A training-free method for classifying scientific research project results

Abstract

A method for classifying scientific research project results that does not require training, comprising: using the title and abstract of the scientific research results as node features and the citation relationship as the edge to construct graph data. Using a pre-trained large language model to extract the vector representation of the results features. Using a parameter-free neighbor aggregation method, the vector information of the node and its neighbors is integrated to obtain the final vector representation. Define the category name of the results and provide a description. Use the same language model to generate vectors for the category name and description. Finally, by calculating the dot product similarity between the node vector and the category vector, select the category with the highest similarity as the predicted category of the result. The method for classifying scientific research project results of the present invention does not require training data, has high accuracy and fast reasoning ability, and provides an efficient solution for the automatic classification of scientific research outputs.

Description

Technical Field

[0001] The present invention relates to the technical field of data mining, and in particular to a scientific research project achievement classification method that does not require training. Background Art

[0002] In recent years, the automatic classification and indexing of scientific research outputs has become an important research direction in the field of information management. With the rapid increase in the number of scientific research papers, how to quickly and accurately classify these research results into appropriate subject categories not only helps academic knowledge sharing, but also improves research efficiency and the possibility of interdisciplinary cooperation. The task of scientific research output classification automatically identifies key information in literature, patents or scientific research reports, such as topics, research methods and application fields, by applying natural language processing (NLP) and machine learning techniques, and then classifies them into predetermined scientific fields or sub-topics. For example, for scientific research funding organizations and institutions, accurate classification of the output of a project is conducive to identifying whether the project output is consistent with the discipline when the project was applied. This helps scientific research funding organizations and institutions identify project risks.

[0003] In addition, the development of scientific research output classification technology is also crucial to the construction of knowledge management systems. For example, large libraries, academic databases, and scientific research institutions can use this technology to manage and retrieve scientific research materials more efficiently. Through accurate classification, researchers can find relevant research faster and avoid duplication of work, while also promoting the innovative process of scientific discovery. With the continuous advancement of artificial intelligence technology, the task of scientific research output classification is expected to be more intelligent in the future, able to handle more complex classification standards and larger data sets.

[0004] Existing research output classification methods are mainly based on text analysis and machine learning techniques. These methods can be divided into several categories: rule-based methods, supervised learning methods, and the recently emerging deep learning methods.

[0005] 1. Rule-based approach: This traditional approach relies on expert systems and handcrafted rules to identify keywords and phrases in texts, thereby classifying research outputs into relevant subject categories. Although this approach is reliable in terms of accuracy, it requires a lot of manual maintenance and is difficult to adapt to the rapid development and changes in subject areas.

[0006] 2. Supervised learning methods: These methods use labeled datasets to train classification models. Common techniques include support vector machines (SVMs), random forests, and gradient boosting machines. These methods can handle a large number of features and perform well on multi-class classification problems, but they usually require a large amount of labeled data and have limited adaptability to emerging topics.

[0007] 3. Deep learning methods: With the improvement of computing power and the availability of big data, deep learning methods have become a hot topic for the classification of scientific research outputs. By applying convolutional neural networks (CNNs), recurrent neural networks (RNNs), graph neural networks (GNNs), and recent transformer models such as BERT, these methods are able to gain a deeper understanding of the semantics of text, extract complex features, and show amazing generalization capabilities through transfer learning on unlabeled data. In addition, deep learning methods have also shown strong potential in processing unstructured data such as images and sounds, which provides new possibilities for the classification of cross-media scientific research outputs.

[0008] In summary, deep learning methods can provide faster speed and accuracy under the premise of collecting sufficient training data. For example, Cora, PubMed, CiteSeer and OGBN-Arxiv are commonly used datasets in scientific publication analysis and graph network research. Each dataset has its own unique characteristics and application areas. The Cora dataset is a citation network of academic papers, containing 2708 scientific research papers in the field of machine learning. These papers are divided into seven categories, including neural networks, probabilistic methods, genetic algorithms, etc. The citation relationship between papers is represented by undirected edges, with a total of 5429 edges. The PubMed dataset includes 19717 scientific research papers from the biomedical field. These papers are classified into three categories: diabetes, viral diseases, and oncology. The citation relationship between papers forms a network. The CiteSeer dataset is a citation graph containing 3312 papers from the field of computer and information science. The papers are divided into six categories, including artificial intelligence, databases, information retrieval, etc. OGBN-Arxiv is a dataset provided by Open Graph Benchmark, containing a citation network from arXiv papers. This dataset includes more than 169,343 papers collected from the field of computer science, covering multiple subfields in computer science. In the above datasets, GNN can achieve the best performance.

[0009] However, in practical applications, due to the lack of sufficient labeled data, the capabilities of deep learning methods cannot be well demonstrated. In addition, the performance of existing methods is often poor when faced with data scarcity and few-sample or zero-sample node classification tasks, making it difficult to meet the requirements of scientific research output classification tasks for high efficiency and high accuracy. In addition, existing methods have limitations in dealing with emerging topics and rapid changes in interdisciplinary fields, which limits the accuracy and generalization ability of scientific research results classification.

[0010] It should be noted that the information disclosed in the above background technology section is only used for understanding the background of the present application, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the invention

[0011] The main purpose of the present invention is to overcome the defects existing in the above-mentioned background technology and provide a scientific research project achievement classification method that does not require training.

[0012] To achieve the above object, the present invention adopts the following technical solutions:

[0013] A method for classifying scientific research project results without training includes the following steps:

[0014] S1: The title and abstract of the results are used as node features, and the citation relationship is used as the rule to establish the edges between nodes, so that the results are established as graph data;

[0015] S2: Use a pre-trained large language model to obtain vector representations of the title and summary of the results;

[0016] S3: The final vector representation of the results obtained using parameter-free neighbor aggregation;

[0017] S4: Define the category name of the outcome and give a description of the category name;

[0018] S5: Obtain vector representations of category names and category name descriptions using the same pre-trained large language model as in step S2;

[0019] S6: Calculate the dot product similarity between the vector obtained in step S3 and the vector obtained in step S5, and take the category with the greatest similarity to the node vector as the predicted category.

[0020] Furthermore, in step S1, the title and summary text of each node are cleaned and converted to lowercase, and the title and summary are concatenated into an input sequence separated by a delimiter.

[0021] Furthermore, in step S2, the input sequence is converted into a series of vector representations through the encoder of the large language model, the vector representation of the [CLS] tag is used as the overall representation of the node, and normalization is performed to ensure that the vectors between different nodes are consistent in scale and size.

[0022] Furthermore, in step S3, the adjacency matrix of the citation network of the results is used to perform parameter-free neighbor aggregation, including normalization of the adjacency matrix and subsequent feature aggregation.

[0023] Furthermore, in step S3, the parameter-free neighbor aggregation specifically includes the following steps:

[0024] Add self-loops to each node in the result graph to integrate the node's own features with those of its neighboring nodes;

[0025] Calculate the normalized adjacency matrix of the subgraph of the result graph by multiplying the adjacency matrix by the inverse square root of the degree matrix;

[0026] Perform iterative feature aggregation, where node features are updated by normalizing the adjacency matrix to aggregate the features of the node itself and its neighboring nodes;

[0027] The iterative process continues for a predefined number of steps, allowing the node embedding to incorporate information from the expanded neighborhood.

[0028] Furthermore, the subgraph of the result graph has an adjacency matrix A, a degree matrix M, and a feature matrix H processed by a large language model;

[0029] The normalized adjacency matrix of a subgraph is calculated as follows:

[0030]

[0031] Using the normalized adjacency matrix, in each iteration, the node features are updated by aggregating features from the node itself and its neighboring nodes, expressed as:

[0032]

[0033] in, Indicates The feature matrix at iteration .

[0034] Furthermore, in step S5, the vector representation of the category name and the category name description is obtained using the same method as step S2, and the category name and description are processed using the same method as step S3 to obtain the category vector.

[0035] Furthermore, in step S6, the dot product similarity is used to represent the similarity between the node representation and the category representation, and the calculation formula is as follows:

[0036]

[0037] in, y ′ represents the predicted category, that is, the category index with the largest similarity is taken as the prediction result; Find the function Reaching the maximum value, is the index variable; Represents a node vector With category vector The similarity between them; N represents the total number of categories.

[0038] Furthermore, the large language model is a BERT model.

[0039] A computer program product includes a computer program, which, when executed by a processor, implements the method for classifying scientific research project results without training.

[0040] The present invention has the following beneficial effects:

[0041] The present invention provides a method for classifying scientific research project results without training, which effectively solves the problem of data scarcity in scientific research results classification tasks and improves classification performance. The method of the present invention does not require any training data, and can achieve or even exceed the effect of methods that require training data in the classification of scientific research results. The present invention abstracts the problem of scientific research project output classification into a node classification problem, obtains the vector representation of scientific research output using a pre-trained language model, and obtains the final scientific research project vector representation through parameter-free neighbor aggregation, showing excellent transferability, and can classify all scientific research project output data without prior training. In addition, the present invention far exceeds existing methods in deployment and reasoning speed, which not only improves efficiency but also reduces costs, especially when processing large-scale data sets. Experimental results show that the accuracy and F1 scores of the present invention on data sets such as Cora, PubMed, CiteSeer and OGBN-Arxiv are better than other methods, and can adapt to different pre-trained language models, proving the effectiveness and robustness of its method. Compared with the existing technology, the present invention can accurately obtain node classification results without training and fine-tuning, and has obvious advantages in technology, cost and efficiency. At the same time, it fills the gap in the market for scientific research results classification methods that do not require training data.

[0042] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is an overall flow chart of the scientific research project achievement classification method that does not require training according to an embodiment of the present invention.

[0044] Figure 2 This is a framework diagram of the scientific research project achievement classification method that does not require training according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is only exemplary and is not intended to limit the scope and application of the present invention.

[0046] See also Figure 1 and Figure 2 The embodiment of the present invention provides a method for classifying scientific research project results without training, comprising the following steps:

[0047] Step S1: Use the title and abstract of the results as node features, and use the reference relationship as the rule for establishing edges between nodes to establish the results as graph data.

[0048] In a preferred embodiment, the title and summary text of each node are cleaned, converted to lower case, and the title and summary are concatenated into an input sequence separated by a delimiter.

[0049] Step S2: Use the pre-trained large language model to obtain the vector representation of the title and summary of the results.

[0050] In a preferred embodiment, the input sequence is converted into a series of vector representations through an encoder of a large language model, the vector representation of the [CLS] tag is used as the overall representation of the node, and normalization is performed to ensure that the vectors between different nodes are consistent in scale and size.

[0051] Step S3: Obtain the final vector representation of the result using parameter-free neighbor aggregation;

[0052] In a preferred embodiment, the adjacency matrix of the citation network of the results is used to perform parameter-free neighbor aggregation, including normalization of the adjacency matrix and subsequent feature aggregation.

[0053] In a preferred embodiment, parameter-free neighbor aggregation specifically includes the following steps: adding self-loops to each node in the result graph to integrate the node's own features with the features of its neighboring nodes; calculating the normalized adjacency matrix of the subgraph of the result graph by multiplying the adjacency matrix by the inverse square root of the degree matrix; performing iterative feature aggregation, in which the node features are updated by the normalized adjacency matrix to aggregate the features of the node itself and its neighboring nodes; the iterative process continues for a predefined number of steps so that the node embedding can integrate information from the extended neighborhood.

[0054] Step S4: Define the category name of the achievement and give a description of the category name.

[0055] Step S5: Use the same pre-trained large language model as in step S2 to obtain vector representations of category names and category name descriptions.

[0056] In some embodiments, the same method as step S2 can be used to obtain a vector representation of the category name and the category name description, and the same method as step S3 can be used to process the category name and description to obtain a category vector.

[0057] Step S6: Calculate the dot product similarity between the vector obtained in step S3 and the vector obtained in step S5, and take the category with the greatest similarity to the node vector as the predicted category.

[0058] In some embodiments, the large language model used in the present invention may be (but not limited to) a BERT model.

[0059] The training-free scientific research project achievement classification method of the present invention effectively solves the data scarcity problem in the classification of scientific research outputs. The method uses a pre-trained large-scale language model to convert the title and abstract of the scientific research achievement into a vector representation, and adopts a parameter-free neighbor aggregation technology to realize the graph data construction and classification of scientific research achievements. This method not only improves the accuracy and efficiency of classification, but also far exceeds the existing technology in deployment and reasoning speed because it does not require training data. The experimental results (which will be further described later) show that the classification performance of the present invention on multiple data sets is better than that of traditional methods, and it has good transferability and can adapt to different pre-trained language models. In addition, the training-free feature of the present invention gives it significant advantages in technology, cost and efficiency, effectively filling the market gap and providing an efficient scientific research achievement management and retrieval tool for scientific research fund issuing organizations, academic databases and scientific research institutions.

[0060] The specific embodiments and experimental verification of the present invention are further described below.

[0061] Scientific research outputs often cite previous work, and data can be represented by graphs. ) and a set of edges (denoted by ) that show the connections or relationships between nodes. Each node in the graph Each with a feature vector This vector captures the properties or characteristics of the node. Taking the paper as an example, the paper collection Each element Represents a paper, category collection ,in is the total number of categories, domain In Indicates the paper Belong to the field . Reference relationship ,in Indicates the paper Cited papers . Feature Matrix ,in is the number of papers, is the dimension of the feature. Label , represents the domain label of each paper. The goal of the classification problem is to learn a mapping , this mapping is able to predict the category of unlabeled papers.

[0062] The present invention provides a scientific research project achievement classification method without training, comprising the following steps:

[0063] Step S1: Use the title and abstract of the achievement as node features, and the reference relationship as the rule for establishing edges between nodes to establish the achievement as graph data.

[0064] Step S2: Use a pre-trained large language model (such as BERT) to obtain vector representations of the title and summary of the results.

[0065] Step S3: Use parameter-free neighbor aggregation to obtain the final vector representation of the results.

[0066] Step S4: Define the category name and give the category name description.

[0067] Step S5: Use the same pre-trained large language model as in step S2 to obtain vector representations of category names and category name descriptions.

[0068] Step S6: Calculate the dot product similarity between the vector obtained in step S3 and the vector obtained in step S5, and take the category with the greatest similarity to the node vector as the predicted category.

[0069] Specifically, in order to use the language model to obtain the vector representation of the node title and summary, the title and summary text of each node are first collected and necessary preprocessing is performed on them. In the preprocessing, data cleaning is first performed to convert all texts to lowercase. Then, the title and summary are concatenated into an input sequence, and a separator (such as [SEP]) is used to separate the title and summary. Next, this input sequence is converted into a series of vector representations through the encoder of the language model. The vector representation of the [CLS] tag in the language model is directly used as the overall representation of the node. In order to further improve the usability of the vector, the vector is standardized to ensure that the vectors between different nodes are consistent in scale and size. After these steps, the node title and summary text are successfully converted into a vector of fixed dimension.

[0070] Secondly, we use the adjacency matrix of the reference network to perform parameter-free neighbor aggregation. First, we assume that a subgraph has an adjacency matrix A, a degree matrix M, and a feature matrix H processed by a language model. Neighborhood aggregation includes normalization of the adjacency matrix and subsequent feature aggregation. First, we add a self-loop to each node in the graph. This step integrates the node's own features with the features of its neighboring nodes.

[0071] Next, the normalized adjacency matrix of the subgraph is calculated as follows:

[0072]

[0073] Using this normalized adjacency matrix, neighborhood aggregation is performed in an iterative manner. In each iteration, node features are updated by aggregating features from its direct neighbors (including the node itself). The mathematical representation of this process is:

[0074]

[0075] in, Indicates The iterative process continues for a predefined number of steps so that the node embedding can gradually integrate information from the expanded neighborhood.

[0076] Next, define the category name. Since the method of the present invention does not require training, the category name can be arbitrarily defined here according to the ideas of human experts. In order to add semantics to enhance the effect of subsequent similarity calculation, it is necessary to provide a description of the category while providing the category name. After defining the category name and description, the category name and description are processed using the same method as the node text to obtain the category vector.

[0077] For each node's final vector representation and the category's vector representation, the dot product similarity is used to represent the similarity between the node representation and the category representation. The following formula is used for calculation:

[0078]

[0079] in, y ′ represents the predicted category, that is, the category index with the largest similarity is taken as the prediction result; Find the function Reaching the maximum value, is the index variable; Represents node vector (scientific research results) With category vector The similarity between them; N represents the total number of categories.

[0080] Finally, the category with the greatest similarity to the node vector is taken as the predicted category.

[0081] Examples

[0082] Taking four real-world paper citation networks as examples, the scientific research project achievement classification process described in the present invention is further described in detail. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments.

[0083] This example uses four datasets, Cora, PubMed, CiteSeer, and OGBN-Arxiv, as examples. Cora, PubMed, and CiteSeer are commonly used datasets in scientific publication analysis and graph network research. Each dataset has its own unique characteristics and application areas. The Cora dataset is a citation network of academic papers, containing 2,708 scientific research papers in the field of machine learning. These papers are divided into seven categories, including neural networks, probabilistic methods, genetic algorithms, etc. The citation relationship between papers is represented by undirected edges, with a total of 5,429 edges. The PubMed dataset includes 19,717 scientific research papers from the biomedical field, which are classified into three categories: diabetes, viral diseases, and oncology. The citation relationships between papers form a network. The CiteSeer dataset is a citation graph containing 3,312 papers from the field of computer and information science. The papers are divided into six categories, including artificial intelligence, databases, information retrieval, etc.

[0084] Experimental Results

[0085] The above methods were deployed on three datasets and compared with four advanced methods. It should be noted that among all the methods, only the method of the present invention does not require training. Accuracy and macro-F1 are used as evaluation indicators. Using Accuracy and macro-F1 as classification evaluation indicators each has its own unique advantages and can provide valuable performance evaluation information in different scenarios. Accuracy is one of the most commonly used classification evaluation indicators, which indicates the proportion of samples predicted correctly by the model to the total samples. Its calculation is simple and clear, so it is easy to understand and apply. Accuracy can intuitively reflect the overall performance of the model for the entire dataset. Accuracy is a useful indicator if the dataset categories are evenly distributed or the overall performance of the model is of primary concern. Accuracy is fast to calculate and does not require complex computing resources, making it suitable for use in real-time or large-scale applications. XWMacro-F1 is a macro average of the F1 score, which averages the F1 scores of each category without considering the size of the category. In this way, each category has the same weight in the evaluation, so it is particularly suitable for datasets with imbalanced categories and can avoid the dominant influence of large categories on the results. Macro-F1 can better reflect the classification performance of minority classes. Even if a category appears less frequently in the dataset, Macro-F1 ensures that the category is not ignored in the overall evaluation. The F1 score combines the information of precision and recall, and can more comprehensively evaluate the classification ability of the model on different categories. Macro-F1 emphasizes the consistent performance of the model on all categories, not just the majority of categories. The results are shown in Table 1, which shows that the method proposed in this invention is superior to the method that requires training.

[0086] Table 1

[0087]

[0088] Table 1 shows that the accuracy Acc and F1 scores of the method without training of the present invention on three data sets are better than those of other methods.

[0089] In addition, as shown in Table 2, different pre-trained language models are used, demonstrating that the method of the present invention is effective under different language models.

[0090] Table 2

[0091]

[0092] Table 2 shows the outstanding performance of the method of the present invention under different language models.

[0093] An embodiment of the present invention further provides a storage medium for storing a computer program, which at least performs the above method when executed.

[0094] An embodiment of the present invention further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute at least the method described above when executing the computer program.

[0095] An embodiment of the present invention further provides a processor, wherein the processor executes a computer program and at least executes the method described above.

[0096] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a ferromagnetic random access memory (Flash Memory), a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk memory or a tape memory. The storage medium described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0097] In the several embodiments provided by the present invention, it should be understood that the disclosed systems 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.

[0098] 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, that is, 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.

[0099] 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.

[0100] A person of ordinary skill in the art can understand that: all or part of the steps of implementing the above method embodiment 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 embodiment; 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, etc. Various media that can store program codes.

[0101] 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.

[0102] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0103] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0104] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0105] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for classifying scientific research project results without training, characterized in that: The steps include: S1: The title and abstract of the results are used as node features, and the citation relationship is used as the rule to establish the edges between nodes, so that the results are established as graph data; S2: Use a pre-trained large language model to obtain vector representations of the title and summary of the results; S3: The final vector representation of the results obtained using parameter-free neighbor aggregation; S4: Define the category name of the outcome and give a description of the category name; S5: Obtain vector representations of category names and category name descriptions using the same pre-trained large language model as in step S2; S6: Calculate the dot product similarity between the vector obtained in step S3 and the vector obtained in step S5, and take the category with the greatest similarity to the node vector as the predicted category; In step S3, the adjacency matrix of the citation network of the result is used to perform parameter-free neighbor aggregation, including normalization of the adjacency matrix and subsequent feature aggregation, which specifically includes the following steps: Add self-loops to each node in the result graph to integrate the node's own features with those of its neighboring nodes; Calculate the normalized adjacency matrix of the subgraph of the result graph by multiplying the adjacency matrix by the inverse square root of the degree matrix; Perform iterative feature aggregation, where node features are updated by normalizing the adjacency matrix to aggregate the features of the node itself and its neighboring nodes; The iterative process continues for a predefined number of steps, allowing the node embedding to incorporate information from the expanded neighborhood.

2. The method for classifying scientific research project results without training as claimed in claim 1, characterized in that: In step S1, the title and summary text of each node are cleaned, converted to lowercase, and the title and summary are concatenated into an input sequence separated by a delimiter.

3. The method for classifying scientific research project results without training as claimed in claim 1, characterized in that: In step S2, the input sequence is converted into a series of vector representations through the encoder of a large language model, and the vector representation of the [CLS] tag is used as the overall representation of the node, and is standardized to ensure that the vectors between different nodes are consistent in scale and size.

4. The method for classifying scientific research project results without training as claimed in claim 1, characterized in that: The subgraph of the result graph has an adjacency matrix A, a degree matrix M, and a feature matrix H processed by a large language model; The normalized adjacency matrix of a subgraph is calculated as follows: ； Using the normalized adjacency matrix, in each iteration, the node features are updated by aggregating features from the node itself and its neighboring nodes, expressed as: ； in, Indicates The feature matrix at iteration .

5. The method for classifying scientific research project results without training as claimed in any one of claims 1 to 4, characterized in that: In step S5, the vector representation of the category name and the category name description is obtained using the same method as step S2, and the category name and description are processed using the same method as step S3 to obtain the category vector.

6. The method for classifying scientific research project results without training as claimed in any one of claims 1 to 4, characterized in that: In step S6, dot product similarity is used to represent the similarity between node representation and category representation, and the calculation formula is as follows: ； in, y ′ represents the predicted category, that is, the category index with the largest similarity is used as the prediction result; argmax i It means to find the function sim( h n , h ci ) reaches the maximum value i value, i is an index variable; sim( h n , h ci ) represents the node vector h n With category vector h ci The similarity between N Indicates the total number of categories.

7. The method for classifying scientific research project results without training as claimed in any one of claims 1 to 4, characterized in that: The large language model is a BERT model.

8. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for classifying scientific research project results without training as described in any one of claims 1 to 7 is implemented.

Images