuning optimization is completed, the parameters of the graph encoder are kept unchanged, and the large language model and the structure-aware projector are optimized based on the optimized projection parameters.

[0059] Specifically, after the fine-tuning optimization in the first stage is completed, the large language model is fine-tuned using task-specific graph instructions in the second stage. Based on the parameters of the structure-aware projector trained in the first stage, the parameters of the graph encoder are kept unchanged, and the large language model and the structure-aware projector are optimized based on the optimized projection parameters. Focus is placed on further optimizing the parameters of the projector and the large language model to generate responses that are more suitable for the current graph learning task, thereby enhancing its adaptability and performance when processing science and technology graph data.

[0060] Furthermore, the terms to be recognized are aggregated through the large language model integrating the graph structure information. Step seven of the present application includes:

[0061] According to the predetermined multidimensional features, multidimensional features are extracted from scientific and technological literature through comparison with the large language model that integrates the graph structure information, and the multidimensional feature extraction results are used as judgment conditions for term alignment, and the judgment conditions are placed in the model problem, wherein the predetermined multidimensional features at least include terms, attributes and relationships; based on the model problem, the terms to be identified are polled and aligned in pairs in turn, and the term aggregation results are output, wherein if the model answer is "yes", the term alignment judgment result is consistent, and if the model answer is "no", the term alignment judgment result is inconsistent.

[0062] Specifically, first, obtain the comparison scientific and technological documents. Then obtain the predetermined multidimensional features, wherein the predetermined multidimensional features at least include terms, attributes and relationships; further, according to the predetermined multidimensional features, extract the multidimensional features of the comparison scientific and technological documents through the large language model that integrates the graph structure information, obtain the multidimensional feature extraction results, and use the multidimensional feature extraction results as the judgment conditions for term alignment, and place the judgment conditions in the model problem. Then, based on the model problem, perform pairwise polling alignment judgments on the terms to be identified in turn, wherein if the model answer is "yes", the term alignment judgment result is consistent, and if the model answer is "no", the term alignment judgment result is inconsistent; finally, multiple scientific and technological terms with "yes" judgment results are aggregated to characterize the multiple scientific and technological terms as the same category, and output the term aggregation results.

[0063] In summary, the technology for mining and aggregating professional technical terms provided by this application has the following technical effects:

[0064] 1. By introducing the intelligent knowledge extraction and complex semantic understanding capabilities of large language models, the accuracy of scientific and technological terminology recognition and the comprehensiveness of terminology knowledge can be improved. At the same time, by comprehensively utilizing graph information and context information, multi-level matching judgments can be carried out to achieve more accurate alignment of scientific and technological terms.

[0065] 2. Through context fusion technology, the extracted fragments are comprehensively analyzed to generate accurate and comprehensive term definition texts to enhance the semantic understanding and description depth of scientific and technological terms. Combined with more contextual information from literature and expert manual review, the term definition texts are optimized and verified to ensure their accuracy and comprehensiveness.

[0066] 3. Based on the semantic understanding ability of the large language model, the similarity of scientific and technological terms in different documents is calculated, and the precise alignment of scientific and technological terms is achieved through multi-level matching strategies such as name matching, attribute feature matching, and relationship matching. The results of the preliminary alignment of the model are verified and adjusted by combining intelligent judgment with manual review to ensure the reliability of the alignment results.

[0067] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0068] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technology, the present application is also intended to include these modifications and variations.

Claims

1. A method for mining and aggregating scientific and technological terms in a professional field, characterized in that: include: Restricted by the scope of professional disciplines, a sample scientific and technological literature set is retrieved and obtained, and the sample scientific and technological literature set is preprocessed to obtain a standard scientific and technological literature set; Using a large language model, extracting scientific and technological entity elements from the standard scientific and technological literature set according to a predetermined extraction strategy, and constructing a target scientific and technological map based on the scientific and technological entity element extraction results; Encode the target technology graph through a graph encoder and output a graph node vector; Based on predetermined recognition rules, scientific and technological term recognition is performed on the scientific and technological entity element extraction results to obtain scientific and technological terms, wherein the predetermined recognition rules include entity type level, term length level and term frequency level; Based on the context-aware mechanism, the context information of the scientific and technological terms in the standard scientific and technological literature set is extracted using keywords, and the term-related definitions are generated in combination with the large language model. The term-related definitions are encoded by the text encoder in the large language model, and the term definition text vector is output; Performing text stepwise alignment of the graph node vector and the term definition text vector, and fine-tuning and optimizing the graph encoder during the text alignment process to obtain an optimized graph encoder; The optimized graph encoder is embedded in a large language model framework to generate a large language model integrating graph structure information, and the to-be-recognized terms are aggregated through the large language model integrating graph structure information; Construct the target technology map based on the extraction results of technology entity elements, including: Configuring a predetermined extraction strategy, wherein the predetermined extraction strategy includes entity extraction, relationship extraction, and attribute extraction; Using a large language model, extracting scientific and technological entity elements from the standard scientific and technological document set according to the predetermined extraction strategy to obtain an element extraction result; Based on the graph computing principle, entities and attributes are set as nodes in the graph, entity relationships are set as edges in the graph, and graph simulation is performed according to the element extraction results to generate the target technology graph; Output graph node vector, including: Constructing an initial graph encoder based on a graph convolutional network, wherein the initial graph encoder includes a plurality of graph convolutional layers; Based on a predetermined loss function, with the goal of minimizing cross entropy loss, supervised training is performed on the initial graph encoder through a sample data set to obtain a graph encoder that meets expected convergence indicators; The target technology graph is encoded by the graph encoder, and the graph node vector is output.

2. The method of mining and aggregating professional technical terms according to claim 1, characterized in that: Output term definition text vectors, including: Through a large language model, the context information is comprehensively analyzed using context fusion technology to generate term-related definitions; Optimizing and correcting the definitions related to the term according to a predetermined verification scheme to obtain optimized term definitions; The optimized term definition is encoded by the text encoder to obtain the term definition text vector, wherein the text encoder is embedded in a large language model.

3. The method of mining and aggregating professional technical terms according to claim 1, characterized in that: Get an optimized graph encoder, including: performing text alignment of the graph node vector and the term definition text vector; In the process of gradual text alignment, the graph encoder is subjected to first-stage fine-tuning optimization and second-stage fine-tuning optimization based on a graph instruction fine-tuning strategy to obtain the optimized graph encoder.

4. The method of mining and aggregating professional technical terms according to claim 3, characterized in that: Conduct the first phase of fine-tuning and optimization, including: Instructing a large language model to gradually align the graph node vector and the term definition text vector through natural language labels; Based on the comparative learning fine-tuning of the graph node vector and the term definition text vector, the parameters of the large language model and the graph encoder are fixed, and the parameters of the structure-aware projector are optimized to obtain optimized projection parameters.

5. The method of mining and aggregating professional technical terms according to claim 4, characterized in that: Conduct the second phase of fine-tuning and optimization, including: After the first stage fine-tuning optimization is completed, the parameters of the graph encoder are kept unchanged, and the large language model and the structure-aware projector are optimized based on the optimized projection parameters.

6. The method of mining and aggregating professional technical terms according to claim 1, characterized in that: The large language model integrating the graph structure information is used to aggregate the terms to be recognized, including: According to the predetermined multidimensional features, multidimensional features of scientific and technological documents are extracted through the large language model comparison of the fusion graph structure information, the multidimensional feature extraction results are used as the judgment conditions for term alignment, and the judgment conditions are placed in the model problem, wherein the predetermined multidimensional features at least include terms, attributes and relationships; Based on the model question, the terms to be identified are aligned and judged in pairs in turn, and the term aggregation results are output. If the model answer is "yes", the term alignment judgment result is consistent; if the model answer is "no", the term alignment judgment result is inconsistent.

Images